Extension of `data.frame`

Fast aggregation of large data (e.g. 100GB in RAM), fast ordered joins, fast add/modify/delete of columns by group using no copies at all, list columns, a fast friendly file reader and parallel file writer. Offers a natural and flexible syntax, for faster development.


  1. fwrite(..., quote='auto') already quoted a field if it contained a sep or \n, or sep2[2] when list columns are present. Now it also quotes a field if it contains a double quote (") as documented, #1925. Thanks to Aki Matsuo for reporting. Tests added. The qmethod tests did test escaping embedded double quotes, but only when sep or \n was present in the field as well to trigger the quoting of the field.

  2. Fixed 3 test failures on Solaris only, #1934. Two were on both sparc and x86 and related to a tzone attribute difference between as.POSIXct and as.POSIXlt even when passed the default tz="". The third was on sparc only: a minor rounding issue in fwrite() of 1e-305.

  3. Regression crash fixed when 0's occur at the end of a non-empty subset of an empty table, #1937. Thanks Arun for tracking down. Tests added. For example, subsetting the empty DT=data.table(a=character()) with DT[c(1,0)] should return a 1 row result with one NA since 1 is past the end of nrow(DT)==0, the same result as DT[1].

  4. Fixed newly reported crash that also occurred in old v1.9.6 when by=.EACHI, nomatch=0, the first item in i has no match AND j has a function call that is passed a key column, #1933. Many thanks to Reino Bruner for finding and reporting with a reproducible example. Tests added.

  5. Fixed fread() error occurring for a subset of Windows users: showProgress is not type integer but type 'logical'., #1944 and #1111. Our tests cover this usage (it is just default usage), pass on AppVeyor (Windows), win-builder (Windows) and CRAN's Windows so perhaps it only occurs on a specific and different version of Windows to all those. Thanks to @demydd for reporting. Fixed by using strictly logical type at R level and Rboolean at C level, consistently throughout.

  6. Combining on= (new in v1.9.6) with by= or keyby= gave incorrect results, #1943. Many thanks to Henrik-P for the detailed and reproducible report. Tests added.

  7. New function rleidv was ignoring its cols argument, #1942. Thanks Josh O'Brien for reporting. Tests added.

  1. It seems OpenMP is not available on CRAN's Mac platform; NOTEs appeared in CRAN checks for v1.9.8. Moved Rprintf from init.c to packageStartupMessage to avoid the NOTE as requested urgently by Professor Ripley. Also fixed the bad grammar of the message: 'single threaded' now 'single-threaded'. If you have a Mac and run macOS or OS X on it (I run Ubuntu on mine) please contact CRAN maintainers and/or Apple if you'd like CRAN's Mac binary to support OpenMP. Otherwise, please follow these instructions for OpenMP on Mac which people have reported success with.

  2. Just to state explicitly: data.table does not now depend on or require OpenMP. If you don't have it (as on CRAN's Mac it appears but not in general on Mac) then data.table should build, run and pass all tests just fine.

  3. There are now 5,910 raw tests as reported by test.data.table(). Tests cover 91% of the 4k lines of R and 89% of the 7k lines of C. These stats are now known thanks to Jim Hester's Covr package and Codecov.io. If anyone is looking for something to help with, creating tests to hit the missed lines shown by clicking the R and src folders at the bottom here would be very much appreciated.

  4. The FAQ vignette has been revised given the changes in v1.9.8. In particular, the very first FAQ.

  5. With hindsight, the last release v1.9.8 should have been named v1.10.0 to convey it wasn't just a patch release from .6 to .8 owing to the 'potentially breaking changes' items. Thanks to @neomantic for correctly pointing out. The best we can do now is now bump to 1.10.0.

  1. By default all columns are now used by unique(), duplicated() and uniqueN() data.table methods, #1284 and #1841. To restore old behaviour: options(datatable.old.unique.by.key=TRUE). In 1 year this option to restore the old default will be deprecated with warning. In 2 years the option will be removed. Please explicitly pass by=key(DT) for clarity. Only code that relies on the default is affected. 266 CRAN and Bioconductor packages using data.table were checked before release. 9 needed to change and were notified. Any lines of code without test coverage will have been missed by these checks. Any packages not on CRAN or Bioconductor were not checked.

  2. A new column is guaranteed with := even when there are no matches or when its RHS is length 0 (e.g. integer(), numeric()) but not NULL. The NA column is created with the same type as the empty RHS. This is for consistency so that whether a new column is added or not does not depend on whether i matched to 1 or more rows or not. See #759 for further details and examples.

  3. When j contains no unquoted variable names (whether column names or not), with= is now automatically set to FALSE. Thus, DT[,1], DT[,"someCol"], DT[,c("colA","colB")] and DT[,100:109] now work as we all expect them to; i.e., returning columns, #1188, #1149. Since there are no variable names there is no ambiguity as to what was intended. DT[,colName1:colName2] no longer needs with=FALSE either since that is also unambiguous. That is a single call to the : function so with=TRUE could make no sense, despite the presence of unquoted variable names. These changes can be made since nobody can be using the existing behaviour of returning back the literal j value since that can never be useful. This provides a new ability and should not break any existing code. Selecting a single column still returns a 1-column data.table (not a vector, unlike data.frame by default) for type consistency for code (e.g. within DT[...][...] chains) that can sometimes select several columns and sometime one, as has always been the case in data.table. In future, DT[,myCols] (i.e. a single variable name) will look for myCols in calling scope without needing to set with=FALSE too, just as a single symbol appearing in i does already. The new behaviour can be turned on now by setting the tersely named option: options(datatable.WhenJisSymbolThenCallingScope=TRUE). The default is currently FALSE to give you time to change your code. In this future state, one way (i.e. DT[,theColName]) to select the column as a vector rather than a 1-column data.table will no longer work leaving the two other ways that have always worked remaining (since data.table is still just a list after all): DT[["someCol"]] and DT$someCol. Those base R methods are faster too (when iterated many times) by avoiding the small argument checking overhead inside the more flexible DT[...] syntax as has been highlighted in example(data.table) for many years. In the next release, DT[,someCol] will continue with old current behaviour but start to warn if the new option is not set. Then the default will change to TRUE to nudge you to move forward whilst still retaining a way for you to restore old behaviour for this feature only, whilst still allowing you to benefit from other new features of the latest release without changing your code. Then finally after an estimated 2 years from now, the option will be removed.

  1. fwrite() - parallel .csv writer:
* Thanks to Otto Seiskari for the initial pull request [#580](https://github.com/Rdatatable/data.table/issues/580) that provided C code, R wrapper, manual page and extensive tests.
* From there Matt parallelized and specialized C functions for writing integer/numeric exactly matching `write.csv` between 2.225074e-308 and 1.797693e+308 to 15 significant figures, dates (between 0000-03-01 and 9999-12-31), times down to microseconds in POSIXct, automatic quoting, `bit64::integer64`, `row.names` and `sep2` for `list` columns where each cell can itself be a vector. See [this blog post](http://blog.h2o.ai/2016/04/fast-csv-writing-for-r/) for implementation details and benchmarks.
* Accepts any `list` of same length vectors; e.g. `data.frame` and `data.table`.
* Caught in development before release to CRAN: thanks to Francesco Grossetti for [#1725](https://github.com/Rdatatable/data.table/issues/1725) (NA handling), Torsten Betz for [#1847](https://github.com/Rdatatable/data.table/issues/1847) (rounding of 9.999999999999998) and @ambils for [#1903](https://github.com/Rdatatable/data.table/issues/1903) (> 1 million columns).
* `fwrite` status was tracked here: [#1664](https://github.com/Rdatatable/data.table/issues/1664)
  1. fread():
* gains `quote` argument. `quote = ""` disables quoting altogether which reads each field *as is*, [#1367](https://github.com/Rdatatable/data.table/issues/1367). Thanks @manimal.
* With [#1462](https://github.com/Rdatatable/data.table/issues/1462) fix, quotes are handled slightly better. Thanks  @Pascal for [posting on SO](http://stackoverflow.com/q/34144314/559784).
* gains `blank.lines.skip` argument that continues reading by skipping empty lines. Default is `FALSE` for backwards compatibility, [#530](https://github.com/Rdatatable/data.table/issues/530). Thanks @DirkJonker. Also closes [#1575](https://github.com/Rdatatable/data.table/issues/1575).
* gains `fill` argument with default `FALSE` for backwards compatibility. Closes [#536](https://github.com/Rdatatable/data.table/issues/536). Also, `fill=TRUE` prioritises maximum cols instead of longest run with identical columns when `fill=TRUE` which allows handle missing columns slightly more robustly, [#1573](https://github.com/Rdatatable/data.table/issues/1573).
* gains `key` argument, [#590](https://github.com/Rdatatable/data.table/issues/590).
* gains `file` argument which expects existing file on input, to ensure no shell commands will be executed when reading file. Closes [#1702](https://github.com/Rdatatable/data.table/issues/1702).
* Column type guessing is improved by testing 100 rows at 10 points rather than 5 rows at 3 points. See point 3 of [convenience features of fread for small data](https://github.com/Rdatatable/data.table/wiki/Convenience-features-of-fread).
  1. Joins:
* Non-equi (or conditional) joins are now possible using the familiar `on=` syntax. Possible binary operators include `>=`, `>`, `<=`, `<` and `==`. For e.g., `X[Y, on=.(a, b>b)]` looks for `X.a == Y.a` first and within those matching rows for rows where`X.b > Y.b`, [#1452](https://github.com/Rdatatable/data.table/issues/1452).
* x's columns can be referred to in `j` using the prefix `x.` at all times. This is particularly useful when it is necessary to x's column that is *also a join column*, [#1615](https://github.com/Rdatatable/data.table/issues/1615). Also closes [#1705](https://github.com/Rdatatable/data.table/issues/1705) (thanks @dbetebenner) and [#1761](https://github.com/Rdatatable/data.table/issues/1761).
* `on=.()` syntax is now posible, e.g., `X[Y, on=.(x==a, y==b)]`, [#1257](https://github.com/Rdatatable/data.table/issues/1257). Thanks @dselivanov.
* Joins using `on=` accepts unnamed columns on ad hoc joins, e.g., X[.(5), on="b"] joins "b" from `X` to "V1" from `i`, partly closes [#1375](https://github.com/Rdatatable/data.table/issues/1375).
* When joining with `on=`, `X[Y, on=c(A="A", b="c")]` can be now specified as `X[Y, on=c("A", b="c")]`, fully closes [#1375](https://github.com/Rdatatable/data.table/issues/1375).
* `on=` joins now provides more friendly error messages when columns aren't found, [#1376](https://github.com/Rdatatable/data.table/issues/1376).
* Joins (and binary search based subsets) using `on=` argument now reuses existing (secondary) indices, [#1439](https://github.com/Rdatatable/data.table/issues/1439). Thanks @jangorecki.
  1. merge.data.table by default also checks for common key columns between the two data.tables before resulting in error when by or by.x, by.y arguments are not provided, #1517. Thanks @DavidArenburg.

  2. Fast set operations fsetdiff, fintersect, funion and fsetequal for data.tables are now implemented, #547.

  3. Added setDTthreads() and getDTthreads() to control the threads used in data.table functions that are now parallelized with OpenMP on all architectures including Windows (fwrite(), fsort() and subsetting). Extra code was required internally to ensure these control data.table only and not other packages using OpenMP. When data.table is used from the parallel package (e.g. mclapply as done by 3 CRAN and Bioconductor packages) data.table automatically switches down to one thread to avoid a deadlock/hang when OpenMP is used with fork(); #1745 and #1727. Thanks to Kontstantinos Tsardounis, Ramon Diaz-Uriarte and Jan Gorecki for testing before release and providing reproducible examples. After parallel::mclapply has finished, data.table reverts to the prior getDTthreads() state. Tests added which will therefore will run every day thanks to CRAN (limited to 2 threads on CRAN which is enough to test).

  4. GForce (See ?`datatable-optimize` for more):

* `dt[, .N, by=cols]` is optimised internally as well, [#1251](https://github.com/Rdatatable/data.table/issues/1251).
* is now also optimised for `median`. Partly addresses [#523](https://github.com/Rdatatable/data.table/issues/523). Check that issue for benchmarks.
* GForce kicks in along with subsets in `i` as well, e.g., `DT[x > 2, mean(y), by=z]`. Partly addresses [#971](https://github.com/Rdatatable/data.table/issues/971).
* GForce is optimised for `head(., 1)` and `tail(., 1)`, where `.` is a column name or `.SD`. Partly addresses [#523](https://github.com/Rdatatable/data.table/issues/523). Check the link for benchmarks.
* GForce is optimised for length-1 subsets, e.g., `.SD[2]`, `col[2]`. Partly addresses [#523](https://github.com/Rdatatable/data.table/issues/523).
* `var`, `sd` and `prod` are all GForce optimised for speed and memory. Partly addresses [#523](https://github.com/Rdatatable/data.table/issues/523). See that post for benchmarks.
  1. Reshaping:
* `dcast.data.table` now allows `drop = c(FALSE, TRUE)` and `drop = c(TRUE, FALSE)`. The former only fills all missing combinations of formula LHS, where as the latter fills only all missing combinations of formula RHS. Thanks to Ananda Mahto for [this SO post](http://stackoverflow.com/q/34830908/559784) and to Jaap for filing [#1512](https://github.com/Rdatatable/data.table/issues/1512).
* `melt.data.table` finds variables provided to `patterns()` when called from within user defined functions, [#1749](https://github.com/Rdatatable/data.table/issues/1749). Thanks to @kendonB for the report.
  1. We can now refer to the columns that are not mentioned in .SD / .SDcols in j as well. For example, DT[, .(sum(v1), lapply(.SD, mean)), by=grp, .SDcols=v2:v3] works as expected, #495. Thanks to @MattWeller for report and to others for linking various SO posts to be updated. Also closes #484.

  2. New functions inrange() and %inrange% are exported. It performs a range join making use of the recently implemented non-equi joins (#1452) #679. Also thanks to @DavidArenburg for #1819.

  3. %between% is vectorised which means we can now do: DT[x %between% list(y,z)] where y and z are vectors, #534. Thanks @MicheleCarriero for filing the issue and the idea.

  4. Most common use case for between(), i.e., lower and upper are length=1, is now implemented in C and parallelised. This results in ~7-10x speed improvement on vectors of length >= 1e6.

  5. Row subset operations of data.table is now parallelised with OpenMP, #1660. See the linked issue page for a rough benchmark on speedup.

  6. tstrsplit gains argument names, #1379. A character vector of column names can be provided as well. Thanks @franknarf1.

  7. tstrsplit gains argument keep which corresponds to the indices of list elements to return from the transposed list.

  8. rowid() and rowidv() - convenience functions for generating a unique row ids within each group, are implemented. rowid() is particularly useful along with dcast(). See ?rowid for more, #1353.

  9. rleid() gains prefix argument, similar to rowid().

  10. shift() understands and operates on list-of-list inputs as well, #1595. Thanks to @enfascination and to @chris for asking on SO.

  11. uniqueN gains na.rm argument, #1455.

  12. first() is now exported to return the first element of vectors, data.frames and data.tables.

  13. New split.data.table method. Faster, more flexible and consistent with data.frame method. Closes #1389. Now also properly preallocate columns, thanks @maverickg for reporting, closes #1908.

  14. rbindlist supports columns of type complex, #1659.

  15. Added second and minute extraction functions which, like extant hour/yday/week/etc, always return an integer, #874. Also added ISO 8601-consistent weeks in isoweek, #1765. Thanks to @bthieurmel and @STATWORX for the FRs and @MichaelChirico for the PRs.

  16. setnames accepts negative indices in old argument, #1443. Thanks @richierocks.

  17. by understands colA:colB syntax now, like .SDcols does, #1395. Thanks @franknarf1.

  18. data.table() function gains stringsAsFactors argument with default FALSE, #643. Thanks to @jangorecki for reviving this issue.

  19. print.data.table now warns when bit64 package isn't loaded but the data.table contains integer64 columns, #975. Thanks to @StephenMcInerney.

  20. New argument print.class for print.data.table allows for including column class under column names (as inspired by tbl_df in dplyr); default (adjustable via "datatable.print.class" option) is FALSE, the inherited behavior. Part of #1523; thanks to @MichaelChirico for the FR & PR.

  21. all.equal.data.table gains check.attributes, ignore.col.order, ignore.row.order and tolerance arguments.

  22. keyby= is now much faster by not doing not needed work; e.g. 25s down to 13s for a 1.5GB DT with 200m rows and 86m groups. With more groups or bigger data, larger speedup factors are possible. Please always use keyby= unless you really need by=. by= returns the groups in first appearance order and takes longer to do that. See #1880 for more info and please register your views there on changing the default.

  1. Now compiles and runs on IBM AIX gcc. Thanks to Vinh Nguyen for investigation and testing, #1351.

  2. as.ITime(NA) works as intended, #1354. Thanks @geneorama.

  3. last() dispatches xts::last() properly again, #1347. Thanks to @JoshuaUlrich for spotting and suggesting the fix.

  4. merge.data.table ignores names when by argument is a named vector, #1352. Thanks @sebastian-c.

  5. melt.data.table names value column correctly when patterns() of length=1 is provided to measure.vars(), #1346. Thanks @jaapwalhout.

  6. Fixed a rare case in melt.data.table not setting variable factor column properly when na.rm=TRUE, #1359. Thanks @mplatzer.

  7. dt[i, .SD] unlocks .SD and overallocates correctly now, #1341. Thanks @marc-outins.

  8. Querying a list column with get(), e.g., dt[, get("c")] is handled properly, #1212. Thanks @DavidArenburg.

  9. Grouping on empty data.table with list col in j works as expected, #1207. Thanks @jangorecki.

  10. Unnamed by/keyby expressions ensure now that the auto generated names are unique, #1334. Thanks @caneff.

  11. melt errors correctly when id.vars or measure.vars are negative values, #1372.

  12. merge.data.table always resets class to c("data.table", "data.frame") in result to be consistent with merge.data.frame, #1378. Thanks @ladida771.

  13. fread reads text input with empty newline but with just spaces properly, for e.g., fread('a,b\n1,2\n '), #1384. Thanks to @ladida771.

  14. fread with stringsAsFactors = TRUE no longer produces factors with NA as a factor level, #1408. Thanks to @DexGroves.

  15. test.data.table no longer raises warning if suggested packages are not available. Thanks to @jangorecki for PR #1403. Closes #1193.

  16. rleid() does not affect attributes of input vector, #1419. Thanks @JanGorecki.

  17. uniqueN() now handles NULL properly, #1429. Thanks @JanGorecki.

  18. GForce min and max functions handle NaN correctly, #1461. Thanks to @LyssBucks for asking on SO.

  19. Warnings on unable to detect column types from middle/last 5 lines are now moved to messages when verbose=TRUE. Closes #1124.

  20. fread converts columns to factor type when used along with colClasses argument, #721. Thanks @AmyMikhail.

  21. Auto indexing handles logical subset of factor column using numeric value properly, #1361. Thanks @mplatzer.

  22. as.data.table.xts handles single row xts object properly, #1484. Thanks Michael Smith and @jangorecki.

  23. data.table now solves the issue of mixed encodings by comparing character columns with marked encodings under UTF8 locale. This resolves issues #66, #69, #469 and #1293. Thanks to @StefanFritsch and @Arthur.

  24. rbindlist handles idcol construction correctly and more efficiently now (logic moved to C), #1432. Thanks to @franknarf1 and @Chris.

  25. CJ sorts correctly when duplicates are found in input values and sorted=TRUE, #1513. Thanks @alexdeng.

  26. Auto indexing returns order of subset properly when input data.table is already sorted, #1495. Thanks @huashan for the nice reproducible example.

  27. [.data.table handles column subsets based on conditions that result in NULL as list elements correctly, #1477. Thanks @MichaelChirico. Also thanks to @Max from DSR for spotting a bug as a result of this fix. Now fixed.

  28. Providing the first argument to .Call, for e.g., .Call("Crbindlist", ...) seems to result in "not resolved in current namespace" error. A potential fix is to simply remove the quotes like so many other calls in data.table. Potentially fixes #1467. Thanks to @rBatt, @rsaporta and @damienchallet.

  29. last function will now properly redirect method if xts is not installed or not attached on search path. Closes #1560.

  30. rbindlist (and rbind) works as expected when fill = TRUE and the first element of input list doesn't have columns present in other elements of the list, #1549. Thanks to @alexkowa.

  31. DT[, .(col), with=FALSE] now returns a meaningful error message, #1440. Thanks to @VasilyA for posting on SO.

  32. Fixed a segault in forder when elements of input list are not of same length, #1531. Thanks to @MichaelChirico.

  33. Reverted support of list-of-lists made in #1224 for consistency.

  34. Fixed an edge case in fread's fill argument, #1503. Thanks to @AnandaMahto.

  35. copy() overallocates properly when input is a list-of-data.tables, #1476. Thanks to @kimiylilammi and @AmitaiPerlstein for the report.

  36. fread() handles embedded double quotes in json fields as expected, #1164. Thanks @richardtessier.

  37. as.data.table.list handles list elements that are matrices/data.frames/data.tables properly, #833. Thanks to @talexand.

  38. data.table(), as.data.table() and [.data.table warn on POSIXlt type column and converts to POSIXct type. setDT() errors when input is list and any column is of type POSIXlt, #646. Thanks to @tdhock.

  39. roll argument handles -ve integer64 values correctly, #1405. Thanks @bryan4887. Also closes #1650, a segfault due to this fix. Thanks @Franknarf1 for filing the issue.

  40. Not join along with mult="first" and mult="last" is handled correctly, #1571.

  41. by=.EACHI works as expected along with mult="first" and mult="last", #1287 and #1271.

  42. Subsets using logical expressions in i (e.g. DT[someCol==3]) no longer return an unintended all-NA row when DT consists of a single row and someCol contains NA, fixing #1252. Thanks to @sergiizaskaleta for reporting. If i is the reserved symbol NA though (i.e. DT[NA]) it is still auto converted to DT[NA_integer_] so that a single NA row is returned as almost surely expected. For consistency with past behaviour and to save confusion when comparing to DT[c(NA,1)].

  43. setattr() catches logical input that points to R's global TRUE value and sets attributes on a copy instead, along with a warning, #1281. Thanks to @tdeenes.

  44. fread respects order of columns provided to argument select in result, and also warns if the column(s) provided is not present, #1445.

  45. DT[, .BY, by=x] and other variants of adding a column using .BY are now handled correctly, #1270.

  46. as.data.table.data.table() method checks and restores over-allocation, #473.

  47. When the number of rows read are less than the number of guessed rows (or allocated), fread() doesn't warn anymore; rather restricts to a verbose message, #1116 and #1239. Thanks to @slowteetoe and @hshipper.

  48. fread() throws an error if input is a directory, #989. Thanks @vlsi.

  49. UTF8 BOM header is excluded properly in fread(), #1087 and #1465. Thanks to @nigmastar and @MichaelChirico.

  50. Joins using on= retains (and discards) keys properly, #1268. Thanks @DouglasClark for this SO post that helped discover the issue.

  51. Secondary keys are properly removed when those columns get updated, #1479. Thanks @fabiangehring for the report, and also @ChristK for the MRE.

  52. dcast no longer errors on tables with duplicate columns that are unused in the call, #1654. Thanks @MichaelChirico for FR&PR.

  53. fread won't use wget for file:// input, #1668; thanks @MichaelChirico for FR&PR.

  54. chmatch() handles nomatch = integer(0) properly, #1672.

  55. dimnames.data.table no longer errors in data.table-unaware environments when a data.table has, e.g., been churned through some dplyr functions and acquired extra classes, #1678. Thanks Daisy Lee on SO for pointing this out and @MichaelChirico for the fix.

  56. fread() did not respect encoding on header column. Now fixed, #1680. Thanks @nachti.

  57. as.data.table's data.table method returns a copy as it should, #1681.

  58. Grouped update operations, e.g., DT[, y := val, by=x] where val is an unsupported type errors without adding an unnamed column, #1676. Thanks @wligtenberg.

  59. Handled use of .I in some GForce operations, #1683. Thanks gibbz00 from SO and @franknarf1 for reporting and @MichaelChirico for the PR.

