Release Notes: Dataflow SDK 1.x for Java

This page documents production updates to the Dataflow SDK 1.x for Java. You can periodically check this page for announcements about new or updated features, bug fixes, known issues, and deprecated functionality. For information about version 2.x of the Dataflow SDK for Java, see the Dataflow SDK 2.x for Java release notes.

The support page contains information about the support status of each release of the Dataflow SDK.

To install and use the Dataflow SDK, see the Dataflow SDK installation guide.

October 16, 2018

Cloud Dataflow SDK 1.x for Java is unsupported as of October 16, 2018. In the near future, the Cloud Dataflow service will reject new Cloud Dataflow jobs that are based on Cloud Dataflow SDK 1.x for Java. See Migrating from Cloud Dataflow SDK 1.x for Java for migration guidance.

1.9.1 (August 28, 2017)

Fixed an issue with Dataflow jobs that read from CompressedSources with compression type set to BZIP2 are potentially losing data during processing. For more information, see Issue #596 on the GitHub repository.

1.9.0 (December 20, 2016)

Identified issue: Dataflow jobs that read from CompressedSources with compression type set to BZIP2 are potentially losing data during processing. For more information, see Issue #596 on the GitHub repository.

Added support for using the Stackdriver Error Reporting Interface.

Added the ValueProvider interface for use in pipeline options. Making an option of type ValueProvider<T> instead of T allows its value to be supplied at runtime (rather than pipeline construction time) and enables Cloud Dataflow templates. Support for ValueProvider has been added to TextIO, PubSubIO, and BigQueryIO and can be added to arbitrary PTransforms as well. See the documentation on Cloud Dataflow templates for more details.

Added the ability to automatically save profiling information to Google Cloud Storage using the --saveProfilesToGcs pipeline option. For more information on profiling pipelines executed by the DataflowPipelineRunner, see GitHub issue #72.

Deprecated the --enableProfilingAgent pipeline option that saved profiles to the individual worker disks. For more information on profiling pipelines executed by the DataflowPipelineRunner, see GitHub issue #72.

Changed FileBasedSource to throw an exception when reading from a file pattern that has no matches. Pipelines will now fail at runtime rather than silently reading no data in this case. This change affects TextIO.Read or AvroIO.Read when configured withoutValidation.

Enhanced Coder validation in the DirectPipelineRunner to catch coders that cannot properly encode and decode their input.

Improved display data throughout core transforms, including properly handling arrays in PipelineOptions.

Improved performance for pipelines using the DataflowPipelineRunner in streaming mode.

Improved scalability of the InProcessRunner, enabling testing with larger datasets.

Improved the cleanup of temporary files created by TextIO, AvroIO, and other FileBasedSource implementations.

1.8.1 (December 12, 2016)

Improved the performance of bounded side inputs in the DataflowPipelineRunner.

1.8.0 (October 3, 2016)

Added support to BigQueryIO.Read for queries in the new BigQuery Standard SQL dialect using .withStandardSQL().

Added support in BigQueryIO for the new BYTES, TIME, DATE, and DATETIME types.

Added support to BigtableIO.Read for reading from a restricted key range using .withKeyRange(ByteKeyRange).

Improved initial splitting of large uncompressed files in CompressedSource, leading to better performance when executing batch pipelines that use TextIO.Read on the Cloud Dataflow service.

Fixed a performance regression when using BigQueryIO.Write in streaming mode.

1.7.0 (September 9, 2016)

Identified issue: We have identified a performance regression in BigQueryIO.Write. When run in streaming mode, users may see a small increase in failed inserts, though no data will be lost or duplicated. For more information, see Issue #451 in the GitHub repository.

Added support for Cloud Datastore API v1 in the new com.google.cloud.dataflow.sdk.io.datastore.DatastoreIO. Deprecated the old DatastoreIO class that supported only the deprecated Cloud Datastore API v1beta2.

Improved DatastoreIO.Read to support dynamic work rebalancing, and added an option to control the number of query splits using withNumQuerySplits.

Improved DatastoreIO.Write to work with an unbounded PCollection, supporting writing to Cloud Datastore when using the DataflowPipelineRunner in streaming mode.

Added the ability to delete Cloud Datastore Entity objects directly using Datastore.v1().deleteEntity or to delete entities by key using Datastore.v1().deleteKey.

Added support for reading from a BoundedSource to the DataflowPipelineRunner in streaming mode. This enables the use of TextIO.Read, AvroIO.Read and other bounded sources in these pipelines.

Added support for optionally writing a header and/or footer to text files produced with TextIO.Write.

Added the ability to control the number of output shards produced when using a Sink.

Added TestStream to enable testing of triggers with multiple panes and late data with the InProcessPipelineRunner.

Added the ability to control the rate at which UnboundedCountingInput produces elements using withRate(long, Duration).

Improved performance and stability for pipelines using the DataflowPipelineRunner in streaming mode.

To support TestStream, reimplemented DataflowAssert to use GroupByKey instead of sideInputs to check assertions. This is an update-incompatible change to DataflowAssert for pipelines run on the DataflowPipelineRunner in streaming mode.

Fixed an issue in which a FileBasedSink would produce no files when writing an empty PCollection.

Fixed an issue in which BigQueryIO.Read could not query a table in a non-US region when using the DirectPipelineRunner or the InProcessPipelineRunner.

Fixed an issue in which the combination of timestamps near the end of the global window and a large allowedLateness could cause an IllegalStateException for pipelines run in the DirectPipelineRunner.

Fixed a NullPointerException that could be thrown during pipeline submission when using an AfterWatermark trigger with no late firings.

1.6.1 (August 8, 2016)

Fixed an issue with Dataflow jobs reading from TextIO with compression type set to GZIP or BZIP2. For more information, see Issue #356 on the GitHub repository.

1.6.0 (June 10, 2016)

Identified issue: Dataflow jobs reading from TextIO, with compression type set to GZIP or BZIP2, are potentially losing data during processing. Users are advised to employ the workarounds discussed in Issue #356 in the GitHub repository.

Added display data, which allows annotating user functions (DoFn, CombineFn, and WindowFn), Sources, and Sinks with static metadata to be displayed in the Dataflow Monitoring Interface. Display data has been implemented for core components and is automatically applied to all PipelineOptions.

Added the methods getSplitPointsConsumed and getSplitPointsRemaining to the BoundedReader API to improve Dataflow's ability to automatically scale a job reading from these sources. Default implementations of these functions have been provided, but reader implementers should override them to provide better information when available.

Added the ability to compose multiple CombineFns into a single CombineFn using CombineFns.compose or CombineFns.composeKeyed.

Added InProcessPipelineRunner, an improvement over the DirectPipelineRunner that better implements the Dataflow model. InProcessPipelineRunner runs on a user's local machine and supports multithreaded execution, unbounded PCollections, and triggers for speculative and late outputs.

Reimplemented BigQueryIO so that the DirectPipelineRunner and InProcessPipelineRunner implementations execute similarly to the DataflowPipelineRunner. You must now specify the --tempLocation execution parameter when using DirectPipelineRunner or InProcessPipelineRunner.

Improved performance of side inputs when using workers with many cores.

Improved efficiency when using CombineFnWithContext.

Fixed several issues related to stability in the streaming mode.

1.5.1 (April 15, 2016)

Fixed an issue that hid BigtableIO.Read.withRowFilter, which allows Cloud Bigtable rows to be filtered in the Read transform.
Fixed support for concatenated GZip files.
Fixed an issue that prevented Write.to to be used with merging windows.
Fixed an issue that caused excessive triggering with repeated composite triggers.
Fixed an issue with merging windows and triggers that finish before the end of the window.

1.5.0 (March 14, 2016)

With this release, we have begun preparing the Dataflow SDK for Java for an eventual move to Apache Beam (incubating). Specifically, we have refactored a number of internal APIs and removed from the SDK classes used only within the worker, which will now be provided by the Google Cloud Dataflow Service during job execution. This refactoring should not affect any user code.

Additionally, the 1.5.0 release includes the following changes:

Enabled an indexed side input format for batch pipelines executed on the Google Cloud Dataflow service. Indexed side inputs significantly increase performance for View.asList, View.asMap, View.asMultimap, and any non-globally-windowed PCollectionViews.
Upgraded to Protocol Buffers version 3.0.0-beta-1. If you use custom Protocol Buffers, you should recompile them with the corresponding version of the protoc compiler. You can continue using both version 2 and 3 of the Protocol Buffers syntax, and no user pipeline code needs to change.
Added ProtoCoder, which is a Coder for Protocol Buffers messages that supports both version 2 and 3 of the Protocol Buffers syntax. This coder can detect when messages can be encoded deterministically. Proto2Coder is now deprecated; we recommend that all users switch to ProtoCoder.
Added withoutResultFlattening to BigQueryIO.Read to disable flattening query results when reading from BigQuery.
Added BigtableIO, enabling support for reading from and writing to Google Cloud Bigtable.
Improved CompressedSource to detect compression format according to the file extension. Added support for reading .gz files that are transparently decompressed by the underlying transport logic.

Apache Beam^™ is a trademark of The Apache Software Foundation or its affiliates in the United States and/or other countries.

1.4.0 (January 22, 2016)

Added a series of batch and streaming example pipelines in a mobile gaming domain that illustrate some advanced topics, including windowing and triggers.
Added support for Combine functions to access pipeline options and side inputs through a context. See GlobalCombineFn and PerKeyCombineFn for further details.
Modified ParDo.withSideInputs() such that successive calls are cumulative.
Modified automatic coder detection of Protocol Buffer messages; such classes now have their coders provided automatically.
Added support for limiting the number of results returned by DatastoreIO.Source. However, when this limit is set, the operation that reads from Cloud Datastore is performed by a single worker rather than executing