Tabular Datasets#

See also

Dataset API reference

The Arrow Datasets library provides functionality to efficiently work with tabular, potentially larger than memory, and multi-file datasets. This includes:

  • A unified interface that supports different sources and file formats and different file systems (local, cloud).

  • Discovery of sources (crawling directories, handling partitioned datasets with various partitioning schemes, basic schema normalization, …)

  • Optimized reading with predicate pushdown (filtering rows), projection (selecting and deriving columns), and optionally parallel reading.

The supported file formats currently are Parquet, Feather / Arrow IPC, CSV and ORC (note that ORC datasets can currently only be read and not yet written). The goal is to expand support to other file formats and data sources (e.g. database connections) in the future.

Reading Datasets#

For the examples below, let’s create a small dataset consisting of a directory with two parquet files:

49// Generate some data for the rest of this example.
50arrow::Result<std::shared_ptr<arrow::Table>> CreateTable() {
51  auto schema =
52      arrow::schema({arrow::field("a", arrow::int64()), arrow::field("b", arrow::int64()),
53                     arrow::field("c", arrow::int64())});
54  std::shared_ptr<arrow::Array> array_a;
55  std::shared_ptr<arrow::Array> array_b;
56  std::shared_ptr<arrow::Array> array_c;
57  arrow::NumericBuilder<arrow::Int64Type> builder;
58  ARROW_RETURN_NOT_OK(builder.AppendValues({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}));
59  ARROW_RETURN_NOT_OK(builder.Finish(&array_a));
60  builder.Reset();
61  ARROW_RETURN_NOT_OK(builder.AppendValues({9, 8, 7, 6, 5, 4, 3, 2, 1, 0}));
62  ARROW_RETURN_NOT_OK(builder.Finish(&array_b));
63  builder.Reset();
64  ARROW_RETURN_NOT_OK(builder.AppendValues({1, 2, 1, 2, 1, 2, 1, 2, 1, 2}));
65  ARROW_RETURN_NOT_OK(builder.Finish(&array_c));
66  return arrow::Table::Make(schema, {array_a, array_b, array_c});
67}
68
69// Set up a dataset by writing two Parquet files.
70arrow::Result<std::string> CreateExampleParquetDataset(
71    const std::shared_ptr<fs::FileSystem>& filesystem, const std::string& root_path) {
72  auto base_path = root_path + "/parquet_dataset";
73  ARROW_RETURN_NOT_OK(filesystem->CreateDir(base_path));
74  // Create an Arrow Table
75  ARROW_ASSIGN_OR_RAISE(auto table, CreateTable());
76  // Write it into two Parquet files
77  ARROW_ASSIGN_OR_RAISE(auto output,
78                        filesystem->OpenOutputStream(base_path + "/data1.parquet"));
79  ARROW_RETURN_NOT_OK(parquet::arrow::WriteTable(
80      *table->Slice(0, 5), arrow::default_memory_pool(), output, /*chunk_size=*/2048));
81  ARROW_ASSIGN_OR_RAISE(output,
82                        filesystem->OpenOutputStream(base_path + "/data2.parquet"));
83  ARROW_RETURN_NOT_OK(parquet::arrow::WriteTable(
84      *table->Slice(5), arrow::default_memory_pool(), output, /*chunk_size=*/2048));
85  return base_path;
86}

(See the full example at bottom: A note on transactions & ACID guarantees.)

Dataset discovery#

A arrow::dataset::Dataset object can be created using the various arrow::dataset::DatasetFactory objects. Here, we’ll use the arrow::dataset::FileSystemDatasetFactory, which can create a dataset given a base directory path:

162// Read the whole dataset with the given format, without partitioning.
163arrow::Result<std::shared_ptr<arrow::Table>> ScanWholeDataset(
164    const std::shared_ptr<fs::FileSystem>& filesystem,
165    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
166  // Create a dataset by scanning the filesystem for files
167  fs::FileSelector selector;
168  selector.base_dir = base_dir;
169  ARROW_ASSIGN_OR_RAISE(
170      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
171                                                       ds::FileSystemFactoryOptions()));
172  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
173  // Print out the fragments
174  ARROW_ASSIGN_OR_RAISE(auto fragments, dataset->GetFragments())
175  for (const auto& fragment : fragments) {
176    std::cout << "Found fragment: " << (*fragment)->ToString() << std::endl;
177  }
178  // Read the entire dataset as a Table
179  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
180  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
181  return scanner->ToTable();
182}

We’re also passing the filesystem to use and the file format to use for reading. This lets us choose between (for example) reading local files or files in Amazon S3, or between Parquet and CSV.

In addition to searching a base directory, we can list file paths manually.

Creating a arrow::dataset::Dataset does not begin reading the data itself. It only crawls the directory to find all the files (if needed), which can be retrieved with arrow::dataset::FileSystemDataset::files():

// Print out the files crawled (only for FileSystemDataset)
for (const auto& filename : dataset->files()) {
  std::cout << filename << std::endl;
}

…and infers the dataset’s schema (by default from the first file):

std::cout << dataset->schema()->ToString() << std::endl;

Using the arrow::dataset::Dataset::NewScan() method, we can build a arrow::dataset::Scanner and read the dataset (or a portion of it) into a arrow::Table with the arrow::dataset::Scanner::ToTable() method:

162// Read the whole dataset with the given format, without partitioning.
163arrow::Result<std::shared_ptr<arrow::Table>> ScanWholeDataset(
164    const std::shared_ptr<fs::FileSystem>& filesystem,
165    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
166  // Create a dataset by scanning the filesystem for files
167  fs::FileSelector selector;
168  selector.base_dir = base_dir;
169  ARROW_ASSIGN_OR_RAISE(
170      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
171                                                       ds::FileSystemFactoryOptions()));
172  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
173  // Print out the fragments
174  ARROW_ASSIGN_OR_RAISE(auto fragments, dataset->GetFragments())
175  for (const auto& fragment : fragments) {
176    std::cout << "Found fragment: " << (*fragment)->ToString() << std::endl;
177  }
178  // Read the entire dataset as a Table
179  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
180  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
181  return scanner->ToTable();
182}

Note

Depending on the size of your dataset, this can require a lot of memory; see Filtering data below on filtering/projecting.

