NEWS.md
There are now two ways to query Arrow data:
dplyr::summarize(), both grouped and ungrouped, is now implemented for Arrow Datasets, Tables, and RecordBatches. Because data is scanned in chunks, you can aggregate over larger-than-memory datasets backed by many files. Supported aggregation functions include n(), n_distinct(), min(), max(), sum(), mean(), var(), sd(), any(), and all(). median() and quantile() with one probability are also supported and currently return approximate results using the t-digest algorithm.
Along with summarize(), you can also call count(), tally(), and distinct(), which effectively wrap summarize().
This enhancement does change the behavior of summarize() and collect() in some cases: see “Breaking changes” below for details.
In addition to summarize(), mutating and filtering equality joins (inner_join(), left_join(), right_join(), full_join(), semi_join(), and anti_join()) with are also supported natively in Arrow.
Grouped aggregation and (especially) joins should be considered somewhat experimental in this release. We expect them to work, but they may not be well optimized for all workloads. To help us focus our efforts on improving them in the next release, please let us know if you encounter unexpected behavior or poor performance.
New non-aggregating compute functions include string functions like str_to_title() and strftime() as well as compute functions for extracting date parts (e.g. year(), month()) from dates. This is not a complete list of additional compute functions; for an exhaustive list of available compute functions see list_compute_functions().
We’ve also worked to fill in support for all data types, such as Decimal, for functions added in previous releases. All type limitations mentioned in previous release notes should be no longer valid, and if you find a function that is not implemented for a certain data type, please report an issue.
If you have the duckdb package installed, you can hand off an Arrow Dataset or query object to DuckDB for further querying using the to_duckdb() function. This allows you to use duckdb’s dbplyr methods, as well as its SQL interface, to aggregate data. Filtering and column projection done before to_duckdb() is evaluated in Arrow, and duckdb can push down some predicates to Arrow as well. This handoff does not copy the data, instead it uses Arrow’s C-interface (just like passing arrow data between R and Python). This means there is no serialization or data copying costs are incurred.
You can also take a duckdb tbl and call to_arrow() to stream data to Arrow’s query engine. This means that in a single dplyr pipeline, you could start with an Arrow Dataset, evaluate some steps in DuckDB, then evaluate the rest in Arrow.
arrange() the query result. For calls to summarize(), you can set options(arrow.summarise.sort = TRUE) to match the current dplyr behavior of sorting on the grouping columns.dplyr::summarize() on an in-memory Arrow Table or RecordBatch no longer eagerly evaluates. Call compute() or collect() to evaluate the query.head() and tail() also no longer eagerly evaluate, both for in-memory data and for Datasets. Also, because row order is no longer deterministic, they will effectively give you a random slice of data from somewhere in the dataset unless you arrange() to specify sorting.sf::st_as_binary(col)) or using the sfarrow package which handles some of the intricacies of this conversion process. We have plans to improve this and re-enable custom metadata like this in the future when we can implement the saving in a safe and efficient way. If you need to preserve the pre-6.0.0 behavior of saving this metadata, you can set options(arrow.preserve_row_level_metadata = TRUE). We will be removing this option in a coming release. We strongly recommend avoiding using this workaround if possible since the results will not be supported in the future and can lead to surprising and inaccurate results. If you run into a custom class besides sf columns that are impacted by this please report an issue.LIBARROW_MINIMAL=true. This will have the core Arrow/Feather components but excludes Parquet, Datasets, compression libraries, and other optional features.create_package_with_all_dependencies() function (also available on GitHub without installing the arrow package) will download all third-party C++ dependencies and bundle them inside the R source package. Run this function on a system connected to the network to produce the “fat” source package, then copy that .tar.gz package to your offline machine and install. Special thanks to @karldw for the huge amount of work on this.libz) by setting ARROW_DEPENDENCY_SOURCE=AUTO. This is not the default in this release (BUNDLED, i.e. download and build all dependencies) but may become the default in the future.read_json_arrow()) are now optional and still on by default; set ARROW_JSON=OFF before building to disable them.options(arrow.use_altrep = FALSE)
Field objects can now be created as non-nullable, and schema() now optionally accepts a list of Fieldswrite_parquet() no longer errors when used with a grouped data.framecase_when() now errors cleanly if an expression is not supported in Arrowopen_dataset() now works on CSVs without header rowsT and t were reversed in read_csv_arrow()
log(..., base = b) where b is something other than 2, e, or 10Table$create() now has alias arrow_table()
This patch version contains fixes for some sanitizer and compiler warnings.