  60. Added +.IDate method so that IDate + integer retains the IDate class, #1528; thanks @MichaelChirico for FR&PR. Similarly, added -.IDate so that IDate - IDate returns a plain integer rather than difftime.

  61. Radix ordering an integer vector containing INTMAX (2147483647) with decreasing=TRUE and na.last=FALSE failed ASAN check and seg faulted some systems. As reported for base R #16925 whose new code comes from data.table. Simplified code, added test and proposed change to base R.

  62. Fixed test in onAttach() for when Packaged field is missing from DESCRIPTION, #1706; thanks @restonslacker for BR&PR.

  63. Adding missing factor levels are handled correctly in case of NAs. This affected a case of join+update operation as shown in #1718. Thanks to @daniellemccool.

  64. foverlaps now raise a meaningful error for duplicate column names, closes #1730. Thanks @rodonn.

  65. na.omit and unique methods now removes indices, closes #1734 and #1760. Thanks @m-dz and @fc9.30.

  66. List of data.tables with custom class is printed properly, #1758. Thanks @fruce-ki.

  67. uniqueN handles na.rm=TRUE argument on sorted inputs correctly, #1771. Thanks @ywhuofu.

  68. get() / mget() play nicely with .SD / .SDcols, #1744. Thanks @franknarf1.

  69. Joins on integer64 columns assigns NA correctly for no matching rows, #1385 and partly #1459. Thanks @dlithio and @abielr.

  70. Added as.IDate.POSIXct to prevent loss of timezone information, #1498. Thanks @dougedmunds for reporting and @MichaelChirico for the investigating & fixing.

  71. Retaining / removing keys is handled better when join is performed on non-key columns using on argument, #1766, #1704 and #1823. Thanks @mllg, @DavidArenburg and @mllg.

  72. rbind for data.tables now coerces non-list inputs to data.tables first before calling rbindlist so that binding list of data.tables and matrices work as expected to be consistent with base's rbind, #1626. Thanks @ems for reporting here on SO.

  73. Subassigning a factor column with NA works as expected. Also, the warning message on coercion is suppressed when RHS is singleton NA, #1740. Thanks @Zus.

  74. Joins on key columns in the presence of on= argument were slightly slower as it was unnecesarily running a check to ensure orderedness. This is now fixed, #1825. Thanks @sz-cgt. See that post for updated benchmark.

  75. keyby= now runs j in the order that the groups appear in the sorted result rather than first appearance order, #606. This only makes a difference in very rare usage where j does something depending on an earlier group's result, perhaps by using <<-. If j is required to be run in first appearance order, then use by= whose behaviour is unchanged. Now we have this option. No existing tests affected. New tests added.

  76. := verbose messages have been corrected and improved, #1808. Thanks to @franknarf1 for reproducible examples. Tests added.

  77. DT[order(colA,na.last=NA)] on a 2-row DT with one NA in colA and na.last=NA (meaning to remove NA) could return a randomly wrong result due to using uninitialized memory. Tests added.

  78. fread is now consistent to read.table on colClasses vector containing NA, also fixes mixed character and factor in colClasses vector. Closes #1910.

  1. Updated error message on invalid joins to reflect the new on= syntax, #1368. Thanks @MichaelChirico.

  2. Fixed test 842 to account for gdata::last as well, #1402. Thanks @JanGorecki.

  3. Fixed tests for fread 1378.2 and 1378.3 with showProgress = FALSE, closes #1397. Thanks to @JanGorecki for the PR.

  4. Worked around auto index error in v1.9.6 to account for indices created with v1.9.4, #1396. Thanks @GRandom.

  5. test.data.table gets new argument silent, if set to TRUE then it will not raise exception but returns TRUE/FALSE based on the test results.

  6. dim.data.table is now implemented in C. Thanks to Andrey Riabushenko.

  7. Better fix to fread's check.names argument using make.names(), #1027. Thanks to @DavidArenberg for spotting the issue with the previous fix using make.unique().

  8. Fixed explanation of skip argument in ?fread as spotted by @aushev, #1425.

  9. Run install_name_tool when building on OS X to ensure that the install name for datatable.so matches its filename. Fixes #1144. Thanks to @chenghlee for the PR.

  10. Updated documentation of i in [.data.table to emphasize the emergence of the new on option as an alternative to keyed joins, #1488. Thanks @MichaelChirico.

  11. Improvements and fixes to ?like #1515. Thanks to @MichaelChirico for the PR.

  12. Several improvements and fixes to ?between #1521. Thanks @MichaelChirico for the PR.

  13. ?shift.Rd is fixed so that it does not get misconstrued to be in a time series sense. Closes #1530. Thanks to @pstoyanov.

  14. ?truelength.Rd is fixed to reflect that over-allocation happens on data.tables loaded from disk only during column additions and not deletions, #1536. Thanks to @Roland and @rajkrpan.

  15. Added \n to message displayed in melt.data.table when duplicate names are found, #1538. Thanks @Franknarf1.

  16. merge.data.table will raise warning if any of data.tables to join has 0 columns. Closes #597.

  17. Travis-CI will now automatically deploy package to drat repository hosted on data.table@gh-pages branch allowing to install latest devel from source via install.packages("data.table", repos = "https://Rdatatable.github.io/data.table", type = "source"). Closes #1505.

  18. Dependency on chron package has been changed to suggested. Closes #1558.

  19. Rnw vignettes are converted to Rmd. The 10 minute quick introduction Rnw vignette has been removed, since almost all of its contents are consolidated into the new intro Rmd vignette. Thanks to @MichaelChirico and @jangorecki.

A quick tour of data.table HTML vignette is in the works in the spirit of the previous 10 minute quick intro PDF guide.

  1. row.names argument to print.data.table can now be changed by default via options("datatable.print.rownames") (TRUE by default, the inherited standard), #1097. Thanks to @smcinerney for the suggestion and @MichaelChirico for the PR.

  2. data.tables with NULL or blank column names now print with blank column names, #545, with minor revision to #97. Thanks to @arunsrinivasan for reporting and @MichaelChirico for the PR.

  3. Added a FAQ entry for the new update to := which sometimes doesn't print the result on the first time, #939.

  4. Added Note section and examples to ?":=" for #905.

  5. Fixed example in ?as.data.table.Rd, #1576. Thanks @MichaelChirico.

  6. Fixed an edge case and added tests for columns of type function, #518.

  7. data.table's dependency has been moved forward from R 2.14.1 to R 3.0.0 (Apr 2013; i.e. 3 years old). We keep this dependency as old as possible for as long as possible as requested by users in managed environments. This bump allows data.table internals to use paste0() for the first time and also allows fsort() to accept vectors of length over 2 billion items. Before release to CRAN our procedures include running the test suite using this stated dependency.

  8. New option options(datatable.use.index = TRUE) (default) gives better control over usage of indices, when combined with options(datatable.auto.index = FALSE) it allows to use only indices created manually with setindex or setindexv. Closes #1422.

  9. The default number of over-allocated spare column pointer slots has been increased from 64 to 1024. The wasted memory overhead (if never used) is insignificant (0.008 MB). The advantage is that adding a large number of columns by reference using := or set() inside a loop will not now saturate as quickly and need reallocating. An alleviation to issue #1633. See ?alloc.col for how to change this default yourself. Accordingly, the warning 'attempt to reduce allocation has been ignored' has been downgraded to a message in verbose mode. That typically occurs when using (not recommended) [<- and $<- methods on data.table. The n= argument to alloc.col() is now simply the number of spare column slots to over-allocate (on creation and reallocation). An expression using ncol(DT) is still ok but now deprecated.

  10. ?IDateTime now makes clear that wday, yday and month are all 1- (not 0- as in POSIXlt) based, #1658; thanks @MichaelChirico.

  11. Fixed misleading documentation of ?uniqueN, #1746. Thanks @SymbolixAU.

  12. melt.data.table restricts column names printed during warning messages to a maximum of five, #1752. Thanks @franknarf1.

  13. data.table's setNumericRounding has a default value of 0, which means ordering, joining and grouping of numeric values will be done at full precision by default. Handles #1642, #1728, #1463, #485.

  14. Subsets with S4 objects in i are now faster, #1438. Thanks @DCEmilberg.

  15. When formula RHS is . and multiple functions are provided to fun.aggregate, column names of the cast data.table columns don't have the . in them, as it doesn't add any useful information really, #1821. Thanks @franknarf1.

  16. Function names are added to column names on cast data.tables only when more than one function is provided, #1810. Thanks @franknarf1.

  17. The option datatable.old.bywithoutby to restore the old default has been removed. As warned 2 years ago in release notes and explicitly warned about for 1 year when used. Search down this file for the text 'bywithoutby' to see previous notes on this topic.

  18. Using with=FALSE together with := was deprecated in v1.9.4 released 2 years ago (Oct 2014). As warned then in release notes (see below) this is now a warning with advice to wrap the LHS of := with parenthesis; e.g. myCols=c("colA","colB"); DT[,(myCols):=1]. In the next release, this warning message will be an error message.

  19. Using nomatch together with := now warns that it is ignored.

  20. Logical i is no longer recycled. Instead an error message if it isn't either length 1 or nrow(DT). This was hiding more bugs than was worth the rare convenience. The error message suggests to recycle explcitly; i.e. DT[rep(<logical>,length=.N),...].

  21. Thanks to Mark Landry and Michael Chirico for finding and reporting a problem in dev before release with auto with=FALSE (item 3 above) when j starts with with ! or -, #1864. Fixed and tests added.

  22. Following latest recommended testthat practices and to avoid a warning that it now issues, inst/tests/testthat has been moved to /tests/testthat. This means that testthat tests won't be installed for use by users by default and that test_package("data.table") will now fail with error No matching test file in dir and also a warning Placing tests in inst/tests/ is deprecated. Please use tests/testthat/ instead. (That warning seems to be misleading since we already have made that move.) To install testthat tests (and this applies to all packages using testthat not just data.table) you need to follow the deleted instructions in testthat's README; i.e., reinstall data.table either with --install-tests passed to R CMD INSTALL or INSTALL_opts = "--install-tests" passed to install.packages(). After that, test_package("data.table") will work. However, the main test suite of data.table (5,000+ tests) doesn't use testthat at all. Those tests are always installed so that test.data.table() can always be run by users at any time to confirm your installation on your platform is working correctly. Sometimes when supporting you, you may be asked to run test.data.table() and provide the output. Particularly now that data.table uses OpenMP. The file /tests/tests.R (which just calls test.data.table()) has been renamed to /tests/main.R to make this clearer to those looking at the GitHub repository and a comment has been added to /tests/main.R pointing to /inst/tests/tests.Rraw where those tests live. Some of these tests test data.table's compability with other packages and that is the reason those packages are listed in DESCRIPTION:Suggests. If you don't have some of those packages installed, test.data.table() will print output that it has skipped tests of compatibility with those packages. On CRAN all Suggests packages are available and data.table's tests of compatibility with them are tested by CRAN every day.

  23. The license field is changed from "GPL (>= 2)" to "GPL-3 | file LICENSE" due to independent communication from two users of data.table at Google. The lack of an explicit license file was preventing them from contributing patches to data.table. Further, Google lawyers require the full text of the license and not a URL to the license. Since this requirement appears to require the choice of one license, we opted for GPL-3 and we checked the GPL-3 is fine by Google for them to use and contribute to. Accordingly, data.table's LICENSE file is an exact duplicate copy of the canonical GPL-3.

  24. Thanks to @rrichmond for finding and reporting a regression in dev before release with roll not respecting fractions in type double, #1904. For example dates like zoo::as.yearmon("2016-11") which is stored as double value 2016.833. Fixed and test added.

  1. fread

    • passes showProgress=FALSE through to download.file() (as quiet=TRUE). Thanks to a pull request from Karl Broman and Richard Scriven for filing the issue, #741.
    • accepts dec=',' (and other non-'.' decimal separators), #917. A new paragraph has been added to ?fread. On Windows this should just-work. On Unix it may just-work but if not you will need to read the paragraph for an extra step. In case it somehow breaks dec='.', this new feature can be turned off with options(datatable.fread.dec.experiment=FALSE).
    • Implemented stringsAsFactors argument for fread(). When TRUE, character columns are converted to factors. Default is FALSE. Thanks to Artem Klevtsov for filing #501, and to @hmi2015 for this SO post.
    • gains check.names argument, with default value FALSE. When TRUE, it uses the base function make.unique() to ensure that the column names of the data.table read in are all unique. Thanks to David Arenburg for filing #1027.
    • gains encoding argument. Acceptable values are "unknown", "UTF-8" and "Latin-1" with default value of "unknown". Closes #563. Thanks to @BenMarwick for the original report and to the many requests from others, and Q on SO.
    • gains col.names argument, and is similar to base::read.table(). Closes #768. Thanks to @dardesta for filing the FR.
  2. DT[column == value] no longer recycles value except in the length 1 case (when it still uses DT's key or an automatic secondary key, as introduced in v1.9.4). If length(value)==length(column) then it works element-wise as standard in R. Otherwise, a length error is issued to avoid common user errors. DT[column %in% values] still uses DT's key (or an an automatic secondary key) as before. Automatic indexing (i.e., optimization of == and %in%) may still be turned off with options(datatable.auto.index=FALSE).

  3. na.omit method for data.table is rewritten in C, for speed. It's ~11x faster on bigger data; see examples under ?na.omit. It also gains two additional arguments a) cols accepts column names (or numbers) on which to check for missing values. 2) invert when TRUE returns the rows with any missing values instead. Thanks to the suggestion and PR from @matthieugomez.

  4. New function shift() implements fast lead/lag of vector, list, data.frames or data.tables. It takes a type argument which can be either "lag" (default) or "lead". It enables very convenient usage along with := or set(). For example: DT[, (cols) := shift(.SD, 1L), by=id]. Please have a look at ?shift for more info.

  5. frank() is now implemented. It's much faster than base::rank and does more. It accepts vectors, lists with all elements of equal lengths, data.frames and data.tables, and optionally takes a cols argument. In addition to implementing all the ties.method methods available from base::rank, it also implements dense rank. It is also capable of calculating ranks by ordering column(s) in ascending or descending order. See ?frank for more. Closes #760 and #771

  6. rleid(), a convenience function for generating a run-length type id column to be used in grouping operations is now implemented. Closes #686. Check ?rleid examples section for usage scenarios.

  7. Efficient convertion of xts to data.table. Closes #882. Check examples in ?as.xts.data.table and ?as.data.table.xts. Thanks to @jangorecki for the PR.

  8. rbindlist gains idcol argument which can be used to generate an index column. If idcol=TRUE, the column is automatically named .id. Instead you can also provide a column name directly. If the input list has no names, indices are automatically generated. Closes #591. Also thanks to @KevinUshey for filing #356.

  9. A new helper function uniqueN is now implemented. It is equivalent to length(unique(x)) but much faster. It handles atomic vectors, lists, data.frames and data.tables as input and returns the number of unique rows. Closes #884. Gains by argument. Closes #1080. Closes #1224. Thanks to @DavidArenburg, @kevinmistry and @jangorecki.

  10. Implemented transpose() to transpose a list and tstrsplit which is a wrapper for transpose(strsplit(...)). This is particularly useful in scenarios where a column has to be split and the resulting list has to be assigned to multiple columns. See ?transpose and ?tstrsplit, #1025 and #1026 for usage scenarios. Closes both #1025 and #1026 issues.

* Implemented `type.convert` as suggested by Richard Scriven. Closes [#1094](https://github.com/Rdatatable/data.table/issues/1094).
  1. melt.data.table

    • can now melt into multiple columns by providing a list of columns to measure.vars argument. Closes #828. Thanks to Ananda Mahto for the extended email discussions and ideas on generating the variable column.
    • also retains attributes wherever possible. Closes #702 and #993. Thanks to @richierocks for the report.
    • Added patterns.Rd. Closes #1294. Thanks to @MichaelChirico.
  2. .SDcols

    • understands ! now, i.e., DT[, .SD, .SDcols=!"a"] now works, and is equivalent to DT[, .SD, .SDcols = -c("a")]. Closes #1066.
    • accepts logical vectors as well. If length is smaller than number of columns, the vector is recycled. Closes #1060. Thanks to @StefanFritsch.
  3. dcast can now:

    • cast multiple value.var columns simultaneously. Closes #739.
    • accept multiple functions under fun.aggregate. Closes #716.
    • supports optional column prefixes as mentioned under this SO post. Closes #862. Thanks to @JohnAndrews.
    • works with undefined variables directly in formula. Closes #1037. Thanks to @DavidArenburg for the MRE.
    • Naming conventions on multiple columns changed according to #1153. Thanks to @MichaelChirico for the FR.
    • also has a sep argument with default _ for backwards compatibility. #1210. Thanks to @dbetebenner for the FR.
  4. .SDcols and with=FALSE understand colA:colB form now. That is, DT[, lapply(.SD, sum), by=V1, .SDcols=V4:V6] and DT[, V5:V7, with=FALSE] works as intended. This is quite useful for interactive use. Closes #748 and #1216. Thanks to @carbonmetrics, @jangorecki and @mtennekes.

  5. setcolorder() and setorder() work with data.frames too. Closes #1018.

  6. as.data.table.* and setDT argument keep.rownames can take a column name as well. When keep.rownames=TRUE, the column will still automatically named rn. Closes #575.

  7. setDT gains a key argument so that setDT(X, key="a") would convert X to a data.table by reference and key by the columns specified. Closes #1121.

  8. setDF also converts list of equal length to data.frame by reference now. Closes #1132.

  9. CJ gains logical unique argument with default FALSE. If TRUE, unique values of vectors are automatically computed and used. This is convenient, for example, DT[CJ(a, b, c, unique=TRUE)] instead of doing DT[CJ(unique(a), unique(b), unique(c))]. Ultimately, unique = TRUE will be default. Closes #1148.

  10. on= syntax: data.tables can join now without having to set keys by using the new on argument. For example: DT1[DT2, on=c(x = "y")] would join column 'y' of DT2 with 'x' of DT1. DT1[DT2, on="y"] would join on column 'y' on both data.tables. Closes #1130 partly.

  11. merge.data.table gains arguments by.x and by.y. Closes #637 and #1130. No copies are made even when the specified columns aren't key columns in data.tables, and therefore much more fast and memory efficient. Thanks to @blasern for the initial PRs. Also gains logical argument sort (like base R). Closes #1282.

  12. setDF() gains rownames argument for ready conversion to a data.frame with user-specified rows. Closes #1320. Thanks to @MichaelChirico for the FR and PR.

  13. print.data.table gains quote argument (defaul=FALSE). This option surrounds all printed elements with quotes, helps make whitespace(s) more evident. Closes #1177; thanks to @MichaelChirico for the PR.

  14. [.data.table now accepts single column numeric matrix in i argument the same way as data.frame. Closes #826. Thanks to @jangorecki for the PR.

  15. setDT() gains check.names argument paralleling that of fread, data.table, and base functionality, allowing poorly declared objects to be converted to tidy data.tables by reference. Closes #1338; thanks to @MichaelChirico for the FR/PR.

  1. if (TRUE) DT[,LHS:=RHS] no longer prints, #869 and #1122. Tests added. To get this to work we've had to live with one downside: if a := is used inside a function with no DT[] before the end of the function, then the next time DT or print(DT) is typed at the prompt, nothing will be printed. A repeated DT or print(DT) will print. To avoid this: include a DT[] after the last := in your function. If that is not possible (e.g., it's not a function you can change) then DT[] at the prompt is guaranteed to print. As before, adding an extra [] on the end of a := query is a recommended idiom to update and then print; e.g. > DT[,foo:=3L][]. Thanks to Jureiss and Jan Gorecki for reporting.

  2. DT[FALSE,LHS:=RHS] no longer prints either, #887. Thanks to Jureiss for reporting.

  3. := no longer prints in knitr for consistency with behaviour at the prompt, #505. Output of a test knit("knitr.Rmd") is now in data.table's unit tests. Thanks to Corone for the illustrated report.

  4. knitr::kable() works again without needing to upgrade from knitr v1.6 to v1.7, #809. Packages which evaluate user code and don't wish to import data.table need to be added to data.table:::cedta.pkgEvalsUserCode and now only the eval part is made data.table-aware (the rest of such package's code is left data.table-unaware). data.table:::cedta.override is now empty and will be deprecated if no need for it arises. Thanks to badbye and Stephanie Locke for reporting.

  5. fread():

    • doubled quotes ("") inside quoted fields including if immediately followed by an embedded newline. Thanks to James Sams for reporting, #489.
    • quoted fields with embedded newlines in the lines used to detect types, #810. Thanks to Vladimir Sitnikov for the scrambled data file which is now included in the test suite.
    • when detecting types in the middle and end of the file, if the jump lands inside a quoted field with (possibly many) embedded newlines, this is now detected.
    • if the file doesn't exist the error message is clearer (#486)
    • system commands are now checked to contain at least one space
    • sep="." now works (when dec!="."), #502. Thanks to Ananda Mahto for reporting.
    • better error message if quoted field is missing an end quote, #802. Thanks to Vladimir Sitnikov for the sample file which is now included in the test suite.
    • providing sep which is not present in the file now reads as if sep="\n" rather than 'sep not found', #738. Thanks to Adam Kennedy for explaining the use-case.
    • seg fault with errors over 1,000 characters (when long lines are included) is fixed, #802. Thanks again to Vladimir Sitnikov.
    • Missing integer64 values are properly assigned NAs. Closes #488. Thanks to @PeterStoyanov and @richierocks for the report.
    • Column headers with empty strings aren't skipped anymore. Closes #483. Thanks to @RobyJoehanes and @kforner.
    • Detects separator correctly when commas also exist in text fields. Closes #923. Thanks to @raymondben for the report.
    • NA values in NA inflated file are read properly. Closes #737. Thanks to Adam Kennedy.
    • correctly handles na.strings argument for all types of columns - it detect possible NA values without coercion to character, like in base read.table. fixes #504. Thanks to @dselivanov for the PR. Also closes #1314, which closes this issue completely, i.e., na.strings = c("-999", "FALSE") etc. also work.
    • deals with quotes more robustly. When reading quoted fields fail, it re-attemps to read the field as if it wasn't quoted. This helps read in those fields that might have unbalanced quotes without erroring immediately, thereby closing issues #568, #1256, #1077, #1079 and #1095. Thanks to @Synergist, @daroczig, @geotheory and @rsaporta for the reports.
    • gains argument strip.white which is TRUE by default (unlike base::read.table). All unquoted columns' leading and trailing white spaces are automatically removed. If \code{FALSE}, only trailing spaces of header is removed. Closes #1113, #1035, #1000, #785, #529 and #956. Thanks to @dmenne, @dpastoor, @GHarmata, @gkalnytskyi, @renqian, @MatthewForrest, @fxi and @heraldb.
    • doesn't warn about empty lines when 'nrow' argument is specified and that many rows are read properly. Thanks to @richierocks for the report. Closes #1330.
    • doesn't error/warn about not being able to read last 5 lines when 'nrow' argument is specified. Thanks to @robbig2871. Closes #773.
  6. Auto indexing:

    • DT[colA == max(colA)] now works again without needing options(datatable.auto.index=FALSE). Thanks to Jan Gorecki and kaybenleroll, #858. Test added.
    • DT[colA %in% c("id1","id2","id2","id3")] now ignores the RHS duplicates (as before, consistent with base R) without needing options(datatable.auto.index=FALSE). Thanks to Dayne Filer for reporting.
    • If DT contains a column class (happens to be a reserved attribute name in R) then DT[class=='a'] now works again without needing options(datatable.auto.index=FALSE). Thanks to sunnyghkm for reporting, #871.
    • := and set* now drop secondary keys (new in v1.9.4) so that DT[x==y] works again after a := or set* without needing options(datatable.auto.index=FALSE). Only setkey() was dropping secondary keys correctly. 23 tests added. Thanks to user36312 for reporting, #885.
    • Automatic indices are not created on .SD so that dt[, .SD[b == "B"], by=a] works correctly. Fixes #958. Thanks to @azag0 for the nice reproducible example.
    • i-operations resulting in 0-length rows ignore j on subsets using auto indexing. Closes #1001. Thanks to @Gsee.
    • POSIXct type columns work as expected with auto indexing. Closes #955. Thanks to @GSee for the minimal report.
    • Auto indexing with ! operator, for e.g., DT[!x == 1] works as intended. Closes #932. Thanks to @matthieugomez for the minimal example.
    • While fixing #932, issues on subsetting NA were also spotted and fixed, for e.g., DT[x==NA] or DT[!x==NA].
    • Works fine when RHS is of list type - quite unusual operation but could happen. Closes #961. Thanks to @Gsee for the minimal report.
    • Auto indexing errored in some cases when LHS and RHS were not of same type. This is fixed now. Closes #957. Thanks to @GSee for the minimal report.
    • DT[x == 2.5] where x is integer type resulted in val being coerced to integer (for binary search) and therefore returned incorrect result. This is now identified using the function isReallyReal() and if so, auto indexing is turned off. Closes #1050.
    • Auto indexing errored during DT[x %in% val] when val has some values not present in x. Closes #1072. Thanks to @CarlosCinelli for asking on StackOverflow.
  7. as.data.table.list with list input having 0-length items, e.g. x = list(a=integer(0), b=3:4). as.data.table(x) recycles item a with NAs to fit the length of the longer column b (length=2), as before now, but with an additional warning message that the item has been recycled with NA. Closes #847. Thanks to @tvinodr for the report. This was a regression from 1.9.2.