Reading different file formats#

The above examples use Parquet files on local disk, but the Dataset API provides a consistent interface across multiple file formats and filesystems. (See Reading from cloud storage for more information on the latter.) Currently, Parquet, ORC, Feather / Arrow IPC, and CSV file formats are supported; more formats are planned in the future.

If we save the table as Feather files instead of Parquet files:

 90// Set up a dataset by writing two Feather files.
 91arrow::Result<std::string> CreateExampleFeatherDataset(
 92    const std::shared_ptr<fs::FileSystem>& filesystem, const std::string& root_path) {
 93  auto base_path = root_path + "/feather_dataset";
 94  ARROW_RETURN_NOT_OK(filesystem->CreateDir(base_path));
 95  // Create an Arrow Table
 96  ARROW_ASSIGN_OR_RAISE(auto table, CreateTable());
 97  // Write it into two Feather files
 98  ARROW_ASSIGN_OR_RAISE(auto output,
 99                        filesystem->OpenOutputStream(base_path + "/data1.feather"));
100  ARROW_ASSIGN_OR_RAISE(auto writer,
101                        arrow::ipc::MakeFileWriter(output.get(), table->schema()));
102  ARROW_RETURN_NOT_OK(writer->WriteTable(*table->Slice(0, 5)));
103  ARROW_RETURN_NOT_OK(writer->Close());
104  ARROW_ASSIGN_OR_RAISE(output,
105                        filesystem->OpenOutputStream(base_path + "/data2.feather"));
106  ARROW_ASSIGN_OR_RAISE(writer,
107                        arrow::ipc::MakeFileWriter(output.get(), table->schema()));
108  ARROW_RETURN_NOT_OK(writer->WriteTable(*table->Slice(5)));
109  ARROW_RETURN_NOT_OK(writer->Close());
110  return base_path;
111}

…then we can read the Feather file by passing an arrow::dataset::IpcFileFormat:

auto format = std::make_shared<ds::ParquetFileFormat>();
// ...
auto factory = ds::FileSystemDatasetFactory::Make(filesystem, selector, format, options)
                   .ValueOrDie();

Customizing file formats#

arrow::dataset::FileFormat objects have properties that control how files are read. For example:

auto format = std::make_shared<ds::ParquetFileFormat>();
format->reader_options.dict_columns.insert("a");

Will configure column "a" to be dictionary-encoded when read. Similarly, setting arrow::dataset::CsvFileFormat::parse_options lets us change things like reading comma-separated or tab-separated data.

Additionally, passing an arrow::dataset::FragmentScanOptions to arrow::dataset::ScannerBuilder::FragmentScanOptions() offers fine-grained control over data scanning. For example, for CSV files, we can change what values are converted into Boolean true and false at scan time.

Filtering data#

So far, we’ve been reading the entire dataset, but if we need only a subset of the data, this can waste time or memory reading data we don’t need. The arrow::dataset::Scanner offers control over what data to read.

In this snippet, we use arrow::dataset::ScannerBuilder::Project() to select which columns to read:

186// Read a dataset, but select only column "b" and only rows where b < 4.
187//
188// This is useful when you only want a few columns from a dataset. Where possible,
189// Datasets will push down the column selection such that less work is done.
190arrow::Result<std::shared_ptr<arrow::Table>> FilterAndSelectDataset(
191    const std::shared_ptr<fs::FileSystem>& filesystem,
192    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
193  fs::FileSelector selector;
194  selector.base_dir = base_dir;
195  ARROW_ASSIGN_OR_RAISE(
196      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
197                                                       ds::FileSystemFactoryOptions()));
198  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
199  // Read specified columns with a row filter
200  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
201  ARROW_RETURN_NOT_OK(scan_builder->Project({"b"}));
202  ARROW_RETURN_NOT_OK(scan_builder->Filter(cp::less(cp::field_ref("b"), cp::literal(4))));
203  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
204  return scanner->ToTable();
205}

Some formats, such as Parquet, can reduce I/O costs here by reading only the specified columns from the filesystem.

A filter can be provided with arrow::dataset::ScannerBuilder::Filter(), so that rows which do not match the filter predicate will not be included in the returned table. Again, some formats, such as Parquet, can use this filter to reduce the amount of I/O needed.

186// Read a dataset, but select only column "b" and only rows where b < 4.
187//
188// This is useful when you only want a few columns from a dataset. Where possible,
189// Datasets will push down the column selection such that less work is done.
190arrow::Result<std::shared_ptr<arrow::Table>> FilterAndSelectDataset(
191    const std::shared_ptr<fs::FileSystem>& filesystem,
192    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
193  fs::FileSelector selector;
194  selector.base_dir = base_dir;
195  ARROW_ASSIGN_OR_RAISE(
196      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
197                                                       ds::FileSystemFactoryOptions()));
198  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
199  // Read specified columns with a row filter
200  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
201  ARROW_RETURN_NOT_OK(scan_builder->Project({"b"}));
202  ARROW_RETURN_NOT_OK(scan_builder->Filter(cp::less(cp::field_ref("b"), cp::literal(4))));
203  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
204  return scanner->ToTable();
205}

Projecting columns#

In addition to selecting columns, arrow::dataset::ScannerBuilder::Project() can also be used for more complex projections, such as renaming columns, casting them to other types, and even deriving new columns based on evaluating expressions.