There are now more than 250 compute functions available for use in dplyr::filter(), mutate(), etc. Additions in this release include:
strsplit() and str_split(); strptime(); paste(), paste0(), and str_c(); substr() and str_sub(); str_like(); str_pad(); stri_reverse()
lubridate methods such as year(), month(), wday(), and so onlog() et al.); trigonometry (sin(), cos(), et al.); abs(); sign(); pmin() and pmax(); ceiling(), floor(), and trunc()
ifelse() and if_else() for all but Decimal types; case_when() for logical, numeric, and temporal types only; coalesce() for all but lists/structs. Note also that in this release, factors/dictionaries are converted to strings in these functions.is.* functions are supported and can be used inside relocate()
arrow_dplyr_query now includes the expression and the resulting type of columns derived by mutate().transmute() now errors if passed arguments .keep, .before, or .after, for consistency with the behavior of dplyr on data.frames.
write_csv_arrow() to use Arrow to write a data.frame to a single CSV filewrite_dataset(format = "csv", ...) to write a Dataset to CSVs, including with partitioningreticulate::py_to_r() and r_to_py() methods. Along with the addition of the Scanner$ToRecordBatchReader() method, you can now build up a Dataset query in R and pass the resulting stream of batches to another tool in process.Array$export_to_c(), RecordBatch$import_from_c()), similar to how they are in pyarrow. This facilitates their use in other packages. See the py_to_r() and r_to_py() methods for usage examples.data.frame to an Arrow Table uses multithreading across columnsoptions(arrow.use_altrep = FALSE)
is.na() now evaluates to TRUE on NaN values in floating point number fields, for consistency with base R.is.nan() now evaluates to FALSE on NA values in floating point number fields and FALSE on all values in non-floating point fields, for consistency with base R.Array, ChunkedArray, RecordBatch, and Table: na.omit() and friends, any()/all()
RecordBatch$create() and Table$create() are recycledarrow_info() includes details on the C++ build, such as compiler versionmatch_arrow() now converts x into an Array if it is not a Scalar, Array or ChunkedArray and no longer dispatches base::match().LIBARROW_MINIMAL=false) includes both jemalloc and mimalloc, and it has still has jemalloc as default, though this is configurable at runtime with the ARROW_DEFAULT_MEMORY_POOL environment variable.LIBARROW_MINIMAL, LIBARROW_DOWNLOAD, and NOT_CRAN are now case-insensitive in the Linux build script.Many more dplyr verbs are supported on Arrow objects:
dplyr::mutate() is now supported in Arrow for many applications. For queries on Table and RecordBatch that are not yet supported in Arrow, the implementation falls back to pulling data into an in-memory R data.frame first, as in the previous release. For queries on Dataset (which can be larger than memory), it raises an error if the function is not implemented. The main mutate() features that cannot yet be called on Arrow objects are (1) mutate() after group_by() (which is typically used in combination with aggregation) and (2) queries that use dplyr::across().dplyr::transmute() (which calls mutate())dplyr::group_by() now preserves the .drop argument and supports on-the-fly definition of columnsdplyr::relocate() to reorder columnsdplyr::arrange() to sort rowsdplyr::compute() to evaluate the lazy expressions and return an Arrow Table. This is equivalent to dplyr::collect(as_data_frame = FALSE), which was added in 2.0.0.Over 100 functions can now be called on Arrow objects inside a dplyr verb:
nchar(), tolower(), and toupper(), along with their stringr spellings str_length(), str_to_lower(), and str_to_upper(), are supported in Arrow dplyr calls. str_trim() is also supported.sub(), gsub(), and grepl(), along with str_replace(), str_replace_all(), and str_detect(), are supported.cast(x, type) and dictionary_encode() allow changing the type of columns in Arrow objects; as.numeric(), as.character(), etc. are exposed as similar type-altering conveniencesdplyr::between(); the Arrow version also allows the left and right arguments to be columns in the data and not just scalarsdplyr verb. This enables you to access Arrow functions that don’t have a direct R mapping. See list_compute_functions() for all available functions, which are available in dplyr prefixed by arrow_.dplyr::filter(arrow_dataset, string_column == 3) will error with a message about the type mismatch between the numeric 3 and the string type of string_column.open_dataset() now accepts a vector of file paths (or even a single file path). Among other things, this enables you to open a single very large file and use write_dataset() to partition it without having to read the whole file into memory.write_dataset() now defaults to format = "parquet" and better validates the format argumentschema in open_dataset() is now correctly handledScanner$Scan() method has been removed; use Scanner$ScanBatches()
value_counts() to tabulate values in an Array or ChunkedArray, similar to base::table().StructArray objects gain data.frame-like methods, including names(), $, [[, and dim().<-) with either $ or [[