  8. DT[i, j] when i returns all FALSE and j contains some length-0 values (ex: integer(0)) now returns an empty data.table as it should. Closes #758 and #813. Thanks to @tunaaa and @nigmastar for the nice reproducible reports.

  9. allow.cartesian is ignored during joins when:

    • i has no duplicates and mult="all". Closes #742. Thanks to @nigmastar for the report.
    • assigning by reference, i.e., j has :=. Closes #800. Thanks to @matthieugomez for the report.

In both these cases (and during a not-join which was already fixed in 1.9.4), allow.cartesian can be safely ignored.

  1. names<-.data.table works as intended on data.table unaware packages with Rv3.1.0+. Closes #476 and #825. Thanks to ezbentley for reporting here on SO and to @narrenfrei.

  2. .EACHI is now an exported symbol (just like .SD,.N,.I,.GRP and .BY already were) so that packages using data.table and .EACHI pass R CMD check with no NOTE that this symbol is undefined. Thanks to Matt Bannert for highlighting.

  3. Some optimisations of .SD in j was done in 1.9.4, refer to #735. Due to an oversight, j-expressions of the form c(lapply(.SD, ...), list(...)) were optimised improperly. This is now fixed. Thanks to @mmeierer for filing #861.

  4. j-expressions in DT[, col := x$y()] (or) DT[, col := x[[1]]()] are now (re)constructed properly. Thanks to @ihaddad-md for reporting. Closes #774.

  5. format.ITime now handles negative values properly. Closes #811. Thanks to @StefanFritsch for the report along with the fix!

  6. Compatibility with big endian machines (e.g., SPARC and PowerPC) is restored. Most Windows, Linux and Mac systems are little endian; type .Platform$endian to confirm. Thanks to Gerhard Nachtmann for reporting and the QEMU project for their PowerPC emulator.

  7. DT[, LHS := RHS] with RHS is of the form eval(parse(text = foo[1])) referring to columns in DT is now handled properly. Closes #880. Thanks to tyner.

  8. subset handles extracting duplicate columns in consistency with data.table's rule - if a column name is duplicated, then accessing that column using column number should return that column, whereas accessing by column name (due to ambiguity) will always extract the first column. Closes #891. Thanks to @jjzz.

  9. rbindlist handles combining levels of data.tables with both ordered and unordered factor columns properly. Closes #899. Thanks to @ChristK.

  10. Updating .SD by reference using set also errors appropriately now; similar to :=. Closes #927. Thanks to @jrowen for the minimal example.

  11. X[Y, .N] returned the same result as X[Y, .N, nomatch=0L]) when Y contained rows that has no matches in X. Fixed now. Closes #963. Thanks to this SO post from @Alex which helped discover the bug.

  12. data.table::dcast handles levels in factor columns properly when drop = FALSE. Closes #893. Thanks to @matthieugomez for the great minimal example.

  13. [.data.table subsets complex and raw type objects again. Thanks to @richierocks for the nice minimal example. Closes #982.

  14. Fixed a bug in the internal optimisation of j-expression with more than one lapply(.SD, function(..) ..) as illustrated here on SO. Closes #985. Thanks to @jadaliha for the report and to @BrodieG for the debugging on SO.

  15. mget fetches columns from the default environment .SD when called from within the frame of DT. That is, DT[, mget(cols)], DT[, lapply(mget(cols), sum), by=.] etc.. work as intended. Thanks to @Roland for filing this issue. Closes #994.

  16. foverlaps() did not find overlapping intervals correctly on numeric ranges in a special case where both start and end intervals had 0.0. This is now fixed. Thanks to @tdhock for the reproducible example. Closes #1006 partly.

  17. When performing rolling joins, keys are set only when we can be absolutely sure. Closes #1010, which explains cases where keys should not be retained.

  18. Rolling joins with -Inf and Inf are handled properly. Closes #1007. Thanks to @tdhock for filing #1006 which lead to the discovery of this issue.

  19. Overlapping range joins with -Inf and Inf and 0.0 in them are handled properly now. Closes #1006. Thanks to @tdhock for filing the issue with a nice reproducible example.

  20. Fixed two segfaults in shift() when number of rows in x is lesser than value for n. Closes #1009 and #1014. Thanks to @jangorecki and @ashinm for the reproducible reports.

  21. Attributes are preserved for sum() and mean() when fast internal (GForce) implementations are used. Closes #1023. Thanks to @DavidArenburg for the nice reproducible example.

  22. lapply(l, setDT) is handled properly now; over-allocation isn't lost. Similarly, for (i in 1:k) setDT(l[[i]]) is handled properly as well. Closes #480.

  23. rbindlist stack imbalance on all NULL list elements is now fixed. Closes #980. Thanks to @ttuggle.

  24. List columns can be assigned to columns of factor type by reference. Closes #936. Thanks to @richierocks for the minimal example.

  25. After setting the datatable.alloccol option, creating a data.table with more than the set truelength resulted in error or segfault. This is now fixed. Closes #970. Thanks to @caneff for the nice minimal example.

  26. Update by reference using := after loading from disk where the data.table exists within a local environment now works as intended. Closes #479. Thanks to @ChongWang for the minimal reproducible example.

  27. Issues on merges involving factor columns with NA and merging factor with character type with non-identical levels are both fixed. Closes #499 and #945. Thanks to @AbielReinhart and @stewbasic for the minimal examples.

  28. as.data.table(ll) returned a data.table with 0-rows when the first element of the list has 0-length, for e.g., ll = list(NULL, 1:2, 3:4). This is now fixed by removing those 0-length elements. Closes #842. Thanks to @Rick for the nice minimal example.

  29. as.datat.able.factor redirects to as.data.table.matrix when input is a matrix, but also of type factor. Closes #868. Thanks to @mgahan for the example.

  30. setattr now returns an error when trying to set data.table and/or data.frame as class to a non-list type object (ex: matrix). Closes #832. Thanks to @Rick for the minimal example.

  31. data.table(table) works as expected. Closes #1043. Thanks to @rnso for the SO post.

  32. Joins and binary search based subsets of the form x[i] where x's key column is integer and i a logical column threw an error before. This is now fixed by converting the logical column to integer type and then performing the join, so that it works as expected.

  33. When by expression is, for example, by = x %% 2, data.table tries to automatically extracts meaningful column names from the expression. In this case it would be x. However, if the j-expression also contains x, for example, DT[, last(x), by= x %% 2], the original x got masked by the expression in by. This is now fixed; by-expressions are not simplified in column names for these cases. Closes #497. Thanks to @GSee for the report.

  34. rbindlist now errors when columns have non-identical class attributes and are not factors, e.g., binding column of class Date with POSIXct. Previously this returned incorrect results. Closes #705. Thanks to @ecoRoland for the minimal report.

  35. Fixed a segfault in melt.data.table when measure.vars have duplicate names. Closes #1055. Thanks to @ChristK for the minimal report.

  36. Fixed another segfault in melt.data.table issue that was caught due to issue in Windows. Closes #1059. Thanks again to @ChristK for the minimal report.

  37. DT[rows, newcol := NULL] resulted in a segfault on the next assignment by reference. Closes #1082. Thanks to @stevenbagley for the MRE.

  38. as.matrix(DT) handles cases where DT contains both numeric and logical columns correctly (doesn't coerce to character columns anymore). Closes #1083. Thanks to @bramvisser for the SO post.

  39. Coercion is handled properly on subsets/joins on integer64 key columns. Closes #1108. Thanks to @vspinu.

  40. setDT() and as.data.table() both strip all classes preceding data.table/data.frame, to be consistent with base R. Closes #1078 and #1128. Thanks to Jan and @helix123 for the reports.

  41. setattr(x, 'levels', value) handles duplicate levels in value appropriately. Thanks to Jeffrey Horner for pointing it out here. Closes #1142.

  42. x[J(vals), .N, nomatch=0L] also included no matches in result, #1074. And x[J(...), col := val, nomatch=0L] returned a warning with incorrect results when join resulted in no matches as well, even though nomatch=0L should have no effect in :=, #1092. Both issues are fixed now. Thanks to @riabusan and @cguill95 for #1092.

  43. .data.table.locked attributes set to NULL in internal function subsetDT. Closes #1154. Thanks to @jangorecki.

  44. Internal function fastmean() retains column attributes. Closes #1160. Thanks to @renkun-ken.

  45. Using .N in i, for e.g., DT[, head(.SD, 3)[1:(.N-1L)]] accessed incorrect value of .N. This is now fixed. Closes #1145. Thanks to @claytonstanley.

  46. setDT handles key= argument properly when input is already a data.table. Closes #1169. Thanks to @DavidArenburg for the PR.

  47. Key is retained properly when joining on factor type columns. Closes #477. Thanks to @nachti for the report.

  48. Over-allocated memory is released more robustly thanks to Karl Millar's investigation and suggested fix.

  49. DT[TRUE, colA:=colA*2] no longer churns through 4 unnecessary allocations as large as one column. This was caused by i=TRUE being recycled. Thanks to Nathan Kurz for reporting and investigating. Added provided test to test suite. Only a single vector is allocated now for the RHS (colA*2). Closes #1249.

  50. Thanks to @and3k for the excellent bug report #1258. This was a result of shallow copy retaining keys when it shouldn't. It affected some cases of joins using on=. Fixed now.

  51. set() and := handle RHS value NA_integer_ on factor types properly. Closes #1234. Thanks to @DavidArenburg.

  52. merge.data.table() didn't set column order (and therefore names) properly in some cases. Fixed now. Closes #1290. Thanks to @ChristK for the minimal example.

  53. print.data.table now works for 100+ rows as intended when row.names=FALSE. Closes #1307. Thanks to @jangorecki for the PR.

  54. Row numbers are not printed in scientific format. Closes #1167. Thanks to @jangorecki for the PR.

  55. Using .GRP unnamed in j now returns a variable named GRP instead of .GRP as the period was causing issues. Same for .BY. Closes #1243; thanks to @MichaelChirico for the PR.

  56. DT[, 0, with=FALSE] returns null data.table to be consistent with data.frame's behaviour. Closes #1140. Thanks to @franknarf1.

  57. Evaluating quoted expressions with . in by works as intended. That is, dt = data.table(a=c(1,1,2,2), b=1:4); expr=quote(.(a)); dt[, sum(b), eval(expr)] works now. Closes #1298. Thanks @eddi.

  58. as.list method for IDate object works properly. Closes #1315. Thanks to @gwerbin.

  1. Clearer explanation of what duplicated() does (borrowed from base). Thanks to @matthieugomez for pointing out. Closes #872.

  2. ?setnames has been updated now that names<- and colnames<- shallow (rather than deep) copy from R >= 3.1.0, #853.

  3. FAQ 1.6 has been embellished, #517. Thanks to a discussion with Vivi and Josh O'Brien.

  4. data.table redefines melt generic and suggests reshape2 instead of import. As a result we don't have to load reshape2 package to use melt.data.table anymore. The reason for this change is that data.table requires R >=2.14, whereas reshape2 R v3.0.0+. Reshape2's melt methods can be used without any issues by loading the package normally.

  5. DT[, j, ] at times made an additional (unnecessary) copy. This is now fixed. This fix also avoids allocating .I when j doesn't use it. As a result := and other subset operations should be faster (and use less memory). Thanks to @szilard for the nice report. Closes #921.

  6. Because reshape2 requires R >3.0.0, and data.table works with R >= 2.14.1, we can not import reshape2 anymore. Therefore we define a melt generic and melt.data.table method for data.tables and redirect to reshape2's melt for other objects. This is to ensure that existing code works fine.

  7. dcast is also a generic now in data.table. So we can use dcast(...) directly, and don't have to spell it out as dcast.data.table(...) like before. The dcast generic in data.table redirects to reshape2::dcast if the input object is not a data.table. But for that you have to load reshape2 before loading data.table. If not, reshape2's dcast overwrites data.table's dcast generic, in which case you will need the :: operator - ex: data.table::dcast(...).

NB: Ideal situation would be for dcast to be a generic in reshape2 as well, but it is not. We have issued a pull request to make dcast in reshape2 a generic, but that has not yet been accepted.

  1. Clarified the use of bit64::integer4 in merge.data.table() and setNumericRounding(). Closes #1093. Thanks to @sfischme for the report.

  2. Removed an unnecessary (and silly) giveNames argument from setDT(). Not sure why I added this in the first place!

  3. options(datatable.prettyprint.char=5L) restricts the number of characters to be printed for character columns. For example:

    options(datatable.prettyprint.char = 5L) DT = data.table(x=1:2, y=c("abcdefghij", "klmnopqrstuv")) DT

    x y

    1: 1 abcde...

    2: 2 klmno...

  4. rolltolast argument in [.data.table is now defunct. It was deprecated in 1.9.4.

  5. data.table's dependency has been moved forward from R 2.14.0 to R 2.14.1, now nearly 4 years old (Dec 2011). As usual before release to CRAN we ensure data.table passes the test suite on the stated dependency and keep this as old as possible for as long as possible. As requested by users in managed environments. For this reason we still don't use paste0() internally, since that was added to R 2.15.0.

  6. Warning about datatable.old.bywithoutby option (for grouping on join without providing by) being deprecated in the next release is in place now. Thanks to @jangorecki for the PR.

  7. Fixed allow.cartesian documentation to nrow(x)+nrow(i) instead of max(nrow(x), nrow(i)). Closes #1123.

  1. by=.EACHI runs j for each group in DT that each row of i joins to.
setkey(DT, ID)
DT[c("id1", "id2"), sum(val)]                # single total across both id1 and id2
DT[c("id1", "id2"), sum(val), by = .EACHI]   # sum(val) for each id
DT[c("id1", "id2"), sum(val), by = key(DT)]  # same
In other words, `by-without-by` is now explicit, as requested by users, [#371](https://github.com/Rdatatable/data.table/issues/371). When `i` contains duplicates, `by=.EACHI` is different to `by=key(DT)`; e.g.,
setkey(DT, ID)
ids = c("id1", "id2", "id1")     # NB: id1 appears twice
DT[ids, sum(val), by = ID]       # 2 rows returned
DT[ids, sum(val), by = .EACHI]   # 3 rows in the order of ids (result 1 and 3 are not merged)
`by=.EACHI` can be useful when `i` is event data, where you don't want the events aggregated by common join values but wish the output to be ordered with repeats, or simply just using join inherited columns as parameters; e.g.;
X[Y, head(.SD, i.top), by = .EACHI]
where `top` is a non-join column in `Y`; i.e., join inherited column. Thanks to many, especially eddi, Sadao Milberg and Gabor Grothendieck for extended discussions. Closes [#538](https://github.com/Rdatatable/data.table/issues/538).
  1. Accordingly, X[Y, j] now does what X[Y][, j] did. To return the old behaviour: options(datatable.old.bywithoutby=TRUE). This is a temporary option to aid migration and will be removed in future. See this, this and this post for discussions and motivation.

  2. Overlap joins (#528) is now here, finally!! Except for type="equal" and maxgap and minoverlap arguments, everything else is implemented. Check out ?foverlaps and the examples there on its usage. This is a major feature addition to data.table.

  3. DT[column==value] and DT[column %in% values] are now optimized to use DT's key when key(DT)[1]=="column", otherwise a secondary key (a.k.a. index) is automatically added so the next DT[column==value] is much faster. No code changes are needed; existing code should automatically benefit. Secondary keys can be added manually using set2key() and existence checked using key2(). These optimizations and function names/arguments are experimental and may be turned off with options(datatable.auto.index=FALSE).

  4. fread():

    • accepts line breaks inside quoted fields. Thanks to Clayton Stanley for highlighting here.
    • accepts trailing backslash in quoted fields. Thanks to user2970844 for highlighting here.
    • Blank and "NA" values in logical columns (T,True,TRUE) no longer cause them to be read as character, #567. Thanks to Adam November for reporting.
    • URLs now work on Windows. R's download.file() converts \r\n to \r\r\n on Windows. Now avoided by downloading in binary mode. Thanks to Steve Miller and Dean MacGregor for reporting, #492.
    • Fixed seg fault in sparse data files when bumping to character, #796 and #722. Thanks to Adam Kennedy and Richard Cotton for the detailed reproducible reports.
    • New argument fread(...,data.table=FALSE) returns a data.frame instead of a data.table. This can be set globally: options(datatable.fread.datatable=FALSE).
  5. .() can now be used in j and is identical to list(), for consistency with i.

DT[,.(MySum=sum(B)),by=...]     # same
DT[,.(colB,colC,colD)]          # same
Similarly, `by=.()` is now a shortcut for `by=list()`, for consistency with `i` and `j`.
  1. rbindlist gains use.names and fill arguments and is now implemented entirely in C. Closes #345:

    • use.names by default is FALSE for backwards compatibility (does not bind by names by default)
    • rbind(...) now just calls rbindlist() internally, except that use.names is TRUE by default, for compatibility with base (and backwards compatibility).
    • fill=FALSE by default. If fill=TRUE, use.names has to be TRUE.
    • When use.names=TRUE, at least one item of the input list has to have non-null column names.
    • When fill=TRUE, all items of the input list has to have non-null column names.
    • Duplicate columns are bound in the order of occurrence, like base.
    • Attributes that might exist in individual items would be lost in the bound result.
    • Columns are coerced to the highest SEXPTYPE when they are different, if possible.
    • And incredibly fast ;).
    • Documentation updated in much detail. Closes #333.
  2. bit64::integer64 now works in grouping and joins, #342. Thanks to James Sams for highlighting UPCs and Clayton Stanley for this SO post. fread() has been detecting and reading integer64 for a while.

  3. setNumericRounding() may be used to reduce to 1 byte or 0 byte rounding when joining to or grouping columns of type 'numeric', #342. See example in ?setNumericRounding and NEWS item below for v1.9.2. getNumericRounding() returns the current setting.

  4. X[Y] now names non-join columns from i that have the same name as a column in x, with an i. prefix for consistency with the i. prefix that has been available in j for some time. This is now documented.

  5. For a keyed table X where the key columns are not at the beginning in order, X[Y] now retains the original order of columns in X rather than moving the join columns to the beginning of the result.

  6. It is no longer an error to assign to row 0 or row NA.

DT[0, colA := 1L]             # now does nothing, silently (was error)
DT[NA, colA := 1L]            # now does nothing, silently (was error)
DT[c(1, NA, 0, 2), colA:=1L]  # now ignores the NA and 0 silently (was error)
DT[nrow(DT) + 1, colA := 1L]  # error (out-of-range) as before
This is for convenience to avoid the need for a switch in user code that evals various `i` conditions in a loop passing in `i` as an integer vector which may containing `0` or `NA`.
  1. A new function setorder is now implemented which uses data.table's internal fast order to reorder rows by reference. It returns the result invisibly (like setkey) that allows for compound statements; e.g., setorder(DT, a, -b)[, cumsum(c), by=list(a,b)]. Check ?setorder for more info.

  2. DT[order(x, -y)] is now by default optimised to use data.table's internal fast order as DT[forder(DT, x, -y)]. It can be turned off by setting datatable.optimize to < 1L or just calling base:::order explicitly. It results in 20x speedup on data.table of 10 million rows with 2 integer columns, for example. To order character vectors in descending order it's sufficient to do DT[order(x, -y)] as opposed to DT[order(x, -xtfrm(y))] in base. This closes #603.

  3. mult="all" -vs- mult="first"|"last" now return consistent types and columns, #340. Thanks to Michele Carriero for highlighting.

  4. duplicated.data.table and unique.data.table gains fromLast = TRUE/FALSE argument, similar to base. Default value is FALSE. Closes #347.

  5. anyDuplicated.data.table is now implemented. Closes #350. Thanks to M C (bluemagister) for reporting.

  6. Complex j-expressions of the form DT[, c(..., lapply(.SD, fun)), by=grp]are now optimised as long as .SD is of the form lapply(.SD, fun) or .SD, .SD[1] or .SD[1L]. This resolves #370. Thanks to Sam Steingold for reporting. This also completes the first two task lists in #735.

## example:
DT[, c(.I, lapply(.SD, sum), mean(x), lapply(.SD, log)), by=grp]
## is optimised to
DT[, list(.I, x=sum(x), y=sum(y), ..., mean(x), log(x), log(y), ...), by=grp]
## and now... these variations are also optimised internally for speed
DT[, c(..., .SD, lapply(.SD, sum), ...), by=grp]
DT[, c(..., .SD[1], lapply(.SD, sum), ...), by=grp]
DT[, .SD, by=grp]
DT[, c(.SD), by=grp]
DT[, .SD[1], by=grp] # Note: but not yet DT[, .SD[1,], by=grp]
DT[, c(.SD[1]), by=grp]
DT[, head(.SD, 1), by=grp] # Note: but not yet DT[, head(.SD, -1), by=grp]
# but not yet optimised
DT[, c(.SD[a], .SD[x>1], lapply(.SD, sum)), by=grp] # where 'a' is, say, a numeric or a data.table, and also for expressions like x>1
The underlying message is that `.SD` is being slowly optimised internally wherever possible, for speed, without compromising in the nice readable syntax it provides.
  1. setDT gains keep.rownames = TRUE/FALSE argument, which works only on data.frames. TRUE retains the data.frame's row names as a new column named rn.

  2. The output of tables() now includes NCOL. Thanks to @dnlbrky for the suggestion.

  3. DT[, LHS := RHS] (or its equivalent in set) now provides a warning and returns DT as it was, instead of an error, when length(LHS) = 0L, #343. For example:

DT[, grep("^b", names(DT)) := NULL] # where no columns start with b
# warns now and returns DT instead of error
  1. GForce now is also optimised for j-expression with .N. Closes #334 and part of #523.
DT[, list(.N, mean(y), sum(y)), by=x] # 1.9.2 - doesn't know to use GForce - will be (relatively) slower
DT[, list(.N, mean(y), sum(y)), by=x] # 1.9.3+ - will use GForce.
  1. setDF is now implemented. It accepts a data.table and converts it to data.frame by reference, #338. Thanks to canneff for the discussion here on data.table mailing list.

  2. .I gets named as I (instead of .I) wherever possible, similar to .N, #344.

  3. setkey on .SD is now an error, rather than warnings for each group about rebuilding the key. The new error is similar to when attempting to use := in a .SD subquery: ".SD is locked. Using set*() functions on .SD is reserved for possible future use; a tortuously flexible way to modify the original data by group." Thanks to Ron Hylton for highlighting the issue on datatable-help here.

  4. Looping calls to unique(DT) such as in DT[,unique(.SD),by=group] is now faster by avoiding internal overhead of calling [.data.table. Thanks again to Ron Hylton for highlighting in the same thread. His example is reduced from 28 sec to 9 sec, with identical results.

  5. Following gsum and gmean, now gmin and gmax from GForce are also implemented. Closes part of #523. Benchmarks are also provided.

DT[, list(sum(x), min(y), max(z), .N), by=...] # runs by default using GForce
  1. setorder() and DT[order(.)] handles integer64 type in descending order as well. Closes #703.