In this case, we pass a vector of expressions used to construct column values and a vector of names for the columns:

209// Read a dataset, but with column projection.
210//
211// This is useful to derive new columns from existing data. For example, here we
212// demonstrate casting a column to a different type, and turning a numeric column into a
213// boolean column based on a predicate. You could also rename columns or perform
214// computations involving multiple columns.
215arrow::Result<std::shared_ptr<arrow::Table>> ProjectDataset(
216    const std::shared_ptr<fs::FileSystem>& filesystem,
217    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
218  fs::FileSelector selector;
219  selector.base_dir = base_dir;
220  ARROW_ASSIGN_OR_RAISE(
221      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
222                                                       ds::FileSystemFactoryOptions()));
223  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
224  // Read specified columns with a row filter
225  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
226  ARROW_RETURN_NOT_OK(scan_builder->Project(
227      {
228          // Leave column "a" as-is.
229          cp::field_ref("a"),
230          // Cast column "b" to float32.
231          cp::call("cast", {cp::field_ref("b")},
232                   arrow::compute::CastOptions::Safe(arrow::float32())),
233          // Derive a boolean column from "c".
234          cp::equal(cp::field_ref("c"), cp::literal(1)),
235      },
236      {"a_renamed", "b_as_float32", "c_1"}));
237  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
238  return scanner->ToTable();
239}

This also determines the column selection; only the given columns will be present in the resulting table. If you want to include a derived column in addition to the existing columns, you can build up the expressions from the dataset schema:

243// Read a dataset, but with column projection.
244//
245// This time, we read all original columns plus one derived column. This simply combines
246// the previous two examples: selecting a subset of columns by name, and deriving new
247// columns with an expression.
248arrow::Result<std::shared_ptr<arrow::Table>> SelectAndProjectDataset(
249    const std::shared_ptr<fs::FileSystem>& filesystem,
250    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
251  fs::FileSelector selector;
252  selector.base_dir = base_dir;
253  ARROW_ASSIGN_OR_RAISE(
254      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
255                                                       ds::FileSystemFactoryOptions()));
256  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
257  // Read specified columns with a row filter
258  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
259  std::vector<std::string> names;
260  std::vector<cp::Expression> exprs;
261  // Read all the original columns.
262  for (const auto& field : dataset->schema()->fields()) {
263    names.push_back(field->name());
264    exprs.push_back(cp::field_ref(field->name()));
265  }
266  // Also derive a new column.
267  names.emplace_back("b_large");
268  exprs.push_back(cp::greater(cp::field_ref("b"), cp::literal(1)));
269  ARROW_RETURN_NOT_OK(scan_builder->Project(exprs, names));
270  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
271  return scanner->ToTable();
272}

Note

When combining filters and projections, Arrow will determine all necessary columns to read. For instance, if you filter on a column that isn’t ultimately selected, Arrow will still read the column to evaluate the filter.

Reading and writing partitioned data#

So far, we’ve been working with datasets consisting of flat directories with files. Oftentimes, a dataset will have one or more columns that are frequently filtered on. Instead of having to read and then filter the data, by organizing the files into a nested directory structure, we can define a partitioned dataset, where sub-directory names hold information about which subset of the data is stored in that directory. Then, we can more efficiently filter data by using that information to avoid loading files that don’t match the filter.

For example, a dataset partitioned by year and month may have the following layout:

dataset_name/
  year=2007/
    month=01/
       data0.parquet
       data1.parquet
       ...
    month=02/
       data0.parquet
       data1.parquet
       ...
    month=03/
    ...
  year=2008/
    month=01/
    ...
  ...

The above partitioning scheme is using “/key=value/” directory names, as found in Apache Hive. Under this convention, the file at dataset_name/year=2007/month=01/data0.parquet contains only data for which year == 2007 and month == 01.

Let’s create a small partitioned dataset. For this, we’ll use Arrow’s dataset writing functionality.

115// Set up a dataset by writing files with partitioning
116arrow::Result<std::string> CreateExampleParquetHivePartitionedDataset(
117    const std::shared_ptr<fs::FileSystem>& filesystem, const std::string& root_path) {
118  auto base_path = root_path + "/parquet_dataset";
119  ARROW_RETURN_NOT_OK(filesystem->CreateDir(base_path));
120  // Create an Arrow Table
121  auto schema = arrow::schema(
122      {arrow::field("a", arrow::int64()), arrow::field("b", arrow::int64()),
123       arrow::field("c", arrow::int64()), arrow::field("part", arrow::utf8())});
124  std::vector<std::shared_ptr<arrow::Array>> arrays(4);
125  arrow::NumericBuilder<arrow::Int64Type> builder;
126  ARROW_RETURN_NOT_OK(builder.AppendValues({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}));
127  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[0]));
128  builder.Reset();
129  ARROW_RETURN_NOT_OK(builder.AppendValues({9, 8, 7, 6, 5, 4, 3, 2, 1, 0}));
130  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[1]));
131  builder.Reset();
132  ARROW_RETURN_NOT_OK(builder.AppendValues({1, 2, 1, 2, 1, 2, 1, 2, 1, 2}));
133  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[2]));
134  arrow::StringBuilder string_builder;
135  ARROW_RETURN_NOT_OK(
136      string_builder.AppendValues({"a", "a", "a", "a", "a", "b", "b", "b", "b", "b"}));
137  ARROW_RETURN_NOT_OK(string_builder.Finish(&arrays[3]));
138  auto table = arrow::Table::Make(schema, arrays);
139  // Write it using Datasets
140  auto dataset = std::make_shared<ds::InMemoryDataset>(table);
141  ARROW_ASSIGN_OR_RAISE(auto scanner_builder, dataset->NewScan());
142  ARROW_ASSIGN_OR_RAISE(auto scanner, scanner_builder->Finish());
143
144  // The partition schema determines which fields are part of the partitioning.
145  auto partition_schema = arrow::schema({arrow::field("part", arrow::utf8())});
146  // We'll use Hive-style partitioning, which creates directories with "key=value" pairs.
147  auto partitioning = std::make_shared<ds::HivePartitioning>(partition_schema);
148  // We'll write Parquet files.
149  auto format = std::make_shared<ds::ParquetFileFormat>();
150  ds::FileSystemDatasetWriteOptions write_options;
151  write_options.file_write_options = format->DefaultWriteOptions();
152  write_options.filesystem = filesystem;
153  write_options.base_dir = base_path;
154  write_options.partitioning = partitioning;
155  write_options.basename_template = "part{i}.parquet";
156  ARROW_RETURN_NOT_OK(ds::FileSystemDataset::Write(write_options, scanner));
157  return base_path;
158}

The above created a directory with two subdirectories (“part=a” and “part=b”), and the Parquet files written in those directories no longer include the “part” column.