Schema can now be edited by assigning in new types. This enables using the CSV reader to detect the schema of a file, modify the Schema object for any columns that you want to read in as a different type, and then use that Schema to read the data.Table with a schema, with columns of different lengths, and with scalar value recycling\0) characters, the error message now informs you that you can set options(arrow.skip_nul = TRUE) to strip them out. It is not recommended to set this option by default since this code path is significantly slower, and most string data does not contain nuls.read_json_arrow() now accepts a schema: read_json_arrow("file.json", schema = schema(col_a = float64(), col_b = string()))
vignette("install", package = "arrow") for details. This allows a faster, smaller package build in cases where that is useful, and it enables a minimal, functioning R package build on Solaris.FORCE_BUNDLED_BUILD=true.arrow now uses the mimalloc memory allocator by default on macOS, if available (as it is in CRAN binaries), instead of jemalloc. There are configuration issues with jemalloc on macOS, and benchmark analysis shows that this has negative effects on performance, especially on memory-intensive workflows. jemalloc remains the default on Linux; mimalloc is default on Windows.ARROW_DEFAULT_MEMORY_POOL environment variable to switch memory allocators now works correctly when the Arrow C++ library has been statically linked (as is usually the case when installing from CRAN).arrow_info() function now reports on the additional optional features, as well as the detected SIMD level. If key features or compression libraries are not enabled in the build, arrow_info() will refer to the installation vignette for guidance on how to install a more complete build, if desired.vignette("developing", package = "arrow").ARROW_HOME to point to a specific directory where the Arrow libraries are. This is similar to passing INCLUDE_DIR and LIB_DIR.flight_get() and flight_put() (renamed from push_data() in this release) can handle both Tables and RecordBatchesflight_put() gains an overwrite argument to optionally check for the existence of a resource with the the same namelist_flights() and flight_path_exists() enable you to see available resources on a Flight serverSchema objects now have r_to_py and py_to_r methods+, *, etc.) are supported on Arrays and ChunkedArrays and can be used in filter expressions in Arrow dplyr pipelines<-) with either $ or [[
names()
.data and .env are now fully supported in Arrow dplyr pipelines.arrow.skip_nul (default FALSE, as in base::scan()) allows conversion of Arrow string (utf8()) type data containing embedded nul \0 characters to R. If set to TRUE, nuls will be stripped and a warning is emitted if any are found.arrow_info() for an overview of various run-time and build-time Arrow configurations, useful for debuggingARROW_DEFAULT_MEMORY_POOL before loading the Arrow package to change memory allocators. Windows packages are built with mimalloc; most others are built with both jemalloc (used by default) and mimalloc. These alternative memory allocators are generally much faster than the system memory allocator, so they are used by default when available, but sometimes it is useful to turn them off for debugging purposes. To disable them, set ARROW_DEFAULT_MEMORY_POOL=system.sf tibbles to faithfully preserved and roundtripped (ARROW-10386).schema() for more details.write_parquet() can now write RecordBatchesreadr’s problems attribute is removed when converting to Arrow RecordBatch and table to prevent large amounts of metadata from accumulating inadvertently (ARROW-10624)SubTreeFileSystem gains a useful print method and no longer errors when printingr-arrow package are available with conda install -c arrow-nightlies -c conda-forge --strict-channel-priority r-arrow
cmake versionsvignette("install", package = "arrow"), especially for known CentOS issuesdistro package. If your OS isn’t correctly identified, please report an issue there.write_dataset() to Feather or Parquet files with partitioning. See the end of vignette("dataset", package = "arrow") for discussion and examples.head(), tail(), and take ([) methods. head() is optimized but the others may not be performant.collect() gains an as_data_frame argument, default TRUE but when FALSE allows you to evaluate the accumulated select and filter query but keep the result in Arrow, not an R data.frame
read_csv_arrow() supports specifying column types, both with a Schema and with the compact string representation for types used in the readr package. It also has gained a timestamp_parsers argument that lets you express a set of strptime parse strings that will be tried to convert columns designated as Timestamp type.libcurl and openssl, as well as a sufficiently modern compiler. See vignette("install", package = "arrow") for details.read_parquet(), write_feather(), et al.), as well as open_dataset() and write_dataset(), allow you to access resources on S3 (or on file systems that emulate S3) either by providing an s3:// URI or by providing a FileSystem$path(). See vignette("fs", package = "arrow") for examples.copy_files() allows you to recursively copy directories of files from one file system to another, such as from S3 to your local machine.Flight is a general-purpose client-server framework for high performance transport of large datasets over network interfaces. The arrow R package now provides methods for connecting to Flight RPC servers to send and receive data. See vignette("flight", package = "arrow") for an overview.