  2. setorder() and setorderv() gain na.last = TRUE/FALSE. Closes #706.

  3. .N is now available in i, FR#724. Thanks to newbie indirectly here and Farrel directly here.

  4. by=.EACHI is now implemented for not-joins as well. Closes #604. Thanks to Garrett See for filing the FR. As an example:

DT = data.table(x=c(1,1,1,1,2,2,3,4,4,4), y=1:10, key="x")
DT[!J(c(1,4)), sum(y), by=.EACHI] # is equivalent to DT[J(c(2,3)), sum(y), by=.EACHI]
  1. When joining to fewer columns than the key has, using one of the later key columns explicitly in j repeated the first value. A problem introduced by v1.9.2 and not caught bythe 1,220 tests, or tests in 37 dependent packages. Test added. Many thanks to Michele Carriero for reporting.
DT = data.table(a=1:2, b=letters[1:6], key="a,b")    # keyed by a and b
DT[.(1), list(b,...)]    # correct result again (joining just to a not b but using b)
  1. setkey works again when a non-key column is type list (e.g. each cell can itself be a vector), #54. Test added. Thanks to James Sams, Michael Nelson and Musx for the reproducible examples.

  2. The warning "internal TRUE value has been modified" with recently released R 3.1 when grouping a table containing a logical column and where all groups are just 1 row is now fixed and tests added. Thanks to James Sams for the reproducible example. The warning is issued by R and we have asked if it can be upgraded to error (UPDATE: change now made for R 3.1.1 thanks to Luke Tierney).

  3. data.table(list()), data.table(data.table()) and data.table(data.frame()) now return a null data.table (no columns) rather than one empty column, #48. Test added. Thanks to Shubh Bansal for reporting.

  4. unique(<NULL data.table>) now returns a null data.table, #44. Thanks to agstudy for reporting.

  5. data.table() converted POSIXlt to POSIXct, consistent with base:::data.frame(), but now also provides a helpful warning instead of coercing silently, #59. Thanks to Brodie Gaslam, Patrick and Ragy Isaac for reporting here and here.

  6. If another class inherits from data.table; e.g. class(DT) == c("UserClass","data.table","data.frame") then DT[...] now retains UserClass in the result. Thanks to Daniel Krizian for reporting, #64. Test added.

  7. An error object '<name>' not found could occur in some circumstances, particularly after a previous error. Reported on SO with non-ASCII characters in a column name, a red herring we hope since non-ASCII characters are fully supported in data.table including in column names. Fix implemented and tests added.

  8. Column order was reversed in some cases by as.data.table.table(), #43. Test added. Thanks to Benjamin Barnes for reporting.

  9. DT[, !"missingcol", with=FALSE] now returns DT (rather than a NULL data.table) with warning that "missingcol" is not present.

  10. DT[,y := y * eval(parse(text="1*2"))] resulted in error unless eval() was wrapped with paranthesis. That is, DT[,y := y * (eval(parse(text="1*2")))], #5423. Thanks to Wet Feet for reporting and to Simon O'Hanlon for identifying the issue here on SO.

  11. Using by columns with attributes (ex: factor, Date) in j did not retain the attributes, also in case of :=. This was partially a regression from an earlier fix (#155) due to recent changes for R3.1.0. Now fixed and clearer tests added. Thanks to Christophe Dervieux for reporting and to Adam B for reporting here on SO. Closes #36.

  12. .BY special variable did not retain names of the grouping columns which resulted in not being able to access .BY$grpcol in j. Ex: DT[, .BY$x, by=x]. This is now fixed. Closes #5415. Thanks to Stephane Vernede for the bug report.

  13. Fixed another issue with eval(parse(...)) in j along with assignment by reference :=. Closes #30. Thanks to Michele Carriero for reporting.

  14. get() in j did not see i's columns when i is a data.table which lead to errors while doing operations like: DT1[DT2, list(get('c'))]. Now, use of get makes all x's and i's columns visible (fetches all columns). Still, as the verbose message states, using .SDcols or eval(macro) would be able to select just the columns used, which is better for efficiency. Closes #34. Thanks to Eddi for reporting.

  15. Fixed an edge case with unique and duplicated, which on empty data.tables returned a 1-row data.table with all NAs. Closes #28. Thanks to Shubh Bansal for reporting.

  16. dcast.data.table resuled in error (because function CJ() was not visible) in packages that "import" data.table. This did not happen if the package "depends" on data.table. Closes bug #31. Thanks to K Davis for the excellent report.

  17. merge(x, y, all=TRUE) error when x is empty data.table is now fixed. Closes #24. Thanks to Garrett See for filing the report.

  18. Implementing #5249 closes bug #26, a case where rbind gave error when binding with empty data.tables. Thanks to Roger for reporting on SO.

  19. Fixed a segfault during grouping with assignment by reference, ex: DT[, LHS := RHS, by=.], where length(RHS) > group size (.N). Closes #25. Thanks to Zachary Long for reporting on datatable-help mailing list.

  20. Consistent subset rules on datat.tables with duplicate columns. In short, if indices are directly provided, 'j', or in .SDcols, then just those columns are either returned (or deleted if you provide -.SDcols or !j). If instead, column names are given and there are more than one occurrence of that column, then it's hard to decide which to keep and which to remove on a subset. Therefore, to remove, all occurrences of that column are removed, and to keep, always the first column is returned each time. Also closes #22 and #86.

    Note that using by= to aggregate on duplicate columns may not give intended result still, as it may not operate on the proper column.

  21. When DT is empty, DT[, newcol:=max(b), by=a] now properly adds the column, #49. Thanks to Shubh bansal for filing the report.

  22. When j evaluates to integer(0)/character(0), DT[, j, with=FALSE] resulted in error, #21. Thanks indirectly to Malcolm Cook for #52, through which this (recent) regression (from 1.9.3) was found.

  23. print(DT) now respects digits argument on list type columns, #37. Thanks to Frank for the discussion on the mailing list and to Matthew Beckers for filing the bug report.

  24. FR # 2551 implemented leniance in warning messages when columns are coerced with DT[, LHS := RHS], when length(RHS)==1. But this was very lenient; e.g., DT[, a := "bla"], where a is a logical column should get a warning. This is now fixed such that only very obvious cases coerces silently; e.g., DT[, a := 1] where a is integer. Closes #35. Thanks to Michele Carriero and John Laing for reporting.

  25. dcast.data.table provides better error message when fun.aggregate is specified but it returns length != 1. Closes #693. Thanks to Trevor Alexander for reporting here on SO.

  26. dcast.data.table tries to preserve attributes whereever possible, except when value.var is a factor (or ordered factor). For factor types, the casted columns will be coerced to type character thereby losing the levels attribute. Closes #688. Thanks to juancentro for reporting.

  27. melt now returns friendly error when meaure.vars are not in data instead of segfault. Closes #699. Thanks to vsalmendra for this post on SO and the subsequent bug report.

  28. DT[, list(m1 = eval(expr1), m2=eval(expr2)), by=val] where expr1 and expr2 are constructed using parse(text=.) now works instead of resulting in error. Closes #472. Thanks to Benjamin Barnes for reporting with a nice reproducible example.

  29. A join of the form X[Y, roll=TRUE, nomatch=0L] where some of Y's key columns occur more than once (duplicated keys) might at times return incorrect join. This was introduced only in 1.9.2 and is fixed now. Closes #700. Thanks to Michael Smith for the very nice reproducible example and nice spotting of such a tricky case.

  30. Fixed an edge case in DT[order(.)] internal optimisation to be consistent with base. Closes #696. Thanks to Michael Smith and Garrett See for reporting.

  31. DT[, list(list(.)), by=.] and DT[, col := list(list(.)), by=.] now return correct results in R >= 3.1.0. The bug was due to a welcome change in R 3.1.0 where list(.) no longer copies. Closes #481. Also thanks to KrishnaPG for filing #728.

  32. dcast.data.table handles fun.aggregate argument properly when called from within a function that accepts fun.aggregate argument and passes to dcast.data.table(). Closes #713. Thanks to mathematicalcoffee for reporting here on SO.

  33. dcast.data.table now returns a friendly error when fun.aggregate value for missing combinations is 0-length, and 'fill' argument is not provided. Closes #715

  34. rbind/rbindlist binds in the same order of occurrence also when binding tables with duplicate names along with 'fill=TRUE' (previously, it grouped all duplicate columns together). This was the underlying reason for #725. Thanks to Stefan Fritsch for the report with a nice reproducible example and discussion.

  35. setDT now provides a friendly error when attempted to change a variable to data.table by reference whose binding is locked (usually when the variable is within a package, ex: CO2). Closes #475. Thanks to David Arenburg for filing the report here on SO.

  36. X[!Y] where X and Y are both data.tables ignores 'allow.cartesian' argument, and rightly so because a not-join (or anti-join) cannot exceed nrow(x). Thanks to @fedyakov for spotting this. Closes #698.

  37. as.data.table.matrix does not convert strings to factors by default. data.table likes and prefers using character vectors to factors. Closes #745. Thanks to @fpinter for reporting the issue on the github issue tracker and to vijay for reporting here on SO.

  38. Joins of the form x[y[z]] resulted in duplicate names when all x, y and z had the same column names as non-key columns. This is now fixed. Closes #471. Thanks to Christian Sigg for the nice reproducible example.

  39. DT[where, someCol:=NULL] is now an error that i is provided since it makes no sense to delete a column for only a subset of rows. Closes #506.

  40. forder did not identify -0 as 0 for numeric types. This is fixed now. Thanks to @arcosdium for nice minimal example. Closes #743.

  41. Segfault on joins of the form X[Y, c(..), by=.EACHI] is now fixed. Closes #744. Thanks to @nigmastar (Michele Carriero) for the excellent minimal example.

  42. Subset on data.table using lapply of the form lapply(L, "[", Time == 3L) works now without error due to [.data.frame redirection. Closes #500. Thanks to Garrett See for reporting.

  43. id.vars and measure.vars default value of NULL was removed to be consistent in behaviour with reshape2:::melt.data.frame. Closes #780. Thanks to @dardesta for reporting.

  44. Grouping using external variables on keyed data.tables did not return correct results at times. Thanks to @colinfang for reporting. Closes #762.

  1. Reminder: using rolltolast still works but since v1.9.2 now issues the following warning:
 > 'rolltolast' has been marked 'deprecated' in ?data.table since v1.8.8 on CRAN 3 Mar 2013, see NEWS. Please change to the more flexible 'rollends' instead. 'rolltolast' will be removed in the next version."
  1. Using with=FALSE with := is now deprecated in all cases, given that wrapping the LHS of := with parentheses has been preferred for some time.
colVar = "col1"
DT[, colVar:=1, with=FALSE]                   # deprecated, still works silently as before
DT[, (colVar):=1]                             # please change to this
DT[, c("col1","col2"):=1]                     # no change
DT[, 2:4 := 1]                                # no change
DT[, c("col1","col2"):=list(sum(a),mean(b)]   # no change
DT[, `:=`(...), by=...]                       # no change
The next release will issue a warning when `with=FALSE` is used with `:=`.
  1. ?duplicated.data.table explained that by=NULL or by=FALSE would use all columns, however by=FALSE resulted in error. by=FALSE is removed from help and duplicated returns an error when by=TRUE/FALSE now. Closes #38.

  2. More info about distinguishing small numbers from 0.0 in v1.9.2+ is here.

  3. ?dcast.data.table now explains how the names are generated for the columns that are being casted. Closes #5676.

  4. dcast.data.table(dt, a ~ ... + b) now generates the column names with values from b coming last. Closes #5675.

  5. Added x[order(.)] internal optimisation, and how to go back to base:::order(.) if one wants to sort by session locale to ?setorder (with alias ?order and ?forder). Closes #478 and also #704. Thanks to Christian Wolf for the report.

  6. Added tests (1351.1 and 1351.2) to catch any future regressions on particular case of binary search based subset reported here on SO. Thanks to Scott for the post. The regression was contained to v1.9.2 AFAICT. Closes #734.

  7. Added an .onUnload method to unload data.table's shared object properly. Since the name of the shared object is 'datatable.so' and not 'data.table.so', 'detach' fails to unload correctly. This was the reason for the issue reported here on SO. Closes #474. Thanks to Matthew Plourde for reporting.

  8. Updated BugReports link in DESCRIPTION. Thanks to @chrsigg for reporting. Closes #754.

  9. Added shiny, rmarkdown and knitr to the data.table whitelist. Packages which take user code as input and run it in their own environment (so do not Depend or Import data.table themselves) either need to be added here, or they can define a variable .datatable.aware <- TRUE in their namepace, so that data.table can work correctly in those packages. Users can also add to data.table's whitelist themselves using assignInNamespace() but these additions upstream remove the need to do that for these packages.

  10. Clarified with=FALSE as suggested in #513.

  11. Clarified .I in ?data.table. Closes #510. Thanks to Gabor for reporting.

  12. Moved ?copy to its own help page, and documented that dt_names <- copy(names(DT)) is necessary for dt_names to be not modified by reference as a result of updating DT by reference (e.g. adding a new column by reference). Closes #512. Thanks to Zach for this SO question and user1971988 for this SO question.

  13. address(x) doesn't increment NAM() value when x is a vector. Using the object as argument to a non-primitive function is sufficient to increment its reference. Closes #824. Thanks to @tarakc02 for the question on twitter and hint from Hadley.

  1. Fast methods of reshape2's melt and dcast have been implemented for data.table, FR #2627. Most settings are identical to reshape2, see ?melt.data.table.
> `melt`: 10 million rows and 5 columns, 61.3 seconds reduced to 1.2 seconds.  
> `dcast`: 1 million rows and 4 columns, 192 seconds reduced to 3.6 seconds.  
  * `melt.data.table` is also capable of melting on columns of type `list`.
  * `melt.data.table` gains `variable.factor` and `value.factor` which by default are TRUE and FALSE respectively for compatibility with `reshape2`. This allows for directly controlling the output type of "variable" and "value" columns (as factors or not).
  * `melt.data.table`'s `na.rm = TRUE` parameter is optimised to remove NAs directly during melt and therefore avoids the overhead of subsetting using `!is.na` afterwards on the molten data. 
  * except for `margins` argument from `reshape2:::dcast`, all features of dcast are intact. `dcast.data.table` can also accept `value.var` columns of type list.

> Reminder of Cologne (Dec 2013) presentation **slide 32** : ["Why not submit a dcast pull request to reshape2?"](https://github.com/Rdatatable/data.table/wiki/talks/CologneR_2013.pdf).
  1. Joins scale better as the number of rows increases. The binary merge used to start on row 1 of i; it now starts on the middle row of i. Many thanks to Mike Crowe for the suggestion. This has been done within column so scales much better as the number of join columns increase, too.
> Reminder: bmerge allows the rolling join feature: forwards, backwards, limited and nearest.  
  1. Sorting (setkey and ad-hoc by=) is faster and scales better on randomly ordered data and now also adapts to almost sorted data. The remaining comparison sorts have been removed. We use a combination of counting sort and forwards radix (MSD) for all types including double, character and integers with range>100,000; forwards not backwards through columns. This was inspired by Terdiman and Herf's (LSD) radix approach for floating point :

  2. unique and duplicated methods for data.table are significantly faster especially for type numeric (i.e. double), and type integer where range > 100,000 or contains negatives.

  3. NA, NaN, +Inf and -Inf are now considered distinct values, may be in keys, can be joined to and can be grouped. data.table defines: NA < NaN < -Inf. Thanks to Martin Liberts for the suggestions, #4684, #4815 and #4883.

  4. Numeric data is still joined and grouped within tolerance as before but instead of tolerance being sqrt(.Machine$double.eps) == 1.490116e-08 (the same as base::all.equal's default) the significand is now rounded to the last 2 bytes, apx 11 s.f. This is more appropriate for large (1.23e20) and small (1.23e-20) numerics and is faster via a simple bit twiddle. A few functions provided a 'tolerance' argument but this wasn't being passed through so has been removed. We aim to add a global option (e.g. 2, 1 or 0 byte rounding) in a future release.

  5. New optimization: GForce. Rather than grouping the data, the group locations are passed into grouped versions of sum and mean (gsum and gmean) which then compute the result for all groups in a single sequential pass through the column for cache efficiency. Further, since the g* function is called just once, we don't need to find ways to speed up calling sum or mean repetitively for each group. Plan is to add gmin, gmax, gsd, gprod, gwhich.min and gwhich.max. Examples where GForce applies now:

DT[,sum(x,na.rm=),by=...]                       # yes
DT[,list(sum(x,na.rm=),mean(y,na.rm=)),by=...]  # yes
DT[,lapply(.SD,sum,na.rm=),by=...]              # yes
DT[,list(sum(x),min(y)),by=...]                 # no. gmin not yet available, only sum and mean so far.
GForce is a level 2 optimization. To turn it off: `options(datatable.optimize=1)`. Reminder: to see the optimizations and other info, set `verbose=TRUE`
  1. fread's integer64 argument implemented. Allows reading of integer64 data as 'double' or 'character' instead of bit64::integer64 (which remains the default as before). Thanks to Chris Neff for the suggestion. The default can be changed globally; e.g, options(datatable.integer64="character")

  2. fread's drop, select and NULL in colClasses are implemented. To drop or select columns by name or by number. See examples in ?fread.

  3. fread now detects T, F, True, False, TRUE and FALSE as type logical, consistent with read.csv, #4766. Thanks to Adam November for highlighting.

  4. fread now accepts quotes (both ' and ") in the middle of fields, whether the field starts with " or not, rather than the 'unbalanced quotes' error, #2694. Thanks to baidao for reporting. It was known and documented at the top of ?fread (now removed). If a field starts with " it must end with " (necessary to include the field separator itself in the field contents). Embedded quotes can be in column names, too. Newlines (\n) still can't be in quoted fields or quoted column names, yet.

  5. fread gains showProgress, default TRUE. The global option is datatable.showProgress.

  6. fread("1.46761e-313\n") detected the ERANGE error, so read as character. It now reads as numeric but with a detailed warning. Thanks to Heather Turner for the detailed report, #4879.

  7. fread now understand system commands; e.g., fread("grep blah file.txt").

  8. as.data.table method for table() implemented, #4848. Thanks to Frank Pinter for suggesting here on SO.

  9. as.data.table methods added for integer, numeric, character, logical, factor, ordered and Date.

  10. DT[i,:=,] now accepts negative indices in i. Thanks to Eduard Antonyan. See also bug fix #2697.

  11. set() is now able to add new columns by reference, #2077.

DT[3:5, newCol := 5L]
set(DT, i=3:5, j="newCol", 5L)   # same
  1. eval will now be evaluated anywhere in a j-expression as long as it has just one argument, #4677. Will still need to use .SD as environment in complex cases. Also fixes bug here on SO.

  2. ! at the head of the expression will no longer trigger a not-join if the expression is logical, #4650. Thanks to Arunkumar Srinivasan for reporting.

  3. rbindlist now chooses the highest type per column, not the first, #2456. Up-conversion follows R defaults, with the addition of factors being the highest type. Also fixes #4981 for the specific case of NA's.

  4. cbind(x,y,z,...) now creates a data.table if x isn't a data.table but y or z is, unless x is a data.frame in which case a data.frame is returned (use data.table(DF,DT) instead for that).

  5. cbind(x,y,z,...) and data.table(x,y,z,...) now retain keys of any data.table inputs directly (no sort needed, for speed). The result's key is c(key(x), key(y), key(z), ...), provided, that the data.table inputs that have keys are not recycled and there are no ambiguities (i.e. duplicates) in column names.

  6. rbind/rbindlist will preserve ordered factors if it's possible to do so; i.e., if a compatible global order exists, #4856 & #5019. Otherwise the result will be a factor and a warning.

  7. rbind now has a fill argument, #4790. When fill=TRUE it will behave in a manner similar to plyr's rbind.fill. This option is incompatible with use.names=FALSE. Thanks to Arunkumar Srinivasan for the base code.

  8. rbind now relies exclusively on rbindlist to bind data.tables together. This makes rbind'ing factors faster, #2115.

  9. DT[, as.factor('x'), with=FALSE] where x is a column in DT is now equivalent to DT[, "x", with=FALSE] instead of ending up with an error, #4867. Thanks to tresbot for reporting here on SO.

  10. format.data.table now understands 'formula' and displays embedded formulas as expected, FR #2591.

  11. {} around := in j now obtain desired result, but with a warning #2496. Now,

DT[, { `:=`(...)}]             # now works
DT[, {`:=`(...)}, by=(...)]    # now works
Thanks to Alex for reporting [here on SO](http://stackoverflow.com/questions/14541959/expression-syntax-for-data-table-in-r).
  1. x[J(2), a], where a is the key column sees a in j, #2693 and FAQ 2.8. Also, x[J(2)] automatically names the columns from i using the key columns of x. In cases where the key columns of x and i are identical, i's columns can be referred to by using i.name; e.g., x[J(2), i.a]. Thanks to mnel and Gabor for the discussion here.

  2. print.data.table gains row.names, default=TRUE. When FALSE, the row names (along with the :) are not printed, #5020. Thanks to Frank Erickson.

  3. .SDcols now is also able to de-select columns. This works both with column names and column numbers.

DT[, lapply(.SD,...), by=..., .SDcols=-c(1,3)]       # .SD all but columns 1 and 3
DT[, lapply(.SD,...), by=..., .SDcols=-c("x", "z")]  # .SD all but columns 'x' and 'z'
DT[..., .SDcols=c(1, -3)]           # can't mix signs, error
DT[, .SD, .SDcols=c("x", -"z")]     # can't mix signs, error
Thanks to Tonny Peterson for filing FR #4979.
  1. as.data.table.list now issues a warning for those items/columns that result in a remainder due to recycling, #4813. data.table() also now issues a warning (instead of an error previously) when recycling leaves a remainder; e.g., data.table(x=1:2, y=1:3).

  2. := now coerces without warning when precision is not lost and length(RHS) == 1, #2551.

DT = data.table(x=1:2, y=c(TRUE, FALSE))
DT[1, x:=1]   # ok, now silent
DT[1, y:=0]   # ok, now silent
DT[1, y:=0L]  # ok, now silent
  1. as.data.table.*(x, keep.rownames=TRUE), where x is a named vector now adds names of x into a new column with default name rn. Thanks to Garrett See for FR #2356.

  2. X[Y, col:=value] when no match exists in the join is now caught early and X is simply returned. Also a message when datatable.verbose is TRUE is provided. In addition, if col is an existing column, since no update actually takes place, the key is now retained. Thanks to Frank Erickson for suggesting, #4996.

  3. New function setDT() takes a list (named and/or unnamed) or data.frame and changes its type by reference to data.table, without any copy. It also has a logical argument giveNames which is used for a list inputs. See ?setDT examples for more. Based on this FR on SO.

  4. setnames(DT,"oldname","newname") no longer complains about any duplicated column names in DT so long as oldname is unique and unambiguous. Thanks to Wet Feet for highlighting here on SO.

  5. last(x) where length(x)=0 now returns 'x' instead of an error, #5152. Thanks to Garrett See for reporting.

  6. as.ITime.character no longer complains when given vector input, and will accept mixed format time entries; e.g., c("12:00", "13:12:25")

  7. Key is now retained in NA subsets; e.g.,

DT = data.table(a=1:3,b=4:6,key="a")
DT[NA]   # 1-row of NA now keyed by 'a'
DT[5]    # 1-row of NA now keyed by 'a'
DT[2:4]  # not keyed as before because NA (last row of result) sorts first in keyed data.table
  1. Each column in the result for each group has always been recycled (if necessary) to match the longest column in that group's result. If it doesn't recycle exactly, though, it was caught gracefully as an error. Now, it is recycled, with remainder with warning.
DT = data.table(a=1:2,b=1:6)
DT[, list(b,1:2), by=a]        # now recycles the 1:2 with warning to length 3
  1. Long outstanding (usually small) memory leak in grouping fixed, #2648. When the last group is smaller than the largest group, the difference in those sizes was not being released. Also evident in non-trivial aggregations where each group returns a different number of rows. Most users run a grouping query once and will never have noticed these, but anyone looping calls to grouping (such as when running in parallel, or benchmarking) may have suffered. Tests added. Thanks to many including vc273 and Y T for reporting here and here on SO.