Reading this dataset, we now specify that the dataset should use a Hive-like partitioning scheme:

276// Read an entire dataset, but with partitioning information.
277arrow::Result<std::shared_ptr<arrow::Table>> ScanPartitionedDataset(
278    const std::shared_ptr<fs::FileSystem>& filesystem,
279    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
280  fs::FileSelector selector;
281  selector.base_dir = base_dir;
282  selector.recursive = true;  // Make sure to search subdirectories
283  ds::FileSystemFactoryOptions options;
284  // We'll use Hive-style partitioning. We'll let Arrow Datasets infer the partition
285  // schema.
286  options.partitioning = ds::HivePartitioning::MakeFactory();
287  ARROW_ASSIGN_OR_RAISE(auto factory, ds::FileSystemDatasetFactory::Make(
288                                          filesystem, selector, format, options));
289  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
290  // Print out the fragments
291  ARROW_ASSIGN_OR_RAISE(auto fragments, dataset->GetFragments());
292  for (const auto& fragment : fragments) {
293    std::cout << "Found fragment: " << (*fragment)->ToString() << std::endl;
294    std::cout << "Partition expression: "
295              << (*fragment)->partition_expression().ToString() << std::endl;
296  }
297  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
298  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
299  return scanner->ToTable();
300}

Although the partition fields are not included in the actual Parquet files, they will be added back to the resulting table when scanning this dataset:

$ ./debug/dataset_documentation_example file:///tmp parquet_hive partitioned
Found fragment: /tmp/parquet_dataset/part=a/part0.parquet
Partition expression: (part == "a")
Found fragment: /tmp/parquet_dataset/part=b/part1.parquet
Partition expression: (part == "b")
Read 20 rows
a: int64
  -- field metadata --
  PARQUET:field_id: '1'
b: double
  -- field metadata --
  PARQUET:field_id: '2'
c: int64
  -- field metadata --
  PARQUET:field_id: '3'
part: string
----
# snip...

We can now filter on the partition keys, which avoids loading files altogether if they do not match the filter:

304// Read an entire dataset, but with partitioning information. Also, filter the dataset on
305// the partition values.
306arrow::Result<std::shared_ptr<arrow::Table>> FilterPartitionedDataset(
307    const std::shared_ptr<fs::FileSystem>& filesystem,
308    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
309  fs::FileSelector selector;
310  selector.base_dir = base_dir;
311  selector.recursive = true;
312  ds::FileSystemFactoryOptions options;
313  options.partitioning = ds::HivePartitioning::MakeFactory();
314  ARROW_ASSIGN_OR_RAISE(auto factory, ds::FileSystemDatasetFactory::Make(
315                                          filesystem, selector, format, options));
316  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
317  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
318  // Filter based on the partition values. This will mean that we won't even read the
319  // files whose partition expressions don't match the filter.
320  ARROW_RETURN_NOT_OK(
321      scan_builder->Filter(cp::equal(cp::field_ref("part"), cp::literal("b"))));
322  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
323  return scanner->ToTable();
324}

Different partitioning schemes#

The above example uses a Hive-like directory scheme, such as “/year=2009/month=11/day=15”. We specified this by passing the Hive partitioning factory. In this case, the types of the partition keys are inferred from the file paths.

It is also possible to directly construct the partitioning and explicitly define the schema of the partition keys. For example:

auto part = std::make_shared<ds::HivePartitioning>(arrow::schema({
    arrow::field("year", arrow::int16()),
    arrow::field("month", arrow::int8()),
    arrow::field("day", arrow::int32())
}));

Arrow supports another partitioning scheme, “directory partitioning”, where the segments in the file path represent the values of the partition keys without including the name (the field names are implicit in the segment’s index). For example, given field names “year”, “month”, and “day”, one path might be “/2019/11/15”.

Since the names are not included in the file paths, these must be specified when constructing a directory partitioning:

auto part = ds::DirectoryPartitioning::MakeFactory({"year", "month", "day"});

Directory partitioning also supports providing a full schema rather than inferring types from file paths.

Partitioning performance considerations#

Partitioning datasets has two aspects that affect performance: it increases the number of files and it creates a directory structure around the files. Both of these have benefits as well as costs. Depending on the configuration and the size of your dataset, the costs can outweigh the benefits.

Because partitions split up the dataset into multiple files, partitioned datasets can be read and written with parallelism. However, each additional file adds a little overhead in processing for filesystem interaction. It also increases the overall dataset size since each file has some shared metadata. For example, each parquet file contains the schema and group-level statistics. The number of partitions is a floor for the number of files. If you partition a dataset by date with a year of data, you will have at least 365 files. If you further partition by another dimension with 1,000 unique values, you will have up to 365,000 files. This fine of partitioning often leads to small files that mostly consist of metadata.

Partitioned datasets create nested folder structures, and those allow us to prune which files are loaded in a scan. However, this adds overhead to discovering files in the dataset, as we’ll need to recursively “list directory” to find the data files. Too fine partitions can cause problems here: Partitioning a dataset by date for a years worth of data will require 365 list calls to find all the files; adding another column with cardinality 1,000 will make that 365,365 calls.

The most optimal partitioning layout will depend on your data, access patterns, and which systems will be reading the data. Most systems, including Arrow, should work across a range of file sizes and partitioning layouts, but there are extremes you should avoid. These guidelines can help avoid some known worst cases:

  • Avoid files smaller than 20MB and larger than 2GB.

  • Avoid partitioning layouts with more than 10,000 distinct partitions.

For file formats that have a notion of groups within a file, such as Parquet, similar guidelines apply. Row groups can provide parallelism when reading and allow data skipping based on statistics, but very small groups can cause metadata to be a significant portion of file size. Arrow’s file writer provides sensible defaults for group sizing in most cases.

Reading from other data sources#

Reading in-memory data#

If you already have data in memory that you’d like to use with the Datasets API (e.g. to filter/project data, or to write it out to a filesystem), you can wrap it in an arrow::dataset::InMemoryDataset:

auto table = arrow::Table::FromRecordBatches(...);
auto dataset = std::make_shared<arrow::dataset::InMemoryDataset>(std::move(table));
// Scan the dataset, filter, it, etc.
auto scanner_builder = dataset->NewScan();

In the example, we used the InMemoryDataset to write our example data to local disk which was used in the rest of the example:

115// Set up a dataset by writing files with partitioning
116arrow::Result<std::string> CreateExampleParquetHivePartitionedDataset(
117    const std::shared_ptr<fs::FileSystem>& filesystem, const std::string& root_path) {
118  auto base_path = root_path + "/parquet_dataset";
119  ARROW_RETURN_NOT_OK(filesystem->CreateDir(base_path));
120  // Create an Arrow Table
121  auto schema = arrow::schema(
122      {arrow::field("a", arrow::int64()), arrow::field("b", arrow::int64()),
123       arrow::field("c", arrow::int64()), arrow::field("part", arrow::utf8())});
124  std::vector<std::shared_ptr<arrow::Array>> arrays(4);
125  arrow::NumericBuilder<arrow::Int64Type> builder;
126  ARROW_RETURN_NOT_OK(builder.AppendValues({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}));
127  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[0]));
128  builder.Reset();
129  ARROW_RETURN_NOT_OK(builder.AppendValues({9, 8, 7, 6, 5, 4, 3, 2, 1, 0}));
130  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[1]));
131  builder.Reset();
132  ARROW_RETURN_NOT_OK(builder.AppendValues({1, 2, 1, 2, 1, 2, 1, 2, 1, 2}));
133  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[2]));
134  arrow::StringBuilder string_builder;
135  ARROW_RETURN_NOT_OK(
136      string_builder.AppendValues({"a", "a", "a", "a", "a", "b", "b", "b", "b", "b"}));
137  ARROW_RETURN_NOT_OK(string_builder.Finish(&arrays[3]));
138  auto table = arrow::Table::Make(schema, arrays);
139  // Write it using Datasets
140  auto dataset = std::make_shared<ds::InMemoryDataset>(table);
141  ARROW_ASSIGN_OR_RAISE(auto scanner_builder, dataset->NewScan());
142  ARROW_ASSIGN_OR_RAISE(auto scanner, scanner_builder->Finish());
143
144  // The partition schema determines which fields are part of the partitioning.
145  auto partition_schema = arrow::schema({arrow::field("part", arrow::utf8())});
146  // We'll use Hive-style partitioning, which creates directories with "key=value" pairs.
147  auto partitioning = std::make_shared<ds::HivePartitioning>(partition_schema);
148  // We'll write Parquet files.
149  auto format = std::make_shared<ds::ParquetFileFormat>();
150  ds::FileSystemDatasetWriteOptions write_options;
151  write_options.file_write_options = format->DefaultWriteOptions();
152  write_options.filesystem = filesystem;
153  write_options.base_dir = base_path;
154  write_options.partitioning = partitioning;
155  write_options.basename_template = "part{i}.parquet";
156  ARROW_RETURN_NOT_OK(ds::FileSystemDataset::Write(write_options, scanner));
157  return base_path;
158}

Reading from cloud storage#

In addition to local files, Arrow Datasets also support reading from cloud storage systems, such as Amazon S3, by passing a different filesystem.

See the filesystem docs for more details on the available filesystems.

A note on transactions & ACID guarantees#

The dataset API offers no transaction support or any ACID guarantees. This affects both reading and writing. Concurrent reads are fine. Concurrent writes or writes concurring with reads may have unexpected behavior. Various approaches can be used to avoid operating on the same files such as using a unique basename template for each writer, a temporary directory for new files, or separate storage of the file list instead of relying on directory discovery.

Unexpectedly killing the process while a write is in progress can leave the system in an inconsistent state. Write calls generally return as soon as the bytes to be written have been completely delivered to the OS page cache. Even though a write operation has been completed it is possible for part of the file to be lost if there is a sudden power loss immediately after the write call.

Most file formats have magic numbers which are written at the end. This means a partial file write can safely be detected and discarded. The CSV file format does not have any such concept and a partially written CSV file may be detected as valid.

