Dataset#

Warning

Experimental: The Java module dataset is currently under early development. API might be changed in each release of Apache Arrow until it gets mature.

Dataset is an universal layer in Apache Arrow for querying data in different formats or in different partitioning strategies. Usually the data to be queried is supposed to be located from a traditional file system, however Arrow Dataset is not designed only for querying files but can be extended to serve all possible data sources such as from inter-process communication or from other network locations, etc.

Getting Started#

Currently supported file formats are:

  • Apache Arrow (.arrow)

  • Apache ORC (.orc)

  • Apache Parquet (.parquet)

  • Comma-Separated Values (.csv)

  • Line-delimited JSON Values (.json)

Below shows a simplest example of using Dataset to query a Parquet file in Java:

// read data from file /opt/example.parquet
String uri = "file:/opt/example.parquet";
ScanOptions options = new ScanOptions(/*batchSize*/ 32768);
try (
    BufferAllocator allocator = new RootAllocator();
    DatasetFactory datasetFactory = new FileSystemDatasetFactory(
            allocator, NativeMemoryPool.getDefault(),
            FileFormat.PARQUET, uri);
    Dataset dataset = datasetFactory.finish();
    Scanner scanner = dataset.newScan(options);
    ArrowReader reader = scanner.scanBatches()
) {
    List<ArrowRecordBatch> batches = new ArrayList<>();
    while (reader.loadNextBatch()) {
        try (VectorSchemaRoot root = reader.getVectorSchemaRoot()) {
            final VectorUnloader unloader = new VectorUnloader(root);
            batches.add(unloader.getRecordBatch());
        }
    }

    // do something with read record batches, for example:
    analyzeArrowData(batches);

    // finished the analysis of the data, close all resources:
    AutoCloseables.close(batches);
} catch (Exception e) {
    e.printStackTrace();
}

Note

ArrowRecordBatch is a low-level composite Arrow data exchange format that doesn’t provide API to read typed data from it directly. It’s recommended to use utilities VectorLoader to load it into a schema aware container VectorSchemaRoot by which user could be able to access decoded data conveniently in Java.

The ScanOptions batchSize argument takes effect only if it is set to a value smaller than the number of rows in the recordbatch.

See also

Load record batches with VectorSchemaRoot.

Schema#

Schema of the data to be queried can be inspected via method DatasetFactory#inspect() before actually reading it. For example:

// read data from local file /opt/example.parquet
String uri = "file:/opt/example.parquet";
BufferAllocator allocator = new RootAllocator(Long.MAX_VALUE);
DatasetFactory factory = new FileSystemDatasetFactory(allocator,
    NativeMemoryPool.getDefault(), FileFormat.PARQUET, uri);

// inspect schema
Schema schema = factory.inspect();

For some of the data format that is compatible with a user-defined schema, user can use method DatasetFactory#inspect(Schema schema) to create the dataset:

Schema schema = createUserSchema()
Dataset dataset = factory.finish(schema);

Otherwise when the non-parameter method DatasetFactory#inspect() is called, schema will be inferred automatically from data source. The same as the result of DatasetFactory#inspect().

Also, if projector is specified during scanning (see next section Projection (Subset of Columns)), the actual schema of output data can be got within method Scanner::schema():

Scanner scanner = dataset.newScan(
    new ScanOptions(32768, Optional.of(new String[] {"id", "name"})));
Schema projectedSchema = scanner.schema();

Projection (Subset of Columns)#

User can specify projections in ScanOptions. For example:

String[] projection = new String[] {"id", "name"};
ScanOptions options = new ScanOptions(32768, Optional.of(projection));

If no projection is needed, leave the optional projection argument absent in ScanOptions:

ScanOptions options = new ScanOptions(32768, Optional.empty());

Or use shortcut constructor:

ScanOptions options = new ScanOptions(32768);

Then all columns will be emitted during scanning.

Projection (Produce New Columns) and Filters#

User can specify projections (new columns) or filters in ScanOptions using Substrait. For example:

ByteBuffer substraitExpressionFilter = getSubstraitExpressionFilter();
ByteBuffer substraitExpressionProject = getSubstraitExpressionProjection();
// Use Substrait APIs to create an Expression and serialize to a ByteBuffer
ScanOptions options = new ScanOptions.Builder(/*batchSize*/ 32768)
             .columns(Optional.empty())
             .substraitExpressionFilter(substraitExpressionFilter)
             .substraitExpressionProjection(getSubstraitExpressionProjection())
             .build();

See also

Executing Projections and Filters Using Extended Expressions

Projections and Filters using Substrait.

Read Data from HDFS#

FileSystemDataset supports reading data from non-local file systems. HDFS support is included in the official Apache Arrow Java package releases and can be used directly without re-building the source code.

To access HDFS data using Dataset API, pass a general HDFS URI to FilesSystemDatasetFactory:

String uri = "hdfs://{hdfs_host}:{port}/data/example.parquet";
BufferAllocator allocator = new RootAllocator(Long.MAX_VALUE);
DatasetFactory factory = new FileSystemDatasetFactory(allocator,
    NativeMemoryPool.getDefault(), FileFormat.PARQUET, uri);

Native Memory Management#

To gain better performance and reduce code complexity, Java FileSystemDataset internally relies on C++ arrow::dataset::FileSystemDataset via JNI. As a result, all Arrow data read from FileSystemDataset is supposed to be allocated off the JVM heap. To manage this part of memory, an utility class NativeMemoryPool is provided to users.

As a basic example, by using a listenable NativeMemoryPool, user can pass a listener hooking on C++ buffer allocation/deallocation:

AtomicLong reserved = new AtomicLong(0L);
ReservationListener listener = new ReservationListener() {
  @Override
  public void reserve(long size) {
    reserved.getAndAdd(size);
  }

  @Override
  public void unreserve(long size) {
    reserved.getAndAdd(-size);
  }
};
NativeMemoryPool pool = NativeMemoryPool.createListenable(listener);
FileSystemDatasetFactory factory = new FileSystemDatasetFactory(allocator,
    pool, FileFormat.PARQUET, uri);

Also, it’s a very common case to reserve the same amount of JVM direct memory for the data read from datasets. For this use a built-in utility class DirectReservationListener is provided:

NativeMemoryPool pool = NativeMemoryPool.createListenable(
    DirectReservationListener.instance());

This way, once the allocated byte count of Arrow buffers reaches the limit of JVM direct memory, OutOfMemoryError: Direct buffer memory will be thrown during scanning.

Note

The default instance NativeMemoryPool.getDefaultMemoryPool() does nothing on buffer allocation/deallocation. It’s OK to use it in the case of POC or testing, but for production use in complex environment, it’s recommended to manage memory by using a listenable memory pool.

Note

The BufferAllocator instance passed to FileSystemDatasetFactory’s constructor is also aware of the overall memory usage of the produced dataset instances. Once the Java buffers are created the passed allocator will become their parent allocator.

Usage Notes#

Native Object Resource Management#

As another result of relying on JNI, all components related to FileSystemDataset should be closed manually or use try-with-resources to release the corresponding native objects after using. For example:

String uri = "file:/opt/example.parquet";
ScanOptions options = new ScanOptions(/*batchSize*/ 32768);
try (
    BufferAllocator allocator = new RootAllocator();
    DatasetFactory factory = new FileSystemDatasetFactory(
            allocator, NativeMemoryPool.getDefault(),
            FileFormat.PARQUET, uri);
    Dataset dataset = factory.finish();
    Scanner scanner = dataset.newScan(options)
) {

    // do something

} catch (Exception e) {
    e.printStackTrace();
}

If user forgets to close them then native object leakage might be caused.

BatchSize#

The batchSize argument of ScanOptions is a limit on the size of an individual batch.

For example, let’s try to read a Parquet file with gzip compression and 3 row groups:

# Let configure ScanOptions as:
ScanOptions options = new ScanOptions(/*batchSize*/ 32768);

$ parquet-tools meta data4_3rg_gzip.parquet
file schema: schema
age:         OPTIONAL INT64 R:0 D:1
name:        OPTIONAL BINARY L:STRING R:0 D:1
row group 1: RC:4 TS:182 OFFSET:4
row group 2: RC:4 TS:190 OFFSET:420
row group 3: RC:3 TS:179 OFFSET:838

Here, we set the batchSize in ScanOptions to 32768. Because that’s greater than the number of rows in the next batch, which is 4 rows because the first row group has only 4 rows, then the program gets only 4 rows. The scanner will not combine smaller batches to reach the limit, but it will split large batches to stay under the limit. So in the case the row group had more than 32768 rows, it would get split into blocks of 32768 rows or less.

On this page
Edit on GitHub