  2. In long running computations where data.table is called many times repetitively the following error could sometimes occur, #2647: "Internal error: .internal.selfref prot is not itself an extptr". Now fixed. Thanks to theEricStone, StevieP and JasonB for (difficult) reproducible examples here.

  3. If fread returns a data error (such as no closing quote on a quoted field) it now closes the file first rather than holding a lock open, a Windows only problem. Thanks to nigmastar for reporting here and Carl Witthoft for the hint. Tests added.

  4. DT[0,col:=value] is now a helpful error rather than crash, #2754. Thanks to Ricardo Saporta for reporting. DT[NA,col:=value]'s error message has also been improved. Tests added.

  5. Assigning to the same column twice in the same query is now an error rather than a crash in some circumstances; e.g., DT[,c("B","B"):=NULL] (delete by reference the same column twice). Thanks to Ricardo (#2751) and matt_k (#2791) for reporting here. Tests added.

  6. Crash and/or incorrect aggregate results with negative indexing in i is fixed, with a warning when the abs(negative index) > nrow(DT), #2697. Thanks to Eduard Antonyan (eddi) for reporting here. Tests added.

  7. head() and tail() handle negative n values correctly now, #2375. Thanks to Garrett See for reporting. Also it results in an error when length(n) != 1. Tests added.

  8. Crash when assigning empty data.table to multiple columns is fixed, #4731. Thanks to Andrew Tinka for reporting. Tests added.

  9. print(DT, digits=2) now heeds digits and other parameters, #2535. Thanks to Heather Turner for reporting. Tests added.

  10. print(data.table(table(1:101))) is now an 'invalid column' error and suggests print(as.data.table(table(1:101))) instead, #4847. Thanks to Frank Pinter for reporting. Test added.

  11. Crash when grouping by character column where i is integer(0) is now fixed. It now returns an appropriate empty data.table. This fixes bug #2440. Thanks to Malcolm Cook for reporting. Tests added.

  12. Grouping when i has value '0' and length(i) > 1 resulted in crash; it is now fixed. It returns a friendly error instead. This fixes bug #2758. Thanks to Garrett See for reporting. Tests added.

  13. := failed while subsetting yielded NA and with=FALSE, #2445. Thanks to Damian Betebenner for reporting.

  14. by=month(date) gave incorrect results if key(DT)=="date", #2670. Tests added.

DT[,,by=month(date)]         # now ok if key(DT)=="date"
DT[,,by=list(month(date))]   # ok before whether or not key(DT)=="date"
  1. rbind and rbindlist could crash if input columns themselves had hidden names, #4890 & #4912. Thanks to Chris Neff and Stefan Fritsch for reporting. Tests added.

  2. data.table(), as.data.table() and other paths to create a data.table now detect and drop hidden names, the root cause of #4890. It was never intended that columns could have hidden names attached.

  3. Cartesian Join (allow.cartesian = TRUE) when both x and i are keyed and length(key(x)) > length(key(i)) set resulting key incorrectly. This is now fixed, #2677. Tests added. Thanks to Shir Levkowitz for reporting.

  4. := (assignment by reference) loses POSIXct or ITime attribute while grouping is now fixed, #2531. Tests added. Thanks to stat quant for reporting here and to Paul Murray for reporting here on SO.

  5. chmatch() didn't always match non-ascii characters, #2538 and #4818. chmatch is used internally so DT[is.na(päs), päs := 99L] now works. Thanks to Benjamin Barnes and Stefan Fritsch for reporting. Tests added.

  6. unname(DT) threw an error when 20 < nrow(DT) <= 100, bug #4934. This is now fixed. Tests added. Thanks to Ricardo Saporta.

  7. A special case of not-join and logical TRUE, DT[!TRUE], gave an error whereas it should be identical to DT[FALSE]. Now fixed and tests added. Thanks once again to Ricardo Saporta for filing #4930.

  8. X[Y,roll=-Inf,rollends=FALSE] didn't roll the middle correctly if Y was keyed. It was ok if Y was unkeyed or rollends left as the default [c(TRUE,FALSE) when roll < 0]. Thanks to user338714 for reporting here. Tests added.

  9. Key is now retained after an order-preserving subset, #295.

  10. Fixed bug #2584. Now columns that had function names, in particular "list" do not pose problems in .SD. Thanks to Zachary Mayer for reporting.

  11. Fixed bug #4927. Unusual column names in normal quotes, ex: by=".Col", now works as expected in by. Thanks to Ricardo Saporta for reporting.

  12. setkey resulted in error when column names contained ",". This is now fixed. Thanks to Corone for reporting here on SO.

  13. rbind when at least one argument was a data.table, but not the first, returned the rbind'd data.table with key. This is now fixed, #4995. Thanks to Frank Erickson for reporting.

  14. That .SD doesn't retain column's class is now fixed (#2530). Thanks to Corone for reporting here.

  15. eval(quote()) returned error when the quoted expression is a not-join, #4994. This is now fixed. Tests added.

  16. DT[, lapply(.SD, function(), by=] did not see columns of DT when optimisation is "on". This is now fixed, #2381. Tests added. Thanks to David F for reporting here on SO.

  17. #4959 - rbind'ing empty data.tables now works

  18. #5005 - some function expressions were not being correctly evaluated in j-expression. Thanks to Tonny Petersen for reporting.

  19. Fixed bug #5007, j did not see variables declared within a local (function) environment properly. Now, DT[, lapply(.SD, function(x) fun_const), by=x] where "fun_const" is a local variable within a function works as expected. Thanks to Ricardo Saporta for catching this and providing a very nice reproducible example.

  20. Fixing #5007 also fixes #4957, where .N was not visible during lapply(.SD, function(x) ...) in j. Thanks to juba for noticing it here on SO.

  21. Fixed another case where function expressions were not constructed properly in j, while fixing #5007. DT[, lapply(.SD, function(x) my_const), by=x] now works as expected instead of ending up in an error.

  22. Fixed #4990, where := did not generate a recycling warning during by, when length(RHS) < group size but not an integer multiple of group size. Now,

DT <- data.table(a=rep(1:2, c(5,2)))
DT[, b := c(1:2), by=a]
   will generate a warning (here for first group as RHS length (2) is not an integer multiple of group size (=5)). 
  1. Fixed #5069 where gdata:::write.fwf returned an error with data.table.

  2. Fixed #5098 where construction of j-expression with a function with no-argument returned the function definition instead of returning the result from executing the function.

  3. Fixed #5106 where DT[, .N, by=y] where y is a vector with length(y) = nrow(DT), but y is not a column in DT. Thanks to colinfang for reporting.

  4. Fixed #5104 which popped out as a side-effect of fixing #2531. := while grouping and assigning columns that are factors resulted in wrong results (and the column not being added). This is now fixed. Thanks to Jonathen Owen for reporting.

  5. Fixed bug #5114 where modifying columns in particular cases resulted in ".SD is locked" error. Thanks to GSee for the bug report.

  6. Implementing FR #4979 lead to a bug when grouping with .SDcols, where .SDcols argument was variable name. This bug #5190 is now fixed.

  7. Fixed #5171 - where setting the attribute name to a non-character type resulted in a segfault. Ex: setattr(x, FALSE, FALSE); x. Now ends up with a friendly error.

  8. Dependent packages using cbind may now Import data.table as intended rather than needing to Depend. There was a missing data.table:: prefix on a call to key(). Thanks to Maarten-Jan Kallen for reporting.

  9. 'chmatch' didn't handle character encodings properly when the string was identical but the encoding were different. For ex: UTF8 and Latin1. This is now fixed (a part of bug #5159). Thanks to Stefan Fritsch for reporting.

  10. Joins (X[Y]) on character columns with different encodings now issue a warning that join may result in unexpected results for those indices with different encodings. That is, when "ä" in X's key column and "ä" in Y's key column are of different encodings, a warning is issued. This takes care of bugs #5266 and other part of #5159 for the moment. Thanks to Stefan Fritsch once again for reporting.

  11. Fixed #5117 - segfault when rbindlist on empty data.tables. Thanks to Garrett See for reporting.

  12. Fixed a rare segfault that occurred on >250m rows (integer overflow during memory allocation); closes #5305. Thanks to Guenter J. Hitsch for reporting.

  13. rbindlist with at least one factor column along with the presence of at least one empty data.table resulted in segfault (or in linux/mac reported an error related to hash tables). This is now fixed, #5355. Thanks to Trevor Alexander for reporting on SO (and mnel for filing the bug report):

  14. CJ() now orders character vectors in a locale consistent with setkey, #5375. Typically this affected whether upper case letters were ordered before lower case letters; they were by setkey() but not by CJ(). This difference started in v1.8.10 with the change "CJ() is 90% faster...", see NEWS below. Test added and avenues for differences closed off and nailed down, with no loss in performance. Many thanks to Malcolm Hawkes for reporting.

  1. Zach Mayer for a reproducible segfault related to radix sorting character strings longer than 20. Test added.

  2. Simon Biggs for reporting a bug in fread'ing logicals. Test added.

  3. Jakub Szewczyk for reporting that where "." is used in formula interface of dcast.data.table along with an aggregate function, it did not result in aggregated result, #5149. Test added.

dcast.data.table(x, a ~ ., mean, value.var="b")
  1. Jonathan Owen for reporting that DT[,sum(.SD),by=] failed with GForce optimization, #5380. Added test and error message redirecting to use DT[,lapply(.SD,sum),by=] or base::sum and how to turn off GForce.

  2. Luke Tierney for guidance in finding a corruption of R_TrueValue which needed --enable-strict-barier, gctorture2 and a hardware watchpoint to ferret out. Started after a change in Rdevel on 11 Feb 2014, r64973.

  3. Minkoo Seo for a new test on rbindlist, #4648.

  4. Gsee for reporting that set() and := could no longer add columns by reference to an object that inherits from data.table; e.g., class = c("myclass", data.table", "data.frame")), #5115.

  5. Clayton Stanley for reporting #5307 here on SO. Aggregating logical types could give wrong results.

  6. New and very welcome ASAN and UBSAN checks on CRAN detected :

    • integer64 overflow in test 899 reading integers longer than apx 18 digits
      fread("Col1\n12345678901234567890")`   # works as before, bumped to character
    • a memory fault in rbindlist when binding ordered factors, and, some items in the list of data.table/list are empty or NULL. In both cases we had anticipated and added tests for these cases, which is why ASAN and UBSAN were able to detect a problem for us.
  7. Karl Millar for reporting a similar fault that ASAN detected, #5042. Also fixed.

  8. Ricardo Saporta for finding a crash when i is empty and a join column is character, #5387. Test added.

  1. If fread detects data which would be lost if the column was read according to type supplied in colClasses, e.g. a numeric column specified as integer in colClasses, the message that it has ignored colClasses is upgraded to warning instead of just a line in verbose=TRUE mode.

  2. ?last has been improved and if xts is needed but not installed the error message is more helpful, #2728. Thanks to Sam Steingold for reporting.

  3. ?between corrected. It returns a logical not integer vector, #2671. Thanks to Michael Nelson for reporting.

  4. .SD, .N, .I, .GRP and .BY are now exported (as NULL). So that NOTEs aren't produced for them by R CMD check or codetools::checkUsage via compiler. utils::globalVariables() was considered, but exporting chosen. Thanks to Sam Steingold for raising, #2723.

  5. When DT is empty, DT[,col:=""] is no longer a warning. The warning was:

    "Supplied 1 items to be assigned to 0 items of column (1 unused)"

  6. Using rolltolast still works but now issues the following warning :

    "'rolltolast' has been marked 'deprecated' in ?data.table since v1.8.8 on CRAN 3 Mar 2013, see NEWS. Please change to the more flexible 'rollends' instead. 'rolltolast' will be removed in the next version."

  7. There are now 1,220 raw tests, as reported by test.data.table().

  8. data.table's dependency has been moved forward from R 2.12.0 to R 2.14.0, now over 2 years old (Oct 2011). As usual before release to CRAN, we ensure data.table passes R CMD check on the stated dependency and keep this as old as possible for as long as possible. As requested by users in managed environments. For this reason we still don't use paste0() internally, since that was added to R 2.15.0.

  • fread :

    • If some column names are blank they are now given default names rather than causing the header row to be read as a data row. Thanks to Simon Judes for suggesting.

    • "+" and "-" are now read as character rather than integer 0. Thanks to Alvaro Gonzalez and Roby Joehanes for reporting, #4814. http://stackoverflow.com/questions/15388714/reading-strand-column-with-fread-data-table-package

    • % progress console meter has been removed. The ouput was inconvenient in batch mode, log files and reports which don't handle \r. It was too difficult to detect where fread is being called from, plus, removing it speeds up fread a little by saving code inside the C for loop (which is why it wasn't made optional instead). Use your operating system's system monitor to confirm fread is progressing. Thanks to Baptiste for highlighting : http://stackoverflow.com/questions/15370993/strange-output-from-fread-when-called-from-knitr

    • colClasses has been added. Same character vector format as read.csv (may be named or unnamed), but additionally may be type list. Type list enables setting ranges of columns by numeric position. NOTE: colClasses is intended for rare overrides, not routine use.

    • fread now supports files larger than 4GB on 64bit Windows (#2767 thanks to Paul Harding) and files between 2GB and 4GB on 32bit Windows (#2655 thanks to Vishal). A C call to GetFileSize() needed to be GetFileSizeEx().

    • When input is the data as a character string, it is no longer truncated to your system's maximum path length, #2649. It was being passed through path.expand() even when it wasn't a filename. Many thanks to Timothee Carayol for the reproducible report. The limit should now be R's character string length limit (2^31-1 bytes = 2GB). Test added.

    • New argument 'skip' overrides automatic banner skipping. When skip>=0, 'autostart' is ignored and line skip+1 will be taken as the first data row, or column names according to header="auto"|TRUE|FALSE as usual. Or, skip="string" uses the first line containing "string" (chosen to be a substring of the column name row unlikely to appear earlier), inspired by read.xls in package gdata. Thanks to Gabor Grothendieck for these suggestions.

    • fread now stops reading if an empty line is encountered, with warning if any text exists after that such as a footer (the first line of which will be included in the warning message).

    • Now reads files that are open in Excel without having to close them first, #2661. And up to 5 attempts are made every 250ms on Windows as recommended here : http://support.microsoft.com/kb/316609.

    • "nan%" observed in output of fread(...,verbose=TRUE) timings are now 0% when fread takes 0.000 seconds.

    • An unintended 50,000 column limit in fread has been removed. Thanks to mpmorley for reporting. Test added. http://stackoverflow.com/questions/18449997/fread-protection-stack-overflow-error

  • unique() and duplicated() methods gain 'by' to allow testing for uniqueness using any subset of columns, not just the keyed columns (if keyed) or all columns (if not). By default by=key(dt) for backwards compatibility. ?duplicated has been revised and tests added. Thanks to Arunkumar Srinivasan, Ricardo Saporta, and Frank Erickson for useful discussions.

  • CJ() is 90% faster on 1e6 rows (for example), #4849. The inputs are now sorted first before combining rather than after combining and uses rep.int instead of rep (thanks to Sean Garborg for the ideas, code and benchmark) and only sorted if is.unsorted(), #2321. Reminder: CJ = Cross Join; i.e., joins to all combinations of its inputs.

  • CJ() gains 'sorted' argument, by default TRUE for backwards compatibility. FALSE retains input order and is faster to create the result of CJ() but then slower to join from since unkeyed.

  • New function address() returns the address in RAM of its argument. Sometimes useful in determining whether a value has been copied or not by R, programatically. http://stackoverflow.com/a/10913296/403310

  • merge no longer returns spurious NA row(s) when y is empty and all.y=TRUE (or all=TRUE), #2633. Thanks to Vinicius Almendra for reporting. Test added. http://stackoverflow.com/questions/15566250/merge-data-table-with-all-true-introduces-na-row-is-this-correct

  • rbind'ing data.tables containing duplicate, "" or NA column names now works, #2726 & #2384. Thanks to Garrett See and Arun Srinivasan for reporting. This also affected the printing of data.tables with duplicate column names since the head and tail are rbind-ed together internally.

  • rbind, cbind and merge on data.table should now work in packages that Import but do not Depend on data.table. Many thanks to a patch to .onLoad from Ken Williams, and related posts from Victor Kryukov : http://r.789695.n4.nabble.com/Import-problem-with-data-table-in-packages-tp4665958.html

  • Mixing adding and updating into one DT[, :=(existingCol=...,newCol=...), by=...] now works without error or segfault, #2778 and #2528. Many thanks to Arunkumar Srinivasan for reporting and for the nice reproducible examples. Tests added.

  • rbindlist() now binds factor columns correctly, #2650. Thanks to many for reporting. Tests added.

  • Deleting a (0-length) factor column using :=NULL on an empty data.table now works, #4809. Thanks to Frank Pinter for reporting. Test added. http://stackoverflow.com/questions/18089587/error-deleting-factor-column-in-empty-data-table

  • Writing FUN= in DT[,lapply(.SD,FUN=...),] now works, #4893. Thanks to Jan Wijffels for reporting and Arun for suggesting and testing a fix. Committed and test added. http://stackoverflow.com/questions/18314757/why-cant-i-used-fun-in-lapply-when-grouping-by-using-data-table

  • The slowness of transform() on data.table has been fixed, #2599. But, please use :=.

  • setkey(DT,Colname with spaces) now works, #2452. setkey(DT,"Colname with spaces") worked already.

  • mean() in j has been optimized since v1.8.2 (see NEWS below) but wasn't respecting na.rm=TRUE (the default). Many thanks to Colin Fang for reporting. Test added. http://stackoverflow.com/questions/18571774/data-table-auto-remove-na-in-by-for-mean-function


  • := on a null data.table now gives a clearer error message : "Cannot use := to add columns to a null data.table (no columns), currently. You can use := to add (empty) columns to an empty data.table (1 or more columns, all 0 length), though." rather than the untrue : "Cannot use := to add columns to an empty data.table, currently"

  • Misuse of := and :=() is now caught in more circumstances and gives a clearer and shorter error message : ":= and :=(...) are defined for use in j, once only and in particular ways. See help(":="). Check is.data.table(DT) is TRUE."

  • data.table(NULL) now prints "Null data.table (0 rows and 0 cols)" and FAQ 2.5 has been improved. Thanks to: http://stackoverflow.com/questions/15317536/is-null-does-not-work-on-null-data-table-in-r-possible-bug

  • The braces {} have been removed from rollends's default, to solve a trace() problem. Thanks to Josh O'Brien's investigation : http://stackoverflow.com/questions/15931801/why-does-trace-edit-true-not-work-when-data-table

  • Tests 617,646 and 647 could sometimes fail (e.g. r-prerel-solaris-sparc on 7 Mar 2013) due to machine tolerance. Fixed.

  • The default for datatable.alloccol has changed from max(100L, 2L*ncol(DT)) to max(100L, ncol(DT)+64L). And a pointer to ?truelength has been added to an error message as suggested and thanks to Roland : http://stackoverflow.com/questions/15436356/potential-problems-from-over-allocating-truelength-more-than-1000-times

  • For packages wishing to use data.table optionally (e.g. according to user of that package) and therefore not wishing to Depend on data.table (which is the normal determination of data.table-awareness via .Depends), .datatable.aware may be set to TRUE in such packages which cedta() will look for, as before. But now it doesn't need to be exported. Thanks to Hadley Wickham for the suggestion and solution.

  • There are now 1,009 raw tests, as reported by test.data.table().

  • Welcome to Arunkumar Srinivasan and Ricardo Saporta who have joined the project and contributed directly by way of commits above.

  • v1.8.9 was on R-Forge only. v1.8.10 was released to CRAN. Odd numbers are development, evens on CRAN.

*   New function fread(), a fast and friendly file reader.
    *  header, skip, nrows, sep and colClasses are all auto detected.
    *  integers>2^31 are detected and read natively as bit64::integer64.
    *  accepts filenames, URLs and "A,B\n1,2\n3,4" directly
    *  new implementation entirely in C
    *  with a 50MB .csv, 1 million rows x 6 columns :
         read.csv("test.csv")                                        # 30-60 sec (varies)
         read.table("test.csv",<all known tricks and known nrows>)   #    10 sec
         fread("test.csv")                                           #     3 sec
    * airline data: 658MB csv (7 million rows x 29 columns)
         read.table("2008.csv",<all known tricks and known nrows>)   #   360 sec
         fread("2008.csv")                                           #    40 sec
    See ?fread. Many thanks to Chris Neff, Garrett See, Hideyoshi Maeda, Patrick
    Nic, Akhil Behl and Aykut Firat for ideas, discussions and beta testing.
    ** The fread function is still under development; e.g., dates are read as
    ** character and embedded quotes ("\"" and """") cause problems.

*   New argument 'allow.cartesian' (default FALSE) added to X[Y] and merge(X,Y), #2464.
    Prevents large allocations due to misspecified joins; e.g., duplicate key values in Y
    joining to the same group in X over and over again. The word 'cartesian' is used loosely
    for when more than max(nrow(X),nrow(Y)) rows would be returned. The error message is
    verbose and includes advice. Thanks to a question by Nick Clark, help from user1935457
    and a detailed reproducible crash report from JR.
    If the new option affects existing code you can set :
    to restore the previous behaviour until you have time to address.

*   In addition to TRUE/FALSE, 'roll' may now be a positive number (roll forwards/LOCF) or
    negative number (roll backwards/NOCB). A finite number limits the distance a value is
    rolled (limited staleness). roll=TRUE and roll=+Inf are equivalent.
    'rollends' is a new parameter holding two logicals. The first observation is rolled
    backwards if rollends[1] is TRUE. The last observation is rolled forwards if rollends[2]
    is TRUE. If roll is a finite number, the same limit applies to the ends.
    New value roll='nearest' joins to the nearest value (either backwards or forwards) when
    the value falls in a gap, and to the end value according to 'rollends'.
    'rolltolast' has been deprecated. For backwards compatibility it is converted to
    This implements [r-forge FR#2300](https://github.com/Rdatatable/data.table/issues/615) & [r-forge FR#206](https://github.com/Rdatatable/data.table/issues/459) and helps several recent S.O. questions.
*   setnames(DT,c(NA,NA)) is now a type error rather than a segfault, #2393.
    Thanks to Damian Betebenner for reporting.

*   rbind() no longers warns about inputs having columns in a different order
    if use.names has been explicitly set TRUE, #2385. Thanks to Simon Judes
    for reporting.

*   := by group with 0 length RHS could crash in some circumstances. Thanks to
    Damien Challet for the reproducible example using obfuscated data and
    pinpointing the version that regressed. Fixed and test added.

*   Error 'attempting to roll join on a factor column' could occur when a non last
    join column was a factor column, #2450. Thanks to Blue Magister for

*   NA in a join column of type double could cause both X[Y] and merge(X,Y)
    to return incorrect results, #2453. Due to an errant x==NA_REAL in the C source
    which should have been ISNA(x). Support for double in keyed joins is a relatively
    recent addition to data.table, but embarrassing all the same. Fixed and tests added.
    Many thanks to statquant for the thorough and reproducible report :

*   setnames() of all column names (such as setnames(DT,toupper(names(DT)))) failed on a
    keyed table where columns 1:length(key) were not the key. Fixed and test added.

*   setkey could sort 'double' columns (such as POSIXct) incorrectly when not the
    last column of the key, #2484. In data.table's C code :
        x[a] > x[b]-tol
    should have been :
        x[a]-x[b] > -tol  [or  x[b]-x[a] < tol ]
    The difference may have been machine/compiler dependent. Many thanks to statquant
    for the short reproducible example. Test added.

*   cbind(DT,1:n) returned an invalid data.table (some columns were empty) when DT
    had one row, #2478. Grouping now warns if j evaluates to an invalid data.table,
    to aid tracing root causes like this in future. Tests added. Many thanks to
    statquant for the reproducible example revealed by his interesting solution
    and to user1935457 for the assistance :

*   merge(...,all.y=TRUE) was 'setcolorder' error if a y column name included a space
    and there were rows in y not in x, #2555. The non syntactically valid column names
    are now preserved as intended. Thanks to Simon Judes for reporting. Tests added.

*   An error in := no longer suppresses the next print, #2376; i.e.,
        > DT[,foo:=colnameTypo+1]
        Error: object 'colnameTypo' not found
        > DT    # now prints DT ok
        > DT    # used to have to type DT a second time to see it
    Many thanks to Charles, Joris Meys, and, Spacedman whose solution is now used
    by data.table internally (http://stackoverflow.com/a/13606880/403310).
*   print(DT,topn=2), where topn is provided explicitly, now prints the top and bottom 2 rows
    even when nrow(x)<100 [options()$datatable.print.nrows]. And the 'topn' argument is now first
    for easier interactive use: print(DT,2), head(DT,2) and tail(DT,2).

*   The J() alias is now removed *outside* DT[...], but will still work inside DT[...];
    i.e., DT[J(...)] is fine. As warned in v1.8.2 (see below in this file) and deprecated
    with warning() in v1.8.4. This resolves the conflict with function J() in package
    XLConnect (#1747) and rJava (#2045).
    Please use data.table() directly instead of J(), outside DT[...].

*   ?merge.data.table and FAQ 1.12 have been improved (#2457), and FAQ 2.24 added.
    Thanks to dnlbrky for highlighting : http://stackoverflow.com/a/14164411/403310.

*   There are now 943 raw tests, as reported by test.data.table().

*   v1.8.7 was on R-Forge only. v1.8.8 was released to CRAN.
    Odd numbers are development, evens on CRAN.
*   A variable in calling scope was not found when combining i, j and by in
    one query, i used that local variable, and that query occurred inside a
    function, #2368. This worked in 1.8.2, a regression. Test added.
*   setnames used paste0() to construct its error messages, a function
    added to R 2.15.0. Reverted to use paste(). Tests added.

*   X[Y] where Y is empty (test 764) failed due to reliance on a pmin()
    enhancement in R 2.15.1. Removed reliance.
*   test.data.table() now passes in 2.12.0, the stated dependency, as well as
    2.14.0, 2.15.0, 2.15.1, 2.15.2 and R-devel.

*   Full R CMD check (i.e. including compatibility tests with the 9 Suggest-ed
    packages and S4 tests run using testthat which in turn depends on packages
    which depend on R >= 2.14.0) passes ok in 2.14.0 onwards.

*   There are now 876 raw tests, as reported by test.data.table().

*   v1.8.5 was on R-Forge only. v1.8.6 was released to CRAN.
    Odd numbers are development, evens on CRAN.
*   New printing options have been added :
    If the table to be printed has more than nrows, the top and bottom topn rows
    are printed. Otherwise, below nrows, the entire table is printed.
    Thanks to Allan Engelhardt and Melanie Bacou for useful discussions :
    and see FAQs 2.11 and 2.22.

*   When one or more rows in i have no match to x and i is unkeyed, i is now
    tested to see if it is sorted. If so, the key is retained. As before, when all rows of
    i match to x, the key is retained if i matches to an ordered subset of keyed x without
    needing to test i, even if i is unkeyed.

*   by on a keyed empty table is now keyed by the by columns, for consistency with
    the non empty case when an ordered grouping is detected.

*   DT[,`:=`(col1=val1, col2=val2, ...)] is now valid syntax rather than a crash, #2254.
    Many thanks to Thell Fowler for the suggestion.

*   with=FALSE is no longer needed to use column names or positions on the LHS of :=, #2120.
        DT[,c("newcol","existingcol"):=list(1L,NULL)]   # with=FALSE not needed
        DT[,`:=`(newcol=1L, existingcol:=NULL)]         # same
    If the LHS is held in a variable, the followed equivalent options are retained :
        mycols = c("existingcol","newcol")
        DT[,eval(mycols):=1L]                # same
        DT[,mycols:=1L,with=FALSE]           # same
        DT[,c("existingcol","newcol"):=1L]   # same (with=FALSE not needed)

*   Multiple LHS:= and `:=`(...) now work by group, and by without by. Implementing
    or fixing, and thanks to, #2215 (Florian Oswald), #1710 (Farrel Buchinsky) and
        DT[,c("newcol1","newcol2"):=list(mean(col1),sd(col1)), by=grp]
                 ...),  by=grp]                        # same but easier to read
        DT[c("grp1","grp2"), `:=`(newcol1=mean(col1),
                                  newcol2=sd(col1))]   # same using by-without-by

*   with=FALSE now works with a symbol LHS of :=, by group (#2120) :
        colname = "newcol"
    Thanks to Alex Chernyakov :

*   .GRP is a new symbol available to j. Value 1 for the first group, 2 for the 2nd, etc. Thanks
    to Josh O'Brien for the suggestion :

*   .I is a new symbol available to j. An integer vector length .N. It contains the group's row
    locations in DT. This implements FR#1962.
        DT[,.I[which.max(colB)],by=colA]      # row numbers of maxima by group

*   A new "!" prefix on i signals 'not-join' (a.k.a. 'not-where'), #1384i.
        DT[-DT["a", which=TRUE, nomatch=0]]   # old not-join idiom, still works
        DT[!"a"]                              # same result, now preferred.
        DT[!J(6),...]                         # !J == not-join
        DT[!2:3,...]                          # ! on all types of i
        DT[colA!=6L | colB!=23L,...]          # multiple vector scanning approach (slow)
        DT[!J(6L,23L)]                        # same result, faster binary search
    '!' has been used rather than '-' :
        * to match the 'not-join'/'not-where' nomenclature
        * with '-', DT[-0] would return DT rather than DT[0] and not be backwards
          compatible. With '!', DT[!0] returns DT both before (since !0 is TRUE in
          base R) and after this new feature.
        * to leave DT[+J...] and DT[-J...] available for future use

*   When with=FALSE, "!" may also be a prefix on j, #1384ii. This selects all but the named columns.
        DF[,-match("somecol",names(DF))]              # works when somecol exists. If not, NA causes an error.
        DF[,-match("somecol",names(DF),nomatch=0)]    # works when somecol exists. Empty data.frame when it doesn't, silently.
        DT[,-match("somecol",names(DT)),with=FALSE]   # same issues.
        DT[,setdiff(names(DT),"somecol"),with=FALSE]  # works but you have to know order of arguments, and no warning if doesn't exist
        - vs -
        DT[,!"somecol",with=FALSE]                    # works and easy to read. With (helpful) warning if somecol isn't there.
    Strictly speaking, this (!j) is a "not-select" (!i is 'not-where'). This has no analogy in SQL.
    Reminder: i is analogous to WHERE, j is analogous to SELECT and `:=` in j changes SELECT to UPDATE.
    !j when j is column positions is very similar to -j.
        DF[,-(2:3),drop=FALSE]           # all but columns 2 and 3. Careful, brackets and drop=FALSE are required.
        DT[,-(2:3),with=FALSE]           # same
        DT[,!2:3,with=FALSE]             # same
        copy(DT)[,2:3:=NULL,with=FALSE]  # same
    !j was introduced for column names really, not positions. It works for both, for consistency :
        toremove = c("somecol","anothercol")
        toremove = 2:3
        DT[,!toremove,with=FALSE]        # same code works without change

*   'which' now accepts NA. This means return the row numbers in i that don't match, #1384iii.
    Thanks to Santosh Srinivas for the suggestion.
        X[Y,which=TRUE]   # row numbers of X that do match, as before
        X[!Y,which=TRUE]  # row numbers of X that don't match
        X[Y,which=NA]     # row numbers of Y that don't match
        X[!Y,which=NA]    # row numbers of Y that do match (for completeness)

*   setnames() now works on data.frame, #2273. Thanks to Christian Hudon for the suggestion.
*   A large slowdown (many minutes instead of a few secs) in X[Y] joins has been fixed, #2216.
    This occurred where the number of rows in i was large, and at least one row joined to
    more than one row in x. Possibly in other similar circumstances too. The workaround was
    to set mult="first" which is no longer required. Test added.
    Thanks to a question and report from Alex Chernyakov :

*   Indexing columns of data.table with a logical vector and `with=FALSE` now works as
    expected, fixing #1797. Thanks to Mani Narayanan for reporting. Test added.

*   In X[Y,cols,with=FALSE], NA matches are now handled correctly. And if cols
    includes join columns, NA matches (if any) are now populated from i. For
    consistency with X[Y] and X[Y,list(...)]. Tests added.

*   "Internal error" when combining join containing missing groups and group by
    is fixed, #2162. For example :
    where Y contains some rows that don't match to X. This bug could also result in a segfault.
    Thanks to Andrey Riabushenko and Michael Schermerhorn for reporting. Tests added.

*   On empty tables, := now changes column type and adds new 0 length columns ok, fixing
    #2274. Tests added.

*   Deleting multiple columns out-of-order is no longer a segfault, #2223. Test added.
    Reminder: deleting columns by reference is relatively instant, regardless of table size.

*   Mixing column adds and deletes in one := gave incorrect results, #2251. Thanks to
    Michael Nelson for reporting. Test added.
        DT[,`:=`(newcol=col1+1L,col1=NULL)]             # same

*   Out of bound positions in the LHS of := are now caught. Root cause of crash in #2254.
    Thanks to Thell Fowler for reporting. Tests added.
        DT[,(ncol(DT)+1):=1L]   # out of bounds error (add new columns by name only)
        DT[,ncol(DT):=1L]       # ok

*   A recycled column plonk RHS of := no longer messes up setkey and := when used on that
    object afterwards, #2298. For example,
        DT = data.table(a=letters[3:1],x=1:3)
        DT[,c("x1","x2"):=x]  # ok (x1 and x2 are now copies of x)
        setkey(DT,a)          # now ok rather than wrong result
    Thanks to Timothee Carayol for reporting. Tests added.

*   Join columns are now named correctly when j is .SD, a subset of .SD, or similar, #2281.
        DT[c("a","b"),.SD[...]]    # result's first column now named key(DT)[1] rather than 'V1'

*   Joining an empty i table now works without error (#2194). It also retains key and has a consistent
    number and type of empty columns as the non empty by-without-by case. Tests added.

*   by-without-by with keyed i where key isn't the 1:n columns of i could crash, #2314. Many thanks
    to Garrett See for reporting with reproducible example data file. Tests added.

*   DT[,col1:=X[Y,col2]] was a crash, #2311. Due to RHS being a data.table. mult="first"
    (or drop=TRUE in future) was likely intended. Thanks to Anoop Shah for reporting with
    reproducible example. Root cause (recycling of list columns) fixed and tests added.

*   Grouping by a column which somehow has names, no longer causes an error, #2307.
      DT = data.table(a=1:3,b=c("a","a","b"))
      setattr(DT$b, "names", c("a","b","c"))  # not recommended, just to illustrate
      DT[,sum(a),by=b]  # now ok

*   gWidgetsWWW wasn't known as data.table aware, even though it mimicks executing
    code in .GlobalEnv, #2340. So, data.table is now gWidgetsWWW-aware. Further packages
    can be added if required by changing a new variable :
    by using assignInNamespace(). Thanks to Zach Waite and Yihui Xie for investigating and
    providing reproducible examples :

*   Optimization of lapply when FUN is a character function name now works, #2212.
        DT[,lapply(.SD, "+", 1), by=id]  # no longer an error
        DT[,lapply(.SD, `+`, 1), by=id]  # same, worked before
    Thanks to Michael Nelson for highlighting. Tests added.

*   Syntactically invalid column names (such as "Some rate (%)") are now preserved in X[Y] joins and
    merge(), as intended. Thanks to George Kaupas (#2193i) and Yang Zhang (#2090) for reporting.
    Tests added.

*   merge() and setcolorder() now check for duplicate column names first rather than a less helpful
    error later, #2193ii. Thanks to Peter Fine for reporting. Tests added.

*   Column attributes (such as 'comment') are now retained by X[Y] and merge(), #2270. Thanks to
    Allan Engelhardt for reporting. Tests added.

*   A matrix RHS of := is now treated as vector, with warning if it has more than 1 column, #2333.
    Thanks to Alex Chernyakov for highlighting. Tests added.
        DT[,b:=scale(a)]   # now works rather than creating an invalid column of type matrix

*   last() is now S3 generic for compatibility with xts::last, #2312. Strictly speaking, for speed,
    last(x) deals with vector, list and data.table inputs directly before falling back to
    S3 dispatch. Thanks to Garrett See for reporting. Tests added.

*   DT[,lapply(.SD,sum)] in the case of no grouping now returns a data.table for consistency, rather
    than list, #2263. Thanks to Justin and mnel for highlighting. Existing test changed.

*   L[[2L]][,newcol:=] now works, where L is a list of data.table objects, #2204. Thanks to Melanie Bacou
    for reporting. Tests added. A warning is issued when the first column is added if L was created with
    list(DT1,DT2) since R's list() copies named inputs. Until reflist() is implemented, this warning can be
    ignored or suppressed.

*   DT[J(data.frame(...))] now works again, giving the same result as DT[data.frame(...)], #2265.
    Thanks to Christian Hudon for reporting. Tests added.

*   A memory leak has been fixed, #2191 and #2284. All data.table objects leaked the over allocated column
    pointer slots; i.e., when a data.table went out of scope or was rm()'d this memory wasn't released and
    gc() would report growing Vcells. For a 3 column data.table with a 100 allocation, the growth was
    1.5Kb per data.table on 64bit (97*8*2 bytes) and 0.75Kb on 32bit (97*4*2 bytes).
    Many thanks to Xavier Saint-Mleux and Sasha Goodman for the reproducible examples and
    assistance. Tests added.

*   rbindlist now skips empty (0 row) items as well as NULL items. So the column types of the result are
    now taken from the first non-empty data.table. Thanks to Garrett See for reporting. Test added.

*   setnames did not update column names correctly when passed integer column positions and those
    column names contained duplicates, fixed. This affected the column names of queries involving
    two or more by expressions with a named list inside {}. Thanks to Steve Lianoglou for finding and
    fixing. Tests added.
        DT[, {list(name1=sum(v),name2=sum(w))}, by="a,b"]  # now ok, no blank column names in result
        DT[, list(name1=sum(v),name2=sum(w)), by="a,b"]    # ok before
*   J() now issues a warning (when used *outside* DT[...]) that using it
    outside DT[...] is deprecated. See item below in v1.8.2.
    Use data.table() directly instead of J(), outside DT[...]. Or, define
    an alias yourself. J() will continue to work *inside* DT[...] as documented.

*   DT[,LHS:=RHS,...] no longer prints DT. This implements #2128 "Try again to get
    DT[i,j:=value] to return invisibly". Thanks to discussions here :
    FAQs 2.21 and 2.22 have been updated.

*   DT[] now returns DT rather than an error that either i or j must be supplied.
    So, ending with [] at the console is a convenience to print the result of :=, rather
    than wrapping with print(); e.g.,
        DT[i,j:=value]...oops forgot print...[]
    is the same as :

*   A warning is now issued when by is set equal to the by-without-by join columns,
    causing x to be subset and then grouped again. The warning suggests removing by or
    changing it, #2282. This can be turned off using options(datatable.warnredundantby=FALSE)
    in case it occurs after upgrading, until those lines can be modified.
    Thanks to Ben Barnes for highlighting :

*   Description of how join columns are determined in X[Y] syntax has been further clarified
    in ?data.table. Thanks to Alex :

*   ?transform and example(transform) has been fixed and embelished, #2316.
    Thanks to Garrett See's suggestion.

*   ?setattr has been updated to document that it takes any input, not just data.table, and
    can be used on columns of a data.frame, for example.

*   Efficiency warnings when joining between a factor column and a character column are now downgraded
    to messages when verbosity is on, #2265i. Thanks to Christian Hudon for the suggestion.
*   Combining a join with mult="first"|"last" followed by by inside the same [...] gave incorrect
    results or a crash, #2303. Many thanks to Garrett See for the reproducible example and
    pinpointing in advance which commit had caused the problem. Tests added.

*   Examples in ?data.table have been updated now that := no longer prints. Thanks to Garrett See.
*   There are now 869 raw tests. test.data.table() should return precisely this number of
    tests passed. If not, then somehow, a slightly stale version from R-Forge is likely
    installed; please reinstall from CRAN.

*   v1.8.3 was an R-Forge only beta release. v1.8.4 was released to CRAN.
*   Numeric columns (type 'double') are now allowed in keys and ad hoc
    by. J() and SJ() no longer coerce 'double' to 'integer'. i join columns
    which mismatch on numeric type are coerced silently to match
    the type of x's join column. Two floating point values
    are considered equal (by grouping and binary search joins) if their
    difference is within sqrt(.Machine$double.eps), by default. See example
    in ?unique.data.table. Completes FRs #951, #1609 and #1075. This paves the
    way for other atomic types which use 'double' (such as POSIXct and bit64).
    Thanks to Chris Neff for beta testing and finding problems with keys
    of two numeric columns (bug #2004), fixed and tests added.

*   := by group is now implemented (FR#1491) and sub-assigning to a new column
    by reference now adds the column automatically (initialized with NA where
    the sub-assign doesn't touch) (FR#1997). := by group can be combined with all
    types of i, so ":= by group" includes grouping by `i` as well as by `by`.
    Since := by group is by reference, it should be significantly faster than any
    method that (directly or indirectly) `cbind`s the grouped results to DT, since
    no copy of the (large) DT is made at all. It's a short and natural syntax that
    can be compounded with other queries.

*   Prettier printing of list columns. The first 6 items of atomic vectors
    are collapsed with "," followed by a trailing "," if there are more than
    6, FR#1608. This difference to data.frame has been added to FAQ 2.17.
    Embedded objects (such as a data.table) print their class name only to avoid
    seemingly mangled output, bug #1803. Thanks to Yike Lu for reporting.
    For example:
    > data.table(x=letters[1:3],
                 y=list( 1:10, letters[1:4], data.table(a=1:3,b=4:6) ))
       x            y
    1: a 1,2,3,4,5,6,
    2: b      a,b,c,d
    3: c <data.table>

*   Warnings added when joining character to factor, and factor to character.
    Character to character is now preferred in joins and needs no coercion.
    Even so, these coercions have been made much more efficient by taking
    a shallow copy of i internally, avoiding a full deep copy of i.

*   Ordered subsets now retain x's key. Always for logical and keyed i, using
    base::is.unsorted() for integer and unkeyed i. Implements FR#295.

*   mean() is now automatically optimized, #1231. This can speed up grouping
    by 20 times when there are a large number of groups. See wiki point 3, which
    is no longer needed to know. Turn off optimization by setting

*   DT[,lapply(.SD,...),by=...] is now automatically optimized, #2067. This can speed
    up applying a function by column by group, by over 20 times. See wiki point 5
    which is no longer needed to know. In other words:
    is now just as fast as :
    Don't forget to use .SDcols when a subset of columns is needed.

*   The package is now Byte Compiled (when installed in R 2.14.0 or later). Several
    internal speed improvements were made in this version too, such as avoiding
    internal copies. If you find 1.8.2 is faster, before attributing that to Byte
    Compilation, please install the package without Byte Compilation and compare
    ceteris paribus. If you find cases where speed has slowed, please let us know.

*   sapply(DT,class) gets a significant speed boost by avoiding a call to unclass()
    in as.list.data.table() called by lapply(DT,...), which copied the entire object.
    Thanks to a question by user1393348 on Stack Overflow, implementing #2000.

*   The J() alias is now deprecated outside DT[...], but will still work inside
    DT[...], as in DT[J(...)].
    J() is conflicting with function J() in package XLConnect (#1747)
    and rJava (#2045). For data.table to change is easier, with some efficiency
    advantages too. The next version of data.table will issue a warning from J()
    when used outside DT[...]. The version after will remove it. Only then will
    the conflict with rJava and XLConnect be resolved.
    Please use data.table() directly instead of J(), outside DT[...].

*   New DT[.(...)] syntax (in the style of package plyr) is identical to
    DT[list(...)], DT[J(...)] and DT[data.table(...)]. We plan to add ..(), too, so
    that .() and ..() are analogous to the file system's ./ and ../; i.e., .()
    evaluates within the frame of DT and ..() in the parent scope.

*   New function rbindlist(l). This does the same as do.call("rbind",l), but much
*   DT[,f(.SD),by=colA] where f(x)=x[,colB:=1L] was a segfault, bug#1727.
    This is now a graceful error to say that using := in .SD's j is
    reserved for future use. This was already caught in most circumstances,
    other than via f(.SD). Thanks to Leon Baum for reporting. Test added.

*   If .N is selected by j it is now renamed "N" (no dot) in the output, to
    avoid a potential conflict in subsequent grouping between a column called
    ".N" and the special .N variable, fixing #1720. ?data.table updated and
    FAQ 4.6 added with detailed examples. Tests added.

*   Moved data.table setup code from .onAttach to .onLoad so that it
    is also run when data.table is simply `import`ed from within a package,
    fixing #1916 related to missing data.table options.

*   Typos fixed in ?":=", thanks to Michael Weylandt for reporting.

*   base::unname(DT) now works again, as needed by plyr::melt(). Thanks to
    Christoph Jaeckel for reporting. Test added.

*   CJ(x=...,y=...) now retains the column names x and y, useful when CJ
    is used independently (since x[CJ(...)] takes join column names from x).
    Restores behaviour lost somewhere between 1.7.1 and 1.8.0, thanks
    to Muhammad Waliji for reporting. Tests added.

*   A column plonk via set() was only possible by passing NULL as i. The default
    for i is now NULL so that missing i invokes a column plonk, too (when length(value)
    == nrow(DT)). A column plonk is much more efficient than creating 1:nrow(DT) and
    passing that as i to set() or DT[i,:=] (almost infinitely faster). Thanks to
    testing by Josh O'Brien in comments on Stack Overflow. Test added.

*   Joining a factor column with unsorted and unused levels to a character column
    now matches properly, fixing #1922. Thanks to Christoph Jäckel for the reproducible
    example. Test added.

*   'by' on an empty table now returns an empty table (#1945) and .N, .SD and .BY are
    now available in the empty case (also #1945). The column names and types of
    the returned empty table are consistent with the non empty case. Thanks to
    Malcolm Cook for reporting. Tests added.

*   DT[NULL] now returns the NULL data.table, rather than an error. Test added.
    Use DT[0] to return an empty copy of DT.

*   .N, .SD and .BY are now available to j when 'by' is missing, "", character()
    and NULL, fixing #1732. For consistency so that j works unchanged when by is
    dynamic and passed one of those values all meaning 'don't group'. Thanks
    to Joseph Voelkel reporting and Chris Neff for further use cases. Tests added.

*   chorder(character()) was a seg fault, #2026. Fixed and test added.

*   When grouping by i, if the first row of i had no match, .N was 1 rather than 0.
    Fixed and tests added. Thanks to a question by user1165199 on Stack Overflow :

*   All object attributes are now retained by grouping; e.g., tzone of POSIXct is no
    longer lost, fixing #1704. Test added. Thanks to Karl Ove Hufthammer for reporting.

*   All object attributes are now retained by recycling assign to a new column (both
    <- and :=); e.g., POSIXct class is no longer lost, fixing #1712. Test added. Thanks
    to Leon Baum for reporting.

*   unique() of ITime no longer coerces to integer, fixing #1719. Test added.

*   rbind() of DT with an irregular list() now recycles the list items correctly,
    #2003. Test added.

*   setcolorder() now produces correct error when passed missing column names. Test added.

*   merge() with common names, and, all.y=TRUE (or all=TRUE) no longer returns an error, #2011.
    Tests added. Thanks to a question by Ina on Stack Overflow :

*   Removing or setting datatable.alloccol to NULL is no longer a memory leak, #2014.
    Tests added. Thanks to a question by Vanja on Stack Overflow :

*   DT[,2:=someval,with=FALSE] now changes column 2 even if column 1 has the same (duplicate)
    name, #2025. Thanks to Sean Creighton for reporting. Tests added.

*   merge() is now correct when all=TRUE but there are no common values in the two
    data.tables, fixing #2114. Thanks to Karl Ove Hufthammer for reporting.  Tests added.

*   An as.data.frame method has been added for ITime, so that ITime can be passed to ggplot2
    without error, #1713. Thanks to Farrel Buchinsky for reporting. Tests added.
    ITime axis labels are still displayed as integer seconds from midnight; we don't know why ggplot2
    doesn't invoke ITime's as.character method. Convert ITime to POSIXct for ggplot2, is one approach.

*   setnames(DT,newnames) now works when DT contains duplicate column names, #2103.
    Thanks to Timothee Carayol for reporting. Tests added.

*   subset() would crash on a keyed table with non-character 'select', #2131. Thanks
    to Benjamin Barnes for reporting. The root cause was non character inputs to chmatch
    and %chin%. Tests added.

*   Non-ascii column names now work when passed as character 'by', #2134. Thanks to
    Karl Ove Hufthammer for reporting. Tests added.
        DT[, mean(foo), by=ÆØÅ]      # worked before
        DT[, mean(foo), by="ÆØÅ"]    # now works too
        DT[, mean(foo), by=colA]     # worked before
        DT[, mean(foo), by="colA"]   # worked before
*   Incorrect syntax error message for := now includes advice to check that
    DT is a data.table rather than a data.frame. Thanks to a comment by
    gkaupas on Stack Overflow.

*   When set() is passed a logical i, the error message now includes advice to
    wrap with which() and take the which() outside the loop (if any) if possible.

*   An empty data.table (0 rows, 1+ cols) now print as "Empty data.table" rather
    than "NULL data.table". A NULL data.table, returned by data.table(NULL) has
    0 rows and 0 cols.  DT[0] returns an empty data.table.

*   0 length by (such as NULL and character(0)) now return a data.table when j
    is vector, rather than vector, for consistency of return types when by
    is dynamic and 'dont group' needs to be represented. Bug fix #1599 in
    v1.7.0 was fixing an error in this case (0 length by).

*   Default column names for unnamed columns are now consistent between 'by' and
    non-'by'; e.g. these two queries now name the columns "V1" and "V2" :
        DT[,list("a","b")]      # used to name the columns 'a' and 'b', oddly.

*   Typing ?merge now asks whether to display ?merge.data.frame or ?merge.data.table,
    and ?merge.data.table works directly. Thanks to Chris Neff for suggesting.

*   Description of how join columns are determined in X[Y] syntax has been clarified
    in ?data.table. Thanks to Juliet Hannah and Yike Lu.

*   DT now prints consistent row numbers when the column names are reprinted at the
    bottom of the output (saves scrolling up). Thanks to Yike Lu for reporting #2015.
    The tail as well as the head of large tables is now printed.
*   Florian Oswald for #2094: DT[,newcol:=NA] now adds a new logical column ok.
    Test added.

*   A large slow down (2s went up to 40s) when iterating calls to DT[...] in a
    for loop, such as in example(":="), was caught and fixed in beta, #2027.
    Speed regression test added.

*   Christoph Jäckel for #2078: by=c(...) with i clause broke. Tests added.

*   Chris Neff for #2065: keyby := now keys, unless, i clause is present or
    keyby is not straightforward column names (in any format). Tests added.

*   :=NULL to delete, following by := by group to add, didn't add the column,
    #2117. Test added.

*   Combining i subset with by gave incorrect results, #2118. Tests added.

*   Benjamin Barnes for #2133: rbindlist not supporting type 'logical'.
    Tests added.

*   Chris Neff for #2146: using := to add a column to the result of a simple
    column subset such as DT[,list(x)], or after changing all column names
    with setnames(), was an error. Fixed and tests added.
*   There are now 717 raw tests, plus S4 tests.

*   v1.8.1 was an R-Forge only beta release. v1.8.2 was released to CRAN.
*   character columns are now allowed in keys and are preferred to
    factor. data.table() and setkey() no longer coerce character to
    factor. Factors are still supported. Implements FR#1493, FR#1224
    and (partially) FR#951.

*   setkey() no longer sorts factor levels. This should be more convenient
    and compatible with ordered factors where the levels are 'labels', in
    some order other than alphabetical. The established advice to paste each
    level with an ordinal prefix, or use another table to hold the factor
    labels instead of a factor column, is no longer needed. Solves FR#1420.
    Thanks to Damian Betebenner and Allan Engelhardt raising on datatable-help
    and their tests have been added verbatim to the test suite.

*   unique(DT) and duplicated(DT) are now faster with character columns,
    on unkeyed tables as well as keyed tables, FR#1724.

*   New function set(DT,i,j,value) allows fast assignment to elements
    of DT. Similar to := but avoids the overhead of [.data.table, so is
    much faster inside a loop. Less flexible than :=, but as flexible
    as matrix subassignment. Similar in spirit to setnames(), setcolorder(),
    setkey() and setattr(); i.e., assigns by reference with no copy at all.

        M = matrix(1,nrow=100000,ncol=100)
        DF = as.data.frame(M)
        DT = as.data.table(M)
        system.time(for (i in 1:1000) DF[i,1L] <- i)   # 591.000s
        system.time(for (i in 1:1000) DT[i,V1:=i])     #   1.158s
        system.time(for (i in 1:1000) M[i,1L] <- i)    #   0.016s
        system.time(for (i in 1:1000) set(DT,i,1L,i))  #   0.027s

*   New functions chmatch() and %chin%, faster versions of match()
    and %in% for character vectors. R's internal string cache is
    utilised (no hash table is built). They are about 4 times faster
    than match() on the example in ?chmatch.

*   Internal function sortedmatch() removed and replaced with chmatch()
    when matching i levels to x levels for columns of type 'factor'. This
    preliminary step was causing a (known) significant slowdown when the number
    of levels of a factor column was large (e.g. >10,000). Exacerbated in
    tests of joining four such columns, as demonstrated by Wes McKinney
    (author of Python package Pandas). Matching 1 million strings of which
    of which 600,000 are unique is now reduced from 16s to 0.5s, for example.
    Background here :

*   rbind.data.table() gains a use.names argument, by default TRUE.
    Set to FALSE to combine columns in order rather than by name. Thanks to
    a question by Zach on Stack Overflow :

*   New argument 'keyby'. An ad hoc by just as 'by' but with an additional setkey()
    on the by columns of the result, for convenience. Not to be confused with a
    'keyed by' such as DT[...,by=key(DT)] which can be more efficient as explained
    by FAQ 3.3.  Thanks to Yike Lu for the suggestion and discussion (FR#1780).

*   Single by (or keyby) expressions no longer need to be wrapped in list(),
    for convenience, implementing FR#1743; e.g., these now works :
    instead of needing :

*   Unnamed 'by' expressions have always been inspected using all.vars() to make
    a guess at a sensible column name for the result. This guess now includes
    function names via all.vars(functions=TRUE), for convenience; e.g.,
    now returns a column called 'month' rather than 'date'. It is more robust to
    explicitly name columns, though; e.g.,
        DT[,sum(v),by=list("Guaranteed name"=month(date))]

*   For a surprising speed boost in some circumstances, default options such as
    'datatable.verbose' are now set when the package loads (unless they are already
    set, by user's profile for example). The 'default' argument of base::getOption()
    was the culprit and has been removed internally from all 11 calls.
*   Fixed a `suffixes` handling bug in merge.data.table that was
    only recently introduced during the recent "fast-merge"-ing reboot.
    Briefly, the bug was only triggered in scenarios where both
    tables had identical column names that were not part of `by` and
    ended with *.1. cf. "merge and auto-increment columns in y[x]"
    test in tests/test-data.frame-like.R for more information.

*   Adding a column using := on a data.table just loaded from disk was
    correctly detected and over allocated, but incorrectly warning about
    a previous copy. Test 462 tested loading from disk, but suppressed
    warnings (sadly). Fixed.

*   data.table unaware packages that use DF[i] and DF[i]<-value syntax
    were not compatible with data.table, fixed. Many thanks to Prasad Chalasani
    for providing a reproducible example with base::droplevels(), and
    Helge Liebert for providing a reproducible example (#1794) with stats::reshape().
    Tests added.

*   as.data.table(DF) already preserved DF's attributes but not any inherited
    classes such as nlme's groupedData, so nlme was incompatible with
    data.table. Fixed. Thanks to Dieter Menne for providing a reproducible
    example. Test added.

*   The internal row.names attribute of .SD (which exists for compatibility with
    data.frame only) was not being updated for each group. This caused length errors
    when calling any non-data.table-aware package from j, by group, when that package
    used length of row.names. Such as the recent update to ggplot2. Fixed.

*   When grouped j consists of a print of an object (such as ggplot2), the print is now
    masked to return NULL rather than the object that ggplot2 returns since the
    recent update v0.9.0. Otherwise data.table tries to accumulate the (albeit
    invisible) print object. The print mask is local to grouping, not generally.

*   'by' was failing (bug #1880) when passed character column names where one or more
    included a space. So, this now works :
        DT[,sum(v),by="column 1"]
    and j retains spaces in column names rather than replacing spaces with "."; e.g.,
        DT[,list("a b"=1)]
    Thanks to Yang Zhang for reporting. Tests added. As before, column names may be
    back ticked in the usual R way (in i, j and by); e.g.,
        DT[,sum(`nicely named var`+1),by=month(`long name for date column`)]

*   unique() on an unkeyed table including character columns now works correctly, fixing
    #1725. Thanks to Steven Bagley for reporting. Test added.

*   %like% now returns logical (rather than integer locations) so that it can be
    combined with other i clauses, fixing #1726. Thanks to Ivan Zhang for reporting. Test
*   Joshua Ulrich for spotting a missing PACKAGE="data.table"
    in .Call in setkey.R, and suggesting as.list.default() and
    unique.default() to avoid dispatch for speed, all implemented.
*   Providing .SDcols when j doesn't use .SD is downgraded from error to warning,
    and verbosity now reports which columns have been detected as used by j.

*   check.names is now FALSE by default, for convenience when working with column
    names with spaces and other special characters, which are now fully supported.
    This difference to data.frame has been added to FAQ 2.17.
*   New function setcolorder() reorders the columns by name
    or by number, by reference with no copy. This is (almost)
    infinitely faster than DT[,neworder,with=FALSE].

*   The prefix i. can now be used in j to refer to join inherited
    columns of i that are otherwise masked by columns in x with
    the same name.
*   tracemem() in example(setkey) was causing CRAN check errors
    on machines where R is compiled without memory profiling available,
    for efficiency. Notably, R for Windows, Ubuntu and Mac have memory
    profiling enabled which may slow down R on those architectures even
    when memory profiling is not being requested by the user. The call to
    tracemem() is now wrapped with try().

*   merge of unkeyed tables now works correctly after breaking in 1.7.8 and
    1.7.9. Thanks to Eric and DM for reporting. Tests added.

*   nomatch=0 was ignored for the first group when j used join inherited
    scope. Fixed and tests added.
*   Updating an existing column using := after a key<- now works without warning
    or error. This can be useful in interactive use when you forget to use setkey()
    but don't mind about the inefficiency of key<-. Thanks to Chris Neff for
    providing a convincing use case. Adding a new column uing := after key<-
    issues a warning, shallow copies and proceeds, as before.

*   The 'datatable.pre.suffixes' option has been removed. It was available to
    obtain deprecated merge() suffixes pre v1.5.4.
  • New function setnames(), referred to in 1.7.8 warning messages. It makes no copy of the whole data object, unlike names<- and colnames<-. It may be more convenient as well since it allows changing a column name, by name; e.g.,

    setnames(DT,"oldcolname","newcolname") # by name; no match() needed setnames(DT,3,"newcolname") # by position setnames(DT,2:3,c("A","B")) # multiple setnames(DT,c("a","b"),c("A","B")) # multiple by name setnames(DT,toupper(names(DT))) # replace all

    setnames() maintains truelength of the over-allocated names vector. This allows := to add columns fully by reference without growing the names vector. As before with names<-, if a key column's name is changed, the "sorted" attribute is updated with the new column name.

  • Incompatibility with reshape() of 3 column tables fixed (introduced by 1.7.8) : Error in setkey(ans, NULL) : x is not a data.table Thanks to Damian Betebenner for reporting and reproducible example. Tests added to catch in future.

  • setattr(DT,...) still returns DT, but now invisibly. It returns DT back again for compound syntax to work; e.g., setattr(DT,...)[i,j,by] Again, thanks to Damian Betebenner for reporting.

  • unique(DT) now works when DT is keyed and a key column is called 'x' (an internal scoping conflict introduced in v1.6.1). Thanks to Steven Bagley for reporting.

  • Errors and seg faults could occur in grouping when j contained character or list columns. Many thanks to Jim Holtman for providing a reproducible example.

  • Setting a key on a table with over 268 million rows (2^31/8) now works (again), #1714. Bug introduced in v1.7.2. setkey works up to the regular R vector limit of 2^31 rows (2 billion). Thanks to Leon Baum for reporting.

  • Checks in := are now made up front (before starting to modify the data.table) so that the data.table isn't left in an invalid state should an error occur, #1711. Thanks to Chris Neff for reporting.

  • The 'Chris crash' is fixed. The root cause was that key<- always copies the whole table. The problem with that copy (other than being slower) is that R doesn't maintain the over allocated truelength, but it looks as though it has. key<- was used internally, in particular in merge(). So, adding a column using := after merge() was a memory overwrite, since the over allocated memory wasn't really there after key<-'s copy.

    data.tables now have a new attribute '.internal.selfref' to catch and warn about such copies in future. All internal use of key<- has been replaced with setkey(), or new function setkeyv() which accepts a vector, and do not copy.

    Many thanks to Chris Neff for extended dialogue, providing a reproducible example and his patience. This problem was not just in pre 2.14.0, but post 2.14.0 as well. Thanks also to Christoph Jäckel, Timothée Carayol and DM for investigations and suggestions, which in combination led to the solution.

  • An example in ?":=" fixed, and j and by descriptions improved in ?data.table. Thanks to Joseph Voelkel for reporting.

  • Multiple new columns can be added by reference using := and with=FALSE; e.g., DT[,c("foo","bar"):=1L,with=FALSE] DT[,c("foo","bar"):=list(1L,2L),with=FALSE]

  • := now recycles vectors of non divisible length, with a warning (previously an error).

  • When setkey coerces a numeric or character column, it no longer makes a copy of the whole table, FR#1744. Thanks to an investigation by DM.

  • New function setkeyv(DT,v) (v stands for vector) replaces key(DT)<-v syntax. Also added setattr(). See ?copy.

  • merge() now uses (manual) secondary keys, for speed.

  • The loc argument of setkey has been removed. This wasn't very useful and didn't warrant a period of deprecation.

  • datatable.alloccol has been removed. That warning is now controlled by datatable.verbose=TRUE. One option is easer.

  • If i is a keyed data.table, it is no longer an error if its key is longer than x's key; the first length(key(x)) columns of i's key are used to join.

  • Previous bug fix for random crash in R <= 2.13.2 related to truelength and over-allocation didn't work, 3rd attempt. Thanks to Chris Neff for his patience and testing. This has shown up consistently as error status on CRAN old-rel checks (windows and mac). So if they pass, this issue is fixed.
  • An empty list column can now be added with :=, and data.table() accepts empty list(). DT[,newcol:=list()] data.table(a=1:3,b=list()) Empty list columns contain NULL for all rows.
  • Adding a column to a data.table loaded from disk could result in a memory corruption in R <= 2.13.2, revealed and thanks to CRAN checks on windows old-rel.

  • Adding a factor column with a RHS to be recycled no longer loses its factor attribute, #1691. Thanks to Damian Betebenner for reporting.

  • merge()-ing a data.table where its key is not the first few columns in order now works correctly and without warning, fixing #1645. Thanks to Timothee Carayol for reporting.

  • Mixing nomatch=0 and mult="last" (or "first") now works, #1661. Thanks to Johann Hibschman for reporting.

  • Join Inherited Scope now respects nomatch=0, #1663. Thanks to Johann Hibschman for reporting.

  • by= could generate a keyed result table with invalid key; e.g., when by= expressions return NA, #1631. Thanks to Muhammad Waliji for reporting.

  • Adding a column to a data.table loaded from disk resulted in an error that truelength(DT)<length(DT).

  • CJ() bogus values and logical error fixed, #1689. Thanks to Damian Betebenner and Chris Neff for reporting.

  • j=list(.SD,newcol=...) now gives friendly error suggesting cbind or merge afterwards until := by group is implemented, rather than treating .SD as a list column, #1647. Thanks to a question by Christoph_J on Stack Overflow.

  • rbind now cross-refs colnames as data.frame does, rather than always binding by column order, FR#1634. A warning is produced when the colnames are not in a consistent order. Thanks to Damian Betebenner for highlighting. rbind an unnamed list to bind columns by position.

  • The 'bysameorder' argument has been removed, as intended and warned in ?data.table.

  • New option datatable.allocwarn. See ?truelength.

  • There are now 472 raw tests, plus S4 tests.
  • v1.7.3 failed CRAN checks (and could crash) in R pre-2.14.0. Over-allocation in v1.7.3 uses truelength which is initialized to 0 by R 2.14.0, but not initialized pre-2.14.0. This was known and coded for but only tested in 2.14.0 before previous release to CRAN.
  • Two unused C variables removed to pass warning from one CRAN check machine (r-devel-fedora). -Wno-unused removed from Makevars to catch this in future before submitting to CRAN.
*   data.table now over-allocates its vector of column pointer slots
    (100 by default). This allows := to add columns fully by
    reference as suggested by Muhammad Waliji, #1646. When the 100
    slots are used up, more space is automatically allocated.

    Over allocation has negligible overhead. It's just the vector
    of column pointers, not the columns themselves.

*   New function alloc.col() pre-allocates column slots. Use
    this before a loop to add many more than 100 columns, for example,
    to avoid the warning as data.table grows its column pointer vector
    every additional 100 columns; e.g.,
        alloc.col(DT,10000)  # reserve 10,000 column slots

*   New function truelength() returns the number of column pointer
    slots allocated, always >= length() other than just after a table
    has been loaded from disk.

*   New option 'datatable.nomatch' allows the default for nomatch
    to be changed from NA to 0, as wished for by Branson Owen.

*   cbind(DT,...) now retains DT's key, as wished for by Chris Neff
    and partly implementing FR#295.
*   Assignment to factor columns (using :=, [<- or $<-) could cause
    'variable not found' errors and a seg fault in some circumstances
    due to a new feature in v1.7.0: "Factor columns on LHS of :=, [<-
    and $<- can now be assigned new levels", fixing #1664. Thanks to
    Daniele Signori for reporting.

*   DT[i,j]<-value no longer crashes when j is a factor column and value
    is numeric, fixing #1656.

*   An unnecessarily strict machine tolerance test failed CRAN checks
    on Mac preventing v1.7.2 availability for Mac (only).
*   := now has its own help page in addition to the examples in ?data.table,
    see help(":=").

*   The error message from X[Y] when X is unkeyed has been lengthened to
    including advice to call setkey first and see ?setkey. Thanks to a
    comment by ilprincipe on Stack Overflow.

*   Deleting a missing column is now a warning rather than error. Thanks
    to Chris Neff for suggesting, #1642.
*   unique and duplicated methods now work on unkeyed tables (comparing
    all columns in that case) and both now respect machine tolerance for
    double precision columns, implementing FR#1626 and fixing bug #1632.
    Their help page has been updated accordingly with detailed examples.
    Thanks to questions by Iterator and comments by Allan Engelhardt on
    Stack Overflow.

*   A new method as.data.table.list has been added, since passing a (pure)
    list to data.table() now creates a single list column.
*   Assigning to a column variable using <- or = in j now
    works (creating a local copy within j), rather than
    persisting from group to group and sometimes causing a crash.
    Non column variables still persist from group to group; e.g.,
    a group counter. This fixes the remainder of #1624 thanks to
    Steve Lianoglou for reporting.

*   A crash bug is fixed when j returns a (strictly) NULL column next
    to a non-empty column, #1633. This case was anticipated and coded
    for but an errant LENGTH() should have been length(). Thanks
    to Dennis Murphy for reporting.

*   The first column of data.table() can now be a list column, fixing
    #1640. Thanks to Stavros Macrakis for reporting.
*   .SD is now locked, partially fixing #1624. It was never
    the intention to allow assignment to .SD. Take a 'copy(.SD)'
    first if needed. Now documented in ?data.table and new FAQ 4.5
    including example. Thanks to Steve Lianoglou for reporting.

*   := now works with a logical i subset; e.g.,
    Thanks to Muhammad Waliji for reporting.
*   Error message "column <name> of i is not internally type integer"
    is now more helpful adding "i doesn't need to be keyed, just
    convert the (likely) character column to factor". Thanks to
    Christoph_J for his SO question.
*   data.table() now accepts list columns directly rather than
    needing to add list columns to an existing data.table; e.g.,

        DT = data.table(x=1:3,y=list(4:6,3.14,matrix(1:12,3)))

    Thanks to Branson Owen for reminding. As before, list columns
    can be created via grouping; e.g.,

        DT = data.table(x=c(1,1,2,2,2,3,3),y=1:7)
        DT2 = DT[,list(list(unique(y))),by=x]
             x      V1
        [1,] 1    1, 2
        [2,] 2 3, 4, 5
        [3,] 3    6, 7

    and list columns can be grouped; e.g.,

             x V1
        [1,] 1 16
        [2,] 0 12

    Accordingly, one item has been added to FAQ 2.17 (differences
    between data.frame and data.table): data.frame(list(1:2,"k",1:4))
    creates 3 columns, data.table creates one list column.

*   subset, transform and within now retain keys when the expression
    does not 'touch' key columns, implemeting FR #1341.

*   Recycling list() items on RHS of := now works; e.g.,

        # set columns 1 and 3 to 1L and remove columns 2 and 4

*   Factor columns on LHS of :=, [<- and $<- can now be assigned
    new levels; e.g.,

        DT = data.table(A=c("a","b"))
        DT[2,"A"] <- "c"  # adds new level automatically
        DT[2,A:="c"]      # same (faster)
        DT$A = "newlevel" # adds new level and recycles it

    Thanks to Damian Betebenner and Chris Neff for highlighting.
    To change the type of a column, provide a full length RHS (i.e.
    'replace' the column).
*   := with i all FALSE no longer sets the whole column, fixing
    bug #1570. Thanks to Chris Neff for reporting.

*   0 length by (such as NULL and character(0)) now behave as
    if by is missing, fixing bug #1599. This is useful when by
    is dynamic and a 'dont group' needs to be represented.
    Thanks to Chris Neff for reporting.

*   NULL j no longer results in 'inconsistent types' error, but
    instead returns no rows for that group, fixing bug #1576.

*   matrix i is now an error rather than using i as if it were a
    vector and obtaining incorrect results. It was undocumented that
    matrix might have been an acceptable type. matrix i is
    still acceptable in [<-; e.g.,
        DT[is.na(DT)] <- 1L
    and this now works rather than assigning to non-NA items in some

*   Inconsistent [<- behaviour is now fixed (#1593) so these examples
    now work :
        DT[x == "a", ]$y <- 0L
        DT["a", ]$y <- 0L
    But, := is highly encouraged instead for speed; i.e.,
        DT[x == "a", y:=0L]
        DT["a", y:=0L]
    Thanks to Leon Baum for reporting.

*   unique on an unsorted table now works, fixing bug #1601.
    Thanks to a question by Iterator on Stack Overflow.

*   Bug fix #1534 in v1.6.5 (see NEWS below) only worked if data.table
    was higher than IRanges on the search() path, despite the item in
    NEWS stating otherwise. Fixed.

*   Compatibility with package sqldf (which can call do.call("rbind",...)
    on an empty "...") is fixed and test added. data.table was switching
    on list(...)[[1]] rather than ..1. Thanks to RYogi for reporting #1623.
*   cbind and rbind are no longer masked. But, please do read FAQ 2.23,
    4.4 and 5.1.
*   Tests using .Call("Rf_setAttrib",...) passed CRAN acceptance
    checks but failed on many (but not all) platforms. Fixed.
    Thanks to Prof Brian Ripley for investigating the issue.
*   The LHS of := may now be column names or positions
    when with=FALSE; e.g.,


    This implements FR#1499 'Ability to efficiently remove a
    vector of column names' by Timothee Carayol in addition to
    creating and assigning to multiple columns. We still plan
    to allow multiple := without needing with=FALSE, in future.

*   setkey(DT,...) now returns DT (invisibly) rather than NULL.
    This is to allow compound statements; e.g.,

*   setkey (and key<-) are now more efficient when the data happens
    to be already sorted by the key columns; e.g., when data is
    loaded from ordered files.

*   If DT is already keyed by the columns passed to setkey (or
    key<-), the key is now rebuilt and checked rather than skipping
    for efficiency. This is to save needing to know to drop the key
    first to rebuild an invalid key. Invalid keys can arise by going
    'under the hood'; e.g., attr(DT,"sorted")="z", or somehow ending
    up with unordered factor levels. A warning is issued so the root
    cause can be fixed. Thanks to Timothee Carayol for highlighting.

*   A new copy() function has been added, FR#1501. This copies a
    data.table (retaining its key, if any) and should now be used to
    copy rather than data.table(). Reminder: data.tables are not
    copied on write by setkey, key<- or :=.
*   DT[,z:=a/b] and DT[a>3,z:=a/b] work again, where a and
    b are columns of DT. Thanks to Chris Neff for reporting,
    and his patience.

*   Numeric columns with class attributes are now correctly
    coerced to integer by setkey and ad hoc by. The error
    similar to 'fractional data cannot be truncated' should now
    only occur when that really is true. A side effect of
    this is that ad hoc by and setkey now work on IDate columns
    which have somehow become numeric; e.g., via rbind(DF,DF)
    as reported by Chris Neff.

*   .N is now 0 (rather than 1) when no rows in x match the
    row in i, fixing bug #1532. Thanks to Yang Zhang for

*   Compatibility with package IRanges has been restored. Both
    data.table and IRanges mask cbind and rbind. When data.table's
    cbind is found first (if it is loaded after IRanges) and the
    first argument is not data.table, it now delegates to the next
    package on the search path (and above that), one or more of which
    may also mask cbind (such as IRanges), rather than skipping
    straight to base::cbind. So, it no longer matters which way around
    data.table and IRanges are loaded, fixing #1534. Thanks to Steve
    Lianoglou for reporting.
*   setkey's verbose messages expanded.
*   DT[colA>3,which=TRUE] now returns row numbers rather
    than a logical vector, for consistency.
*   Changing a keyed column name now updates the key, too,
    so an invalid key no longer arises, fixing #1495.
    Thanks to Chris Neff for reporting.

*   := already warned when a numeric RHS is coerced to
    match an integer column's type. Now it also warns when
    numeric is coerced to logical, and integer is coerced
    to logical, fixing #1500. Thanks to Chris Neff for

*   The result of DT[,newcol:=3.14] now includes the new
    column correctly, as well as changing DT by reference,
    fixing #1496. Thanks to Chris Neff for reporting.

*   :=NULL to remove a column (instantly, regardless of table
    size) now works rather than causing a segfault in some
    circumstances, fixing #1497. Thanks to Timothee Carayol
    for reporting.

*   Previous within() and transform() behaviour restored; e.g.,
    can handle multiple columns again. Thanks to Timothee Carayol
    for reporting.

*   cbind(DT,DF) now works, as does rbind(DT,DF), fixing #1512.
    Thanks to Chris Neff for reporting. This was tricky to fix due
    to nuances of the .Internal dispatch code in cbind and rbind,
    preventing S3 methods from working in all cases.
    R will now warn that cbind and rbind have been masked when
    the data.table package is loaded. These revert to base::cbind
    and base::rbind when the first argument is not data.table.

*   Removing multiple columns now works (again) using
    DT[,c("a","b")]=NULL, or within(DT,rm(a,b)), fixing #1510.
    Thanks to Timothee Carayol for reporting.
*   The package uses two features (packageVersion() and \href in Rd)
    added to R 2.12.0 and is therefore dependent on that release.
    A 'spurious warning' when checking a package using \href was
    fixed in R 2.12.2 patched but we believe that warning can safely
    be ignored in versions >= 2.12.0 and < 2.12.2 patched.
*   Ad hoc grouping now returns results in the same order each
    group first appears in the table, rather than sorting the
    groups. Thanks to Steve Lianoglou for highlighting. The order
    of the rows within each group always has and always will be
    preserved. For larger datasets a 'keyed by' is still faster;
    e.g., by=key(DT).

*   The 'key' argument of data.table() now accepts a vector of
    column names in addition to a single comma separated string
    of column names, for consistency. Thanks to Steve Lianoglou
    for highlighting.

*   A new argument '.SDcols' has been added to [.data.table. This
    may be character column names or numeric positions and
    specifies the columns of x included in .SD. This is useful
    for speed when applying a function through a subset of
    (possibly very many) columns; e.g.,

*   as(character, "IDate") and as(character, "ITime") coercion
    functions have been added. Enables the user to declaring colClasses
    as "IDate" and "ITime" in the various read.table (and sister)
    functions. Thanks to Chris Neff for the suggestion.

*   DT[i,j]<-value is now handled by data.table in C rather
    than falling through to data.frame methods, FR#200. Thanks to
    Ivo Welch for raising speed issues on r-devel, to Simon Urbanek
    for the suggestion, and Luke Tierney and Simon for information
    on R internals.

    [<- syntax still incurs one working copy of the whole
    table (as of R 2.13.1) due to R's [<- dispatch mechanism
    copying to `*tmp*`, so, for ultimate speed and brevity,
    the operator := may now be used in j as follows.

*   := is now available to j and means assign to the column by
    reference; e.g.,


    This syntax makes no copies of any part of memory at all.

    m = matrix(1,nrow=100000,ncol=100)
    DF = as.data.frame(m)
    DT = as.data.table(m)

    system.time(for (i in 1:1000) DF[i,1] <- i)
         user  system elapsed
      287.062 302.627 591.984

    system.time(for (i in 1:1000) DT[i,V1:=i])
         user  system elapsed
        1.148   0.000   1.158     ( 511 times faster )

    := in j can be combined with all types of i, such as binary
    search, and used to add and remove columns efficiently.
    Fast assigning within groups will be implemented in future.

    Reminder that data.frame and data.table both allow columns
    of mixed types, including columns which themselves may be
    type list; matrix may be one (atomic) type only.

    *Please note*, := is new and experimental.
*   merge()ing two data.table's with user-defined `suffixes`
    was getting tripped up when column names in x ended in
    '.1'. This resulted in the `suffixes` parameter being

*   Mistakenly wrapping a j expression inside quotes; e.g.,
    was appearing to work, but with wrong column names. This
    now returns a character column (the quotes should not
    be used). Thanks to Joseph Voelkel for reporting.

*   setkey has been made robust in several ways to fix issues
    introduced in 1.6.2: #1465 ('R crashes after setkey')
    reported by Eugene Tyurin and similar bug #1387 ('paste()
    by group to create long comma separated strings can crash')
    reported by Nicolas Servant and Jean-Francois Rami. This
    bug was not reproducible so we are especially grateful for
    the patience of these people in helping us find, fix and
    test it.

*   Combining a join, j and by together in one query now works
    rather than giving an error, fixing bug #1468. Discovered
    indirectly thanks to a post from Jelmer Ypma.

*   Invalid keys no longer arise when a non-data.table-aware
    package reorders the data; e.g.,
        plyr::arrange(DT,y)       # same as DT[order(y)]
    This now drops the key to avoid incorrect results being
    returned the next time the invalid key is joined to. Thanks
    to Chris Neff for reporting.
*   The startup banner has been shortened to one line.

*   data.table does not support POSIXlt. Almost unbelievably
    POSIXlt uses 40 bytes to store a single datetime. If it worked
    before, that was unintentional. Please see ?IDateTime, or any
    other date class that uses a single atomic vector. This is
    regardless of whether the POSIXlt is a key column, or not. This
    resolves bug #1481 by documenting non support in ?data.table.
  • Use of the DT() alias in j is no longer caught for backwards compatibility and is now fully removed. As warned in NEWS for v1.5.3, v1.4, and FAQs 2.6 and 2.7.
  • setkey no longer copies the whole table and should be faster for large tables. Each column is reordered by reference (in C) using one column of working memory, FR#1006. User defined attributes on the original table are now also retained (thanks to Thell Fowler for reporting).

  • A new symbol .N is now available to j, containing the number of rows in the group. This may be useful when the column names are not known in advance, for convenience generally, and for efficiency.

  • j's environment is now consistently reused so that local variables may be set which persist from group to group; e.g., incrementing a group counter : DT[,list(z,groupInd<-groupInd+1),by=x] Thanks to Andreas Borg for reporting.

  • A new symbol .BY is now available to j, containing 1 row of the current 'by' variables, type list. 'by' variables may also be used by name, and are now length 1 too. This implements FR#1313. FAQ 2.10 has been updated accordingly. Some examples : DT[,sum(x)*.BY[[1]],by=eval(byexp)] DT[,sum(x)*mylookuptable[J(y),z],by=y] DT[,list(sum(unlist(.BY)),sum(z)),by=list(x,y%%2)]

  • i may now be type list, and works the same as when i is type data.table. This saves needing J() in as many situations and may be a little more efficient. One application is using .BY directly in j to join to a relatively small lookup table, once per group, for space and time efficiency. For example : DT[,list(GROUPDATA[.BY]$name,sum(v)),by=grp]

  • A 'by' character vector of column names now works when there are less rows than columns; e.g., DT[,sum(x),by=key(DT)] where nrow(DT)==1. Many thanks to Andreas Borg for report, proposed fix and tests.

  • Zero length columns in j no longer cause a crash in some circumstances. Empty columns are filled with NA to match the length of the longest column in j. Thanks to Johann Hibschman for bug report #1431.

  • unique.data.table now calls the same internal code (in C) that grouping calls. This fixes a bug when unique is called directly by user, and, NA exist in the key (which might be quite rare). Thanks to Damian Betebenner for bug report. unique should also now be faster.

  • Variables in calling scope can now be used in j when i is logical or integer, fixing bug #1421. Thanks to Alexander Peterhansl for reporting.

*   ?data.table now documents that logical i is not quite
    the same as i in [.data.frame. NA are treated as FALSE,
    and DT[NA] returns 1 row of NA, unlike [.data.frame.
    Three points have been added to FAQ 2.17. Thanks to
    Johann Hibschman for highlighting.

*   Startup banner now uses packageStartupMessage() so the
    banner can be suppressed by those annoyed by banners,
    whilst still being helpful to new users.
  • data.table now plays nicely with S4 classes. Slots can be defined to be S4 objects, S4 classes can inherit from data.table, and S4 function dispatch works on data.table objects. See the tests in inst/tests/test-S4.R, and from the R prompt: ?"data.table-class"

  • merge.data.table now works more like merge.data.frame: (i) suffixes are consistent with merge.data.frame; existing users may set options(datatable.pre.suffixes=TRUE) for backwards compatibility. (ii) support for 'by' argument added (FR #1315). However, X[Y] syntax is preferred; some users never use merge.

  • by=key(DT) now works when the number of rows is not divisible by the number of groups (#1298, an odd bug). Thanks to Steve Lianoglou for reporting.

  • Combining i and by where i is logical or integer subset now works, fixing bug #1294. Thanks to Johann Hibschman for contributing a new test.

  • Variable scope inside [[...]] now works without a workaround required. This can be useful for looking up which function to call based on the data e.g. DT[,fns[fn],by=ID]. Thanks to Damian Betebenner for reporting.

  • Column names in self joins such as DT[DT] are no longer duplicated, fixing bug #1340. Thanks to Andreas Borg for reporting.

  • Additions and updates to FAQ vignette. Thanks to Dennis Murphy for his thorough proof reading.

  • Welcome to Steve Lianoglou who joins the project contributing S4-ization, testing using testthat, and more.

  • IDateTime is now linked from ?data.table. data.table users unaware of IDateTime, please do take a look. Tom added IDateTime in v1.5 (see below).

  • .SD no longer includes 'by' columns, FR#978. This resolves the long standing annoyance of duplicated 'by' columns when the j expression returns a subset of rows from .SD. For example, the following query no longer contains a redundant 'colA.1' duplicate. DT[,.SD[2],by=colA] # 2nd row of each group Any existing code that uses .SD may require simple changes to remove workarounds.

  • 'by' may now be a character vector of column names. This allows syntax such as DT[,sum(x),by=key(DT)].

  • X[Y] now includes Y's non-join columns, as most users naturally expect, FR#746. Please do use j in one step (i.e. X[Y,j]) since that merges just the columns j uses and is much more efficient than X[Y][,j] or merge(X,Y)[,j].

  • The 'Join Inherited Scope' feature is back on, FR#1095. This is consistent with X[Y] including Y's non-join columns, enabling natural progression from X[Y] to X[Y,j]. j sees columns in X first then Y. If the same column name exists in both X and Y, the data in Y can be accessed via a prefix "i." (not yet implemented).

  • Ad hoc by now coerces double to integer (provided they are all.equal) and character to factor, FR#1051, as setkey already does.

  • The default for mult is now "all", as planned and prior notice given in FAQ 2.2.

  • ?[.data.table has been merged into ?data.table and updated, simplified, corrected and formatted.

  • The DT() alias is now fully deprecated, as warned in NEWS for v1.4, and FAQs 2.6 and 2.7.
  • 'by' now works when DT contains list() columns i.e. where each value in a column may itself be vector or where each value is a different type. FR#1092.

  • The result from merge() is now keyed. FR#1244.

*   eval of parse()-ed expressions now works without
    needing quote() in the expression, bug #1243. Thanks
    to Joseph Voelkel for reporting.

*   the result from the first group alone may be bigger
    than the table itself, bug #1245. Thanks to
    Steve Lianoglou for reporting.

*   merge on a data.table with a single key'd column only
    and all=TRUE now works, bug #1241. Thanks to
    Joseph Voelkel for reporting.

*   merge()-ing by a column called "x" now works, bug
    #1229 related to variable scope. Thanks to Steve
    Lianoglou for reporting.
*   Fixed inheritance for other packages importing or depending
    on data.table, bugs #1093 and #1132. Thanks to Koert Kuipers
    for reporting.

*   data.table queries can now be used at the debugger() prompt,
    fixing bug #1131 related to inheritance from data.frame.
*   data.table now *inherits* from data.frame, for functions and
    packages which _only_ accept data.frame, saving time and
    memory of conversion. A data.table is a data.frame too;
    is.data.frame() now returns TRUE.

*   Integer-based date and time-of-day classes have been
    introduced. This allows dates and times to be used as keys
    more easily. See as.IDate, as.ITime, and IDateTime.
    Conversions to and from POSIXct, Date, and chron are

*   [<-.data.table and $<-.data.table were revised to check for
    changes to the key-ed columns. [<-.data.table also now allows
    data.table-style indexing for i. Both of these changes may
    introduce incompatibilities for existing code.

*   Logical columns are now allowed in keys and in 'by', as are expressions
    that evaluate to logical. Thanks to David Winsemius for highlighting.
*   DT[,5] now returns 5 as FAQ 1.1 says, for consistency
    with DT[,c(5)] and DT[,5+0]. DT[,"region"] now returns
    "region" as FAQ 1.2 says. Thanks to Harish V for reporting.

*   When a quote()-ed expression q is passed to 'by' using
    by=eval(q), the group column names now come from the list
    in the expression rather than the name 'q' (bug #974) and,
    multiple items work (bug #975). Thanks to Harish V for

*   quote()-ed i and j expressions receive similar fixes, bugs
    #977 and #1058. Thanks to Harish V and Branson Owen for

*   Multiple errors (grammar, format and spelling) in intro.Rnw
    and faqs.Rnw corrected by Dennis Murphy. Thank you.

*   Memory is now reallocated in rare cases when the up front
    allocate for the result of grouping is insufficient. Bug
    #952 raised by Georg V, and also reported by Harish. Thank

*   A function call foo(arg=sum(b)) now finds b in DT when foo
    contains DT[,eval(substitute(arg)),by=a], fixing bug #1026.
    Thanks to Harish V for reporting.

*   If DT contains column 'a' then DT[J(unique(a))] now finds
    'a', fixing bug #1005. Thanks to Branson Owen for reporting.

*   'by' on no data (for example when 'i' returns no rows) now
    works, fixing bug #709.

*   'by without by' now heeds nomatch=NA, fixing bug #1015.
    Thanks to Harish V for reporting.

*   DT[NA] now returns 1 row of NA rather than the whole table
    via standard NA logical recycling. A single NA logical is
    a special case and is now replaced by NA_integer_. Thanks
    to Branson Owen for highlighting the issue.

*   NROW removed from data.table, since the is.data.frame() in
    base::NROW now returns TRUE due to inheritance. Fixes bug
    #1039 reported by Bradley Buchsbaum. Thank you.

*   setkey() now coerces character to factor and double to
    integer (provided they are all.equal), fixing bug #953.
    Thanks to Steve Lianoglou for reporting.

*   'by' now accepts lists from the calling scope without the
    work around of wrapping with as.list() or {}, fixing bug
    #1060. Thanks to Johann Hibschman for reporting.
*   The package uses the 'default' option of base::getOption,
    and is therefore dependent on R 2.10.0. Updated DESCRIPTION
    file accordingly. Thanks to Christian Hudon for reporting.
*   Vignettes tidied up.
*   Out of order levels in key columns are now sorted by
    setkey. Thanks to Steve Lianoglou for reporting.
*   'by' faster. Memory is allocated first for the result, then
populated directly by the result of j for each group. Can be 10
or more times faster than tapply() and aggregate(), see
timings vignette.

*   j should now be a list(), not DT(), of expressions. Use of
j=DT(...) is caught internally and replaced with j=list(...).

*   'by' may be a list() of expressions. A single column name
is automatically list()-ed for convenience. 'by' may still be
a comma separated character string, as before.
    DT[,sum(x),by=region]                     # new
    DT[,sum(x),by=list(region,month(date))]   # new
    DT[,sum(x),by="region"]                   # old, ok too
    DT[,sum(x),by="region,month(date)"]       # old, ok too

*   key() and key<- added. More R-style alternatives to getkey()
and setkey().

*   haskey() added. Returns TRUE if a table has a key.

*   radix sorting is now column by column where possible, was
previously all or nothing. Helps with keys of many columns.

*   Added format method.

*   22 tests added to test.data.table(), now 149.

*   Three vignettes added : FAQ, Intro & Timings
*   The DT alias is removed. Use 'data.table' instead to create
objects. See 2nd new feature above.

*   RUnit framework removed.
test.data.table() is called from examples in .Rd so 'R CMD check'
will run it. Simpler. An eval(body(test.data.table))
is also in the .Rd, to catch namespace issues.

*   Dependency on package 'ref' removed.

*   Arguments removed:  simplify, incbycols and byretn.
Grouping is simpler now, these are superfluous.
*   Column classes are now retained by subset and grouping.

*   tail no longer fails when a column 'x' exists.
*   Minor : Join Inherited Scope not working, contrary
    to the documentation.
*   v1.4 was essentially the branch at rev 44, reintegrated
at rev 78.
*   Radix sorting added. Speeds up setkey and add-hoc 'by'
by factor of 10 or more.

*   Merge method added, much faster than base::merge method
of data.frame.

*   'by' faster. Logic moved from R into C. Memory is
allocated for the largest group only, then re-used.

*   The Sub Data is accessible as a whole by j using object
.SD. This should only be used in rare circumstances. See FAQ.

*   Methods added : duplicated, unique, transform, within,
[<-, t, Math, Ops, is.na, na.omit, summary

*   Column name rules improved e.g. dots now allowed.

*   as.data.frame.data.table rownames improved.

*   29 tests added to test.data.table(), now 127.
*   Default of mb changed, now tables(mb=TRUE)
*   ... removed in [.data.table.
j may not be a function, so this is now superfluous.
*   Incorrect version warning with R 2.10+ fixed.

*   j enclosure raised one level. This fixes some bugs
where the j expression previously saw internal variable
names. It also speeds up grouping a little.
*   v1.3 was not released to CRAN. R-Forge repository only.

Reference manual

It appears you don't have a PDF plugin for this browser. You can click here to download the reference manual.


1.10.4 by Matt Dowle, 20 days ago


Report a bug at https://github.com/Rdatatable/data.table/issues

Browse source code at https://github.com/cran/data.table

Authors: Matt Dowle [aut, cre], Arun Srinivasan [aut], Jan Gorecki [ctb], Tom Short [ctb], Steve Lianoglou [ctb], Eduard Antonyan [ctb]

Documentation:   PDF Manual  

Task views: Empirical Finance, High-Performance and Parallel Computing with R

GPL-3 | file LICENSE license

Imports methods

Suggests bit64, knitr, nanotime, chron, ggplot2, plyr, reshape, reshape2, testthat, hexbin, fastmatch, nlme, xts, gdata, GenomicRanges, caret, curl, zoo, plm, rmarkdown, parallel

Imported by AFM, APSIM, BTYDplus, CatEncoders, Causata, ChannelAttributionApp, DataCombine, DataExplorer, Epi, FAOSTAT, FedData, Fgmutils, GDINA, GGIR, GSODR, Hmisc, IPtoCountry, ITGM, Information, InformationValue, Kmisc, LogisticDx, Luminescence, OpenML, PGRdup, PSF, PTXQC, PopGenReport, R6Frame, RAM, RAPIDR, RBMRB, RDML, REDCapR, RNewsflow, RbioRXN, Rdice, RecordLinkage, RndTexExams, RxnSim, SGP, SGPdata, SWMPr, SafeQuant, SciencesPo, SetRank, SocialMediaLab, SpATS, SpaDES, StatMeasures, TSTr, TideCurves, TideTables, aLFQ, alakazam, assertable, bedr, benford.analysis, biomartr, bold, brainGraph, bridger2, caretEnsemble, chromer, cleanEHR, clickstream, dat, datadr, dataview, dtplyr, eclust, ecoengine, edarf, edeaR, eeptools, eiCompare, emil, emuR, extracat, fbRads, finch, flippant, freqweights, fst, futureheatwaves, gasfluxes, gems, genderizeR, getlandsat, githubinstall, gofastr, grattan, gridsample, gunsales, healthcareai, hisse, hoardeR, hurricaneexposure, inbreedR, isdparser, joinXL, lar, ldamatch, lidR, livechatR, lllcrc, mapview, metaMix, mlr, mltools, mmpf, mregions, msmtools, mtconnectR, mvQuad, natserv, ndjson, noncompliance, optiRum, originr, partialAR, pcrsim, polmineR, popEpi, pxweb, qdapTools, quanteda, r2glmm, rAmCharts, rBayesianOptimization, rFTRLProximal, rIsing, randomNames, rangeMapper, rbi, rdpla, rdrop2, readability, repmis, rerddap, rfisheries, rgbif, rio, ritis, rlas, rlist, rnpn, rplexos, rpubchem, sdcMicro, sdcTable, sentimentr, shazam, simPH, simcausal, solarius, speaq2, spellcheckr, spocc, squid, starmie, statar, stdReg, stremr, strvalidator, survMisc, syllable, tableMatrix, taxize, tcR, text2vec, textTinyR, textshape, tmlenet, toxboot, traits, translateSPSS2R, treemap, trelliscope, unpivotr, xgboost, yCrypticRNAs.

Depended on by AF, ARTP2, BuyseTest, EurosarcBayes, GenomicTools, JWileymisc, LSPFP, LabourMarketAreas, QuantTools, RSauceLabs, RWildbook, Rbitcoin, VIM, batchtools, bea.R, cffdrs, classifierplots, cvAUC, drgee, edgeRun, fdq, gbp, greport, haploReconstruct, ie2misc, iemisc, lookupTable, musica, oec, orgR, panelaggregation, partools, psidR, robCompositions, scanstatistics, simPop, simstudy, sitree, skm, slim, sparseFLMM, splitstackshape, surveyplanning, synthACS, tcpl, ttwa, validateRS, vardpoor, word.alignment.

Suggested by ALA4R, BEDMatrix, Lahman, NEArender, ProjectTemplate, Rblpapi, Xmisc, assertive.types, bikeshare14, checkmate, csvy, ggswissmaps, googleVis, h2o, ie2miscdata, iemiscdata, installr, itsadug, lava, nanotime, padr, propr, rbokeh, rpivotTable, rtop, seleniumPipes, sticky, taRifx, vaersNDvax, vaersvax.

Enhanced by dendextend, repr.

See at CRAN