Version 0.9.6 is a minor update to ‘hotfix’ an issue with the use of All calls to now explicitly reference a length one size object.

Version 0.9.4 is a minor update to ‘hotfix’ an issue with suggested packages as reported by CRAN.

Bugs solved

  • Packages nanotime, bit64 and lintr are now used conditionally as is required for packages suggested in DESCRIPTION, see ‘Writing R Extensions’ (thanks @CRAN for reporting)

Version 0.9.2 of the fst package brings support for zero-row table serialization and compression for long vectors. In addition, fst was prepared for the change in the default settings for the stringsAsFactors argument (data.frame) in R 4.0.0.

Library updates

  • Library fstlib updated to version 0.1.6
  • Library LZ4 updated to version 1.9.2
  • library ZSTD updated to version 1.4.4


  • Incorrect column selection gets a more informative error message (issue #138, thanks Jean-Luc and @Moohan for reporting).
  • Long raw vectors can be hashed with hash_fst (issue #202)
  • Empty tables are serialized correctly using write_fst() (issue #99)

Bugs solved

  • Ellipsis is forwarded in method str.fst_table() (issue #219, thanks @nbenn for reporting).
  • Coloring is turned off for terminals that don’t support it or when package crayon is not installed (issue #198, thanks @muschellij2 for reporting and the code fix).
  • Method metadata_fst() correctly displays the key of the data table if column names are not in alphabetical order (issue #199, thanks @renkun-ken for the pull request).
  • stringsAsFactors argument defaults to FALSE for upcoming R 4.0.0 (issue #234, thanks @CRAN for reporting)

Version 0.9.0 of the fst package addresses the request from CRAN maintainers to fix issues identified by rchk. These issues result from PROTECT / UNPROTECT pairs called in the constructor / destructor pairs of C++ classes. rchk (rightfully) warns about those because it can’t determine from the code if pairs are properly matched. With this submission the relevant SEXP classes are protected by containing them in SEXP classes that are already PROTECTED, which allows for removal of the PROTECT / UNPROTECT pairs in question.

As of fst version 0.9.0, support for fst files generated with fst package versions lower than 0.8.0 has been deprecated. This significantly reduces the (C++) code base and prepares fst for future code changes.

Library updates

  • Library fstlib updated to version 0.1.1


  • Method setnrofthreads returns invisible result to avoid printing unwanted output (thanks @renkun-ken for the pull request)

Bugs solved


Various documentation issues have been fixed (thanks @ginberg and @renkun-ken for the pull requests).

Version 0.8.10 of the fst package is an intermediate release designed to update the incorporated C++ libraries to their latest versions and to fix reported issues. Also, per request of CRAN maintainers, the OpenMP build option was moved to the correct flag in the Makevars file, resolving a warning in the package check.

Library updates

  • Library fstlib updated to version 0.1.0

  • Library ZSTD updated to version 1.3.7

  • Library LZ4 updated to version 1.8.3

Bugs solved

  • Method compress_fst() can handle vectors with sizes larger than 4 GB (issue #176, thanks @bwlewis for reporting)

  • A fst file is correctly read from a subfolder on a network drive where the user does not have access to the top-level folder (issues #136 and #175, thanks @xiaodaigh for reporting).

  • The suggested data.table dependency is now properly escaped (issue #181, thanks @jangorecki for the pull request)


Version 0.8.8 of the fst package is an intermediate release designed to fix valgrind warnings reported on CRAN builds (per request of CRAN maintainers). These warnings were due to fst writing uninitialized data buffers to file, which was done to maximize speed. To fix these warnings (and for safety), all memory blocks are now initialized to zero before being written to disk.

Version 0.8.6 of the fst package brings clearer printing of fst_table objects. It also includes optimizations for controlling the number of threads used by the package during reads and writes and after a fork has ended. The LZ4 and ZSTD compression libraries are updated to their latest (and fastest) releases. UTF-8 encoded column names are now correctly stored in the fst format.

New features

  • More advanced printing generic of the fst_table reference object, showing column types, (possible) keys, and the table header and footer data (issue #131, thanks @renkun-ken for reporting and discussions).

  • User has more control over the number of threads used by fst. Option ‘fst_threads’ can now be used to initialize the number of threads when the package is first loaded (issue #132, thanks to @karldw for the pull request).

  • Option ‘fst_restore_after_fork’ can be used to select the threading behavior after a fork has ended. Like the data.table package, fst switches back to a single thread when a fork is detected (using OpenMP in a fork can lead to problems). Unlike data.table, the fst package restores the number of threads to it’s previous setting when the fork ends. If this leads to unexpected problems, the user can set the ‘fst_restore_after_fork’ option to FALSE to disable that.

Bugs solved

  • Character encoding of column names correctly stored in the fst format (issue #144, thanks @shrektan for reporting and discussions).


  • Improved accuracy of fst_table documentation regarding random row access (issue #143, thanks @martinblostein for pointed out the unclarity)

  • Improved documentation on background threads during write_fst() and read_fst() (issue #121, thanks @krlmlr for suggestions and discussion)

The v0.8.4 release brings a data.frame interface to the fst package. Column and row selection can now be done directly from the [ operator. In addition, it fixes some issues and prepares the package for the next build toolchain of CRAN.

New features

  • A data.frame interface was added to the package. The user can create a reference object to a fst file with method fst. That reference can be used like a data.frame and will automatically make column- and row- selections in the referenced fst file.

Bugs solved

  • Build issues with the dev build of R have been fixed. In particular, fst now builds correctly with the Clang 6.0 toolchain which will be released by CRAN shortly (thanks @kevinushey for reporting the problem and CRAN maintainers for the advance warning.

  • An error was fixed where compressing a short factor column with 128 to 32767 levels but only a single value, returned incorrect results (issue #128, thanks @martinblostein for reporting and help fixing the problem).

  • An error was fixed where columns f type ‘ITime’ were incorrectly serialized (issue #126, thanks @Giqles for reporting the problem).

  • An error was fixed where using fst as a dependency in another package and building that package in RStudio, crashed RStudio. The problem was that RStudio uses a fork to build or document a package. That fork made fst use OpenMP library methods, which leads to crashes on macOS. After the fix, no calls to any OpenMP library method are now made from fst when it’s run from a forked process (issue #100 and issue #109, thanks to @eipi10, @PeteHaitch, @kevinushey, @thierrygosselin, @xiaodaigh and @jzzcutler for reporting the problem and help fix it).


  • Documentation for method write_fst was improved (issue #123, thanks @krlmlr for reporting and submitting a pull request).

New features

  • Package fst has support for multi-threading using OpenMP. Compression, decompression and disk IO have been largely parallelized for (much) improved performance.

  • Many new column types are now supported by the fst format (where appropriate, both the double and integer variants are supported):

    • raw
    • DateTime
    • integer64
    • nanotime
    • POSIXct
    • ordered factors
    • difftime

    Thanks @arunsrinivasan, @derekholmes, @phillc73, @HughParsonage, @statquant, @eddelbuettel, @eipi10, and @verajosemanuel for feature requests and helpful discussions.

  • Multi-threaded LZ4 and ZSTD compression using methods compress_fst and decompress_fst. These methods provide a direct API to the LZ4 and ZSTD compressors at speeds of multiple GB/s. A specific block format is used to facilitate parallel processing. For additional stability, hashes can be calculated if required.

  • Method hash_fst provides an extremely fast multi-threaded 64-bit hashing algorithm based on xxHash. Speeds up to the memory bandwidth can be achieved.

  • Faster conversion to data.table in read_fst. Thanks @dselivanov

  • Package data.table is now an optional dependency. Thanks @jimhester. Note that in the near future, a dependency on data.table will probably be introduced again, as fst will get a data.table-like interface.

  • The fst format has a magic number to be able to identify a fst file without actually opening the file or requiring the fstlib library. Thanks @davidanthoff.

  • For development versions, the build number is now shown when fst is loaded. Thanks @skanskan.

  • Character encodings are preserved for character and factor columns. Thanks @carioca67 and @adrianadermon

  • Naming of fst methods is now consistent. Thanks @jimhester and @brinkhuis.

  • The core C++ code with the API to read and write fst files, and use compression and hashing now lives in a separate library called fstlib. Although not visible to the user, this is a major development allowing fst to be implemented for other languages than R (with comparable performance).

Bugs solved

  • Tilde-expansion in write_fst not correctly processed. Thanks @HughParsonage, @PoGibas.

  • Writing more than INT_MAX rows crashes fst. Thanks @wei-wu-nyc

  • Incorrect fst file is created when an empty data.table is saved. Thanks @wei-wu-nyc.

  • Error/crash when saving factor column with 0 factor levels. Thanks @martinblostein.

  • No warning was given when disk runs out of space during a fstwrite operation.

  • A data.table warning message was given on modification of columns of a sorted table. Thanks @martinblostein.

  • Stack imbalance warnings under certain conditions. Thanks @ryankennedyio


Thanks to @mattdowle, @st-pasha, @phillc73 for valuable discussions on fst benchmarks and how to accurately perform (and present) them.

Additional credits