Full Example#

  1// Licensed to the Apache Software Foundation (ASF) under one
  2// or more contributor license agreements. See the NOTICE file
  3// distributed with this work for additional information
  4// regarding copyright ownership. The ASF licenses this file
  5// to you under the Apache License, Version 2.0 (the
  6// "License"); you may not use this file except in compliance
  7// with the License. You may obtain a copy of the License at
  8//
  9// http://www.apache.org/licenses/LICENSE-2.0
 10//
 11// Unless required by applicable law or agreed to in writing,
 12// software distributed under the License is distributed on an
 13// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
 14// KIND, either express or implied. See the License for the
 15// specific language governing permissions and limitations
 16// under the License.
 17
 18// This example showcases various ways to work with Datasets. It's
 19// intended to be paired with the documentation.
 20
 21#include <arrow/api.h>
 22#include <arrow/compute/cast.h>
 23#include <arrow/dataset/dataset.h>
 24#include <arrow/dataset/discovery.h>
 25#include <arrow/dataset/file_base.h>
 26#include <arrow/dataset/file_ipc.h>
 27#include <arrow/dataset/file_parquet.h>
 28#include <arrow/dataset/scanner.h>
 29#include <arrow/filesystem/filesystem.h>
 30#include <arrow/ipc/writer.h>
 31#include <arrow/util/iterator.h>
 32#include <parquet/arrow/writer.h>
 33#include "arrow/compute/expression.h"
 34
 35#include <iostream>
 36#include <vector>
 37
 38namespace ds = arrow::dataset;
 39namespace fs = arrow::fs;
 40namespace cp = arrow::compute;
 41
 42/**
 43 * \brief Run Example
 44 *
 45 * ./debug/dataset-documentation-example file:///<some_path>/<some_directory> parquet
 46 */
 47
 48// (Doc section: Reading Datasets)
 49// Generate some data for the rest of this example.
 50arrow::Result<std::shared_ptr<arrow::Table>> CreateTable() {
 51  auto schema =
 52      arrow::schema({arrow::field("a", arrow::int64()), arrow::field("b", arrow::int64()),
 53                     arrow::field("c", arrow::int64())});
 54  std::shared_ptr<arrow::Array> array_a;
 55  std::shared_ptr<arrow::Array> array_b;
 56  std::shared_ptr<arrow::Array> array_c;
 57  arrow::NumericBuilder<arrow::Int64Type> builder;
 58  ARROW_RETURN_NOT_OK(builder.AppendValues({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}));
 59  ARROW_RETURN_NOT_OK(builder.Finish(&array_a));
 60  builder.Reset();
 61  ARROW_RETURN_NOT_OK(builder.AppendValues({9, 8, 7, 6, 5, 4, 3, 2, 1, 0}));
 62  ARROW_RETURN_NOT_OK(builder.Finish(&array_b));
 63  builder.Reset();
 64  ARROW_RETURN_NOT_OK(builder.AppendValues({1, 2, 1, 2, 1, 2, 1, 2, 1, 2}));
 65  ARROW_RETURN_NOT_OK(builder.Finish(&array_c));
 66  return arrow::Table::Make(schema, {array_a, array_b, array_c});
 67}
 68
 69// Set up a dataset by writing two Parquet files.
 70arrow::Result<std::string> CreateExampleParquetDataset(
 71    const std::shared_ptr<fs::FileSystem>& filesystem, const std::string& root_path) {
 72  auto base_path = root_path + "/parquet_dataset";
 73  ARROW_RETURN_NOT_OK(filesystem->CreateDir(base_path));
 74  // Create an Arrow Table
 75  ARROW_ASSIGN_OR_RAISE(auto table, CreateTable());
 76  // Write it into two Parquet files
 77  ARROW_ASSIGN_OR_RAISE(auto output,
 78                        filesystem->OpenOutputStream(base_path + "/data1.parquet"));
 79  ARROW_RETURN_NOT_OK(parquet::arrow::WriteTable(
 80      *table->Slice(0, 5), arrow::default_memory_pool(), output, /*chunk_size=*/2048));
 81  ARROW_ASSIGN_OR_RAISE(output,
 82                        filesystem->OpenOutputStream(base_path + "/data2.parquet"));
 83  ARROW_RETURN_NOT_OK(parquet::arrow::WriteTable(
 84      *table->Slice(5), arrow::default_memory_pool(), output, /*chunk_size=*/2048));
 85  return base_path;
 86}
 87// (Doc section: Reading Datasets)
 88
 89// (Doc section: Reading different file formats)
 90// Set up a dataset by writing two Feather files.
 91arrow::Result<std::string> CreateExampleFeatherDataset(
 92    const std::shared_ptr<fs::FileSystem>& filesystem, const std::string& root_path) {
 93  auto base_path = root_path + "/feather_dataset";
 94  ARROW_RETURN_NOT_OK(filesystem->CreateDir(base_path));
 95  // Create an Arrow Table
 96  ARROW_ASSIGN_OR_RAISE(auto table, CreateTable());
 97  // Write it into two Feather files
 98  ARROW_ASSIGN_OR_RAISE(auto output,
 99                        filesystem->OpenOutputStream(base_path + "/data1.feather"));
100  ARROW_ASSIGN_OR_RAISE(auto writer,
101                        arrow::ipc::MakeFileWriter(output.get(), table->schema()));
102  ARROW_RETURN_NOT_OK(writer->WriteTable(*table->Slice(0, 5)));
103  ARROW_RETURN_NOT_OK(writer->Close());
104  ARROW_ASSIGN_OR_RAISE(output,
105                        filesystem->OpenOutputStream(base_path + "/data2.feather"));
106  ARROW_ASSIGN_OR_RAISE(writer,
107                        arrow::ipc::MakeFileWriter(output.get(), table->schema()));
108  ARROW_RETURN_NOT_OK(writer->WriteTable(*table->Slice(5)));
109  ARROW_RETURN_NOT_OK(writer->Close());
110  return base_path;
111}
112// (Doc section: Reading different file formats)
113
114// (Doc section: Reading and writing partitioned data)
115// Set up a dataset by writing files with partitioning
116arrow::Result<std::string> CreateExampleParquetHivePartitionedDataset(
117    const std::shared_ptr<fs::FileSystem>& filesystem, const std::string& root_path) {
118  auto base_path = root_path + "/parquet_dataset";
119  ARROW_RETURN_NOT_OK(filesystem->CreateDir(base_path));
120  // Create an Arrow Table
121  auto schema = arrow::schema(
122      {arrow::field("a", arrow::int64()), arrow::field("b", arrow::int64()),
123       arrow::field("c", arrow::int64()), arrow::field("part", arrow::utf8())});
124  std::vector<std::shared_ptr<arrow::Array>> arrays(4);
125  arrow::NumericBuilder<arrow::Int64Type> builder;
126  ARROW_RETURN_NOT_OK(builder.AppendValues({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}));
127  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[0]));
128  builder.Reset();
129  ARROW_RETURN_NOT_OK(builder.AppendValues({9, 8, 7, 6, 5, 4, 3, 2, 1, 0}));
130  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[1]));
131  builder.Reset();
132  ARROW_RETURN_NOT_OK(builder.AppendValues({1, 2, 1, 2, 1, 2, 1, 2, 1, 2}));
133  ARROW_RETURN_NOT_OK(builder.Finish(&arrays[2]));
134  arrow::StringBuilder string_builder;
135  ARROW_RETURN_NOT_OK(
136      string_builder.AppendValues({"a", "a", "a", "a", "a", "b", "b", "b", "b", "b"}));
137  ARROW_RETURN_NOT_OK(string_builder.Finish(&arrays[3]));
138  auto table = arrow::Table::Make(schema, arrays);
139  // Write it using Datasets
140  auto dataset = std::make_shared<ds::InMemoryDataset>(table);
141  ARROW_ASSIGN_OR_RAISE(auto scanner_builder, dataset->NewScan());
142  ARROW_ASSIGN_OR_RAISE(auto scanner, scanner_builder->Finish());
143
144  // The partition schema determines which fields are part of the partitioning.
145  auto partition_schema = arrow::schema({arrow::field("part", arrow::utf8())});
146  // We'll use Hive-style partitioning, which creates directories with "key=value" pairs.
147  auto partitioning = std::make_shared<ds::HivePartitioning>(partition_schema);
148  // We'll write Parquet files.
149  auto format = std::make_shared<ds::ParquetFileFormat>();
150  ds::FileSystemDatasetWriteOptions write_options;
151  write_options.file_write_options = format->DefaultWriteOptions();
152  write_options.filesystem = filesystem;
153  write_options.base_dir = base_path;
154  write_options.partitioning = partitioning;
155  write_options.basename_template = "part{i}.parquet";
156  ARROW_RETURN_NOT_OK(ds::FileSystemDataset::Write(write_options, scanner));
157  return base_path;
158}
159// (Doc section: Reading and writing partitioned data)
160
161// (Doc section: Dataset discovery)
162// Read the whole dataset with the given format, without partitioning.
163arrow::Result<std::shared_ptr<arrow::Table>> ScanWholeDataset(
164    const std::shared_ptr<fs::FileSystem>& filesystem,
165    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
166  // Create a dataset by scanning the filesystem for files
167  fs::FileSelector selector;
168  selector.base_dir = base_dir;
169  ARROW_ASSIGN_OR_RAISE(
170      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
171                                                       ds::FileSystemFactoryOptions()));
172  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
173  // Print out the fragments
174  ARROW_ASSIGN_OR_RAISE(auto fragments, dataset->GetFragments())
175  for (const auto& fragment : fragments) {
176    std::cout << "Found fragment: " << (*fragment)->ToString() << std::endl;
177  }
178  // Read the entire dataset as a Table
179  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
180  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
181  return scanner->ToTable();
182}
183// (Doc section: Dataset discovery)
184
185// (Doc section: Filtering data)
186// Read a dataset, but select only column "b" and only rows where b < 4.
187//
188// This is useful when you only want a few columns from a dataset. Where possible,
189// Datasets will push down the column selection such that less work is done.
190arrow::Result<std::shared_ptr<arrow::Table>> FilterAndSelectDataset(
191    const std::shared_ptr<fs::FileSystem>& filesystem,
192    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
193  fs::FileSelector selector;
194  selector.base_dir = base_dir;
195  ARROW_ASSIGN_OR_RAISE(
196      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
197                                                       ds::FileSystemFactoryOptions()));
198  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
199  // Read specified columns with a row filter
200  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
201  ARROW_RETURN_NOT_OK(scan_builder->Project({"b"}));
202  ARROW_RETURN_NOT_OK(scan_builder->Filter(cp::less(cp::field_ref("b"), cp::literal(4))));
203  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
204  return scanner->ToTable();
205}
206// (Doc section: Filtering data)
207
208// (Doc section: Projecting columns)
209// Read a dataset, but with column projection.
210//
211// This is useful to derive new columns from existing data. For example, here we
212// demonstrate casting a column to a different type, and turning a numeric column into a
213// boolean column based on a predicate. You could also rename columns or perform
214// computations involving multiple columns.
215arrow::Result<std::shared_ptr<arrow::Table>> ProjectDataset(
216    const std::shared_ptr<fs::FileSystem>& filesystem,
217    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
218  fs::FileSelector selector;
219  selector.base_dir = base_dir;
220  ARROW_ASSIGN_OR_RAISE(
221      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
222                                                       ds::FileSystemFactoryOptions()));
223  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
224  // Read specified columns with a row filter
225  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
226  ARROW_RETURN_NOT_OK(scan_builder->Project(
227      {
228          // Leave column "a" as-is.
229          cp::field_ref("a"),
230          // Cast column "b" to float32.
231          cp::call("cast", {cp::field_ref("b")},
232                   arrow::compute::CastOptions::Safe(arrow::float32())),
233          // Derive a boolean column from "c".
234          cp::equal(cp::field_ref("c"), cp::literal(1)),
235      },
236      {"a_renamed", "b_as_float32", "c_1"}));
237  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
238  return scanner->ToTable();
239}
240// (Doc section: Projecting columns)
241
242// (Doc section: Projecting columns #2)
243// Read a dataset, but with column projection.
244//
245// This time, we read all original columns plus one derived column. This simply combines
246// the previous two examples: selecting a subset of columns by name, and deriving new
247// columns with an expression.
248arrow::Result<std::shared_ptr<arrow::Table>> SelectAndProjectDataset(
249    const std::shared_ptr<fs::FileSystem>& filesystem,
250    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
251  fs::FileSelector selector;
252  selector.base_dir = base_dir;
253  ARROW_ASSIGN_OR_RAISE(
254      auto factory, ds::FileSystemDatasetFactory::Make(filesystem, selector, format,
255                                                       ds::FileSystemFactoryOptions()));
256  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
257  // Read specified columns with a row filter
258  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
259  std::vector<std::string> names;
260  std::vector<cp::Expression> exprs;
261  // Read all the original columns.
262  for (const auto& field : dataset->schema()->fields()) {
263    names.push_back(field->name());
264    exprs.push_back(cp::field_ref(field->name()));
265  }
266  // Also derive a new column.
267  names.emplace_back("b_large");
268  exprs.push_back(cp::greater(cp::field_ref("b"), cp::literal(1)));
269  ARROW_RETURN_NOT_OK(scan_builder->Project(exprs, names));
270  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
271  return scanner->ToTable();
272}
273// (Doc section: Projecting columns #2)
274
275// (Doc section: Reading and writing partitioned data #2)
276// Read an entire dataset, but with partitioning information.
277arrow::Result<std::shared_ptr<arrow::Table>> ScanPartitionedDataset(
278    const std::shared_ptr<fs::FileSystem>& filesystem,
279    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
280  fs::FileSelector selector;
281  selector.base_dir = base_dir;
282  selector.recursive = true;  // Make sure to search subdirectories
283  ds::FileSystemFactoryOptions options;
284  // We'll use Hive-style partitioning. We'll let Arrow Datasets infer the partition
285  // schema.
286  options.partitioning = ds::HivePartitioning::MakeFactory();
287  ARROW_ASSIGN_OR_RAISE(auto factory, ds::FileSystemDatasetFactory::Make(
288                                          filesystem, selector, format, options));
289  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
290  // Print out the fragments
291  ARROW_ASSIGN_OR_RAISE(auto fragments, dataset->GetFragments());
292  for (const auto& fragment : fragments) {
293    std::cout << "Found fragment: " << (*fragment)->ToString() << std::endl;
294    std::cout << "Partition expression: "
295              << (*fragment)->partition_expression().ToString() << std::endl;
296  }
297  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
298  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
299  return scanner->ToTable();
300}
301// (Doc section: Reading and writing partitioned data #2)
302
303// (Doc section: Reading and writing partitioned data #3)
304// Read an entire dataset, but with partitioning information. Also, filter the dataset on
305// the partition values.
306arrow::Result<std::shared_ptr<arrow::Table>> FilterPartitionedDataset(
307    const std::shared_ptr<fs::FileSystem>& filesystem,
308    const std::shared_ptr<ds::FileFormat>& format, const std::string& base_dir) {
309  fs::FileSelector selector;
310  selector.base_dir = base_dir;
311  selector.recursive = true;
312  ds::FileSystemFactoryOptions options;
313  options.partitioning = ds::HivePartitioning::MakeFactory();
314  ARROW_ASSIGN_OR_RAISE(auto factory, ds::FileSystemDatasetFactory::Make(
315                                          filesystem, selector, format, options));
316  ARROW_ASSIGN_OR_RAISE(auto dataset, factory->Finish());
317  ARROW_ASSIGN_OR_RAISE(auto scan_builder, dataset->NewScan());
318  // Filter based on the partition values. This will mean that we won't even read the
319  // files whose partition expressions don't match the filter.
320  ARROW_RETURN_NOT_OK(
321      scan_builder->Filter(cp::equal(cp::field_ref("part"), cp::literal("b"))));
322  ARROW_ASSIGN_OR_RAISE(auto scanner, scan_builder->Finish());
323  return scanner->ToTable();
324}
325// (Doc section: Reading and writing partitioned data #3)
326
327arrow::Status RunDatasetDocumentation(const std::string& format_name,
328                                      const std::string& uri, const std::string& mode) {
329  std::string base_path;
330  std::shared_ptr<ds::FileFormat> format;
331  std::string root_path;
332  ARROW_ASSIGN_OR_RAISE(auto fs, fs::FileSystemFromUri(uri, &root_path));
333
334  if (format_name == "feather") {
335    format = std::make_shared<ds::IpcFileFormat>();
336    ARROW_ASSIGN_OR_RAISE(base_path, CreateExampleFeatherDataset(fs, root_path));
337  } else if (format_name == "parquet") {
338    format = std::make_shared<ds::ParquetFileFormat>();
339    ARROW_ASSIGN_OR_RAISE(base_path, CreateExampleParquetDataset(fs, root_path));
340  } else if (format_name == "parquet_hive") {
341    format = std::make_shared<ds::ParquetFileFormat>();
342    ARROW_ASSIGN_OR_RAISE(base_path,
343                          CreateExampleParquetHivePartitionedDataset(fs, root_path));
344  } else {
345    std::cerr << "Unknown format: " << format_name << std::endl;
346    std::cerr << "Supported formats: feather, parquet, parquet_hive" << std::endl;
347    return arrow::Status::ExecutionError("Dataset creating failed.");
348  }
349
350  std::shared_ptr<arrow::Table> table;
351  if (mode == "no_filter") {
352    ARROW_ASSIGN_OR_RAISE(table, ScanWholeDataset(fs, format, base_path));
353  } else if (mode == "filter") {
354    ARROW_ASSIGN_OR_RAISE(table, FilterAndSelectDataset(fs, format, base_path));
355  } else if (mode == "project") {
356    ARROW_ASSIGN_OR_RAISE(table, ProjectDataset(fs, format, base_path));
357  } else if (mode == "select_project") {
358    ARROW_ASSIGN_OR_RAISE(table, SelectAndProjectDataset(fs, format, base_path));
359  } else if (mode == "partitioned") {
360    ARROW_ASSIGN_OR_RAISE(table, ScanPartitionedDataset(fs, format, base_path));
361  } else if (mode == "filter_partitioned") {
362    ARROW_ASSIGN_OR_RAISE(table, FilterPartitionedDataset(fs, format, base_path));
363  } else {
364    std::cerr << "Unknown mode: " << mode << std::endl;
365    std::cerr
366        << "Supported modes: no_filter, filter, project, select_project, partitioned"
367        << std::endl;
368    return arrow::Status::ExecutionError("Dataset reading failed.");
369  }
370  std::cout << "Read " << table->num_rows() << " rows" << std::endl;
371  std::cout << table->ToString() << std::endl;
372  return arrow::Status::OK();
373}
374
375int main(int argc, char** argv) {
376  if (argc < 3) {
377    // Fake success for CI purposes.
378    return EXIT_SUCCESS;
379  }
380
381  std::string uri = argv[1];
382  std::string format_name = argv[2];
383  std::string mode = argc > 3 ? argv[3] : "no_filter";
384
385  auto status = RunDatasetDocumentation(format_name, uri, mode);
386  if (!status.ok()) {
387    std::cerr << status.ToString() << std::endl;
388    return EXIT_FAILURE;
389  }
390  return EXIT_SUCCESS;
391}