==, >, etc.) and boolean (&, |, !) operations, along with is.na, %in% and match (called match_arrow()), on Arrow Arrays and ChunkedArrays are now implemented in the C++ library.min(), max(), and unique() are implemented for Arrays and ChunkedArrays.dplyr filter expressions on Arrow Tables and RecordBatches are now evaluated in the C++ library, rather than by pulling data into R and evaluating. This yields significant performance improvements.dim() (nrow) for dplyr queries on Table/RecordBatch is now supportedarrow now depends on cpp11, which brings more robust UTF-8 handling and faster compilationInt64 type when all values fit with an R 32-bit integer now correctly inspects all chunks in a ChunkedArray, and this conversion can be disabled (so that Int64 always yields a bit64::integer64 vector) by setting options(arrow.int64_downcast = FALSE).ParquetFileReader has additional methods for accessing individual columns or row groups from the fileParquetFileWriter; invalid ArrowObject pointer from a saved R object; converting deeply nested structs from Arrow to Rproperties and arrow_properties arguments to write_parquet() are deprecated%in% expression now faithfully returns all relevant rows. or _; files and subdirectories starting with those prefixes are still ignoredopen_dataset("~/path") now correctly expands the pathversion option to write_parquet() is now correctly implementedparquet-cpp library has been fixedcmake is more robust, and you can now specify a /path/to/cmake by setting the CMAKE environment variablevignette("arrow", package = "arrow") includes tables that explain how R types are converted to Arrow types and vice versa.uint64, binary, fixed_size_binary, large_binary, large_utf8, large_list, list of structs.character vectors that exceed 2GB are converted to Arrow large_utf8 typePOSIXlt objects can now be converted to Arrow (struct)attributes() are preserved in Arrow metadata when converting to Arrow RecordBatch and table and are restored when converting from Arrow. This means that custom subclasses, such as haven::labelled, are preserved in round trip through Arrow.batch$metadata$new_key <- "new value"
int64, uint32, and uint64 now are converted to R integer if all values fit in boundsdate32 is now converted to R Date with double underlying storage. Even though the data values themselves are integers, this provides more strict round-trip fidelityfactor, dictionary ChunkedArrays that do not have identical dictionaries are properly unifiedRecordBatch{File,Stream}Writer will write V5, but you can specify an alternate metadata_version. For convenience, if you know the consumer you’re writing to cannot read V5, you can set the environment variable ARROW_PRE_1_0_METADATA_VERSION=1 to write V4 without changing any other code.ds <- open_dataset("s3://..."). Note that this currently requires a special C++ library build with additional dependencies–this is not yet available in CRAN releases or in nightly packages.sum() and mean() are implemented for Array and ChunkedArray
dimnames() and as.list()
reticulate
coerce_timestamps option to write_parquet() is now correctly implemented.type definition if provided by the userread_arrow and write_arrow are now deprecated; use the read/write_feather() and read/write_ipc_stream() functions depending on whether you’re working with the Arrow IPC file or stream format, respectively.FileStats, read_record_batch, and read_table have been removed.jemalloc included, and Windows packages use mimalloc
CC and CXX values that R usesdplyr 1.0reticulate::r_to_py() conversion now correctly works automatically, without having to call the method yourselfThis release includes support for version 2 of the Feather file format. Feather v2 features full support for all Arrow data types, fixes the 2GB per-column limitation for large amounts of string data, and it allows files to be compressed using either lz4 or zstd. write_feather() can write either version 2 or version 1 Feather files, and read_feather() automatically detects which file version it is reading.
Related to this change, several functions around reading and writing data have been reworked. read_ipc_stream() and write_ipc_stream() have been added to facilitate writing data to the Arrow IPC stream format, which is slightly different from the IPC file format (Feather v2 is the IPC file format).
Behavior has been standardized: all read_<format>() return an R data.frame (default) or a Table if the argument as_data_frame = FALSE; all write_<format>() functions return the data object, invisibly. To facilitate some workflows, a special write_to_raw() function is added to wrap write_ipc_stream() and return the raw vector containing the buffer that was written.
To achieve this standardization, read_table(), read_record_batch(), read_arrow(), and write_arrow() have been deprecated.
The 0.17 Apache Arrow release includes a C data interface that allows exchanging Arrow data in-process at the C level without copying and without libraries having a build or runtime dependency on each other. This enables us to use reticulate to share data between R and Python (pyarrow) efficiently.
See vignette("python", package = "arrow") for details.
dim() method, which sums rows across all files (#6635, @boshek)UnionDataset with the c() methodNA as FALSE, consistent with dplyr::filter()
vignette("dataset", package = "arrow") now has correct, executable codeNOT_CRAN=true. See vignette("install", package = "arrow") for details and more options.unify_schemas() to create a Schema containing the union of fields in multiple schemasread_feather() and other reader functions close any file connections they openR.oo package is also loadedFileStats is renamed to FileInfo, and the original spelling has been deprecatedinstall_arrow() now installs the latest release of arrow, including Linux dependencies, either for CRAN releases or for development builds (if nightly = TRUE)LIBARROW_DOWNLOAD or NOT_CRAN environment variable is setwrite_feather(), write_arrow() and write_parquet() now return their input, similar to the write_* functions in the readr package (#6387, @boshek)list and create a ListArray when all list elements are the same type (#6275, @michaelchirico)This release includes a dplyr interface to Arrow Datasets, which let you work efficiently with large, multi-file datasets as a single entity. Explore a directory of data files with open_dataset() and then use dplyr methods to select(), filter(), etc. Work will be done where possible in Arrow memory. When necessary, data is pulled into R for further computation. dplyr methods are conditionally loaded if you have dplyr available; it is not a hard dependency.
See vignette("dataset", package = "arrow") for details.
A source package installation (as from CRAN) will now handle its C++ dependencies automatically. For common Linux distributions and versions, installation will retrieve a prebuilt static C++ library for inclusion in the package; where this binary is not available, the package executes a bundled script that should build the Arrow C++ library with no system dependencies beyond what R requires.
See vignette("install", package = "arrow") for details.
Tables and RecordBatches also have dplyr methods.dplyr, [ methods for Tables, RecordBatches, Arrays, and ChunkedArrays now support natural row extraction operations. These use the C++ Filter, Slice, and Take methods for efficient access, depending on the type of selection vector.array_expression class has also been added, enabling among other things the ability to filter a Table with some function of Arrays, such as arrow_table[arrow_table$var1 > 5, ] without having to pull everything into R first.write_parquet() now supports compressioncodec_is_available() returns TRUE or FALSE whether the Arrow C++ library was built with support for a given compression library (e.g. gzip, lz4, snappy)character (as R factor levels are required to be) instead of raising an errorClass$create() methods. Notably, arrow::array() and arrow::table() have been removed in favor of Array$create() and Table$create(), eliminating the package startup message about masking base functions. For more information, see the new vignette("arrow").ARROW_PRE_0_15_IPC_FORMAT=1.as_tibble argument in the read_*() functions has been renamed to as_data_frame (ARROW-6337, @jameslamb)arrow::Column class has been removed, as it was removed from the C++ libraryTable and RecordBatch objects have S3 methods that enable you to work with them more like data.frames. Extract columns, subset, and so on. See ?Table and ?RecordBatch for examples.read_csv_arrow() supports more parsing options, including col_names, na, quoted_na, and skip
read_parquet() and read_feather() can ingest data from a raw vector (ARROW-6278)~/file.parquet (ARROW-6323)double()), and time types can be created with human-friendly resolution strings (“ms”, “s”, etc.). (ARROW-6338, ARROW-6364)Initial CRAN release of the arrow package. Key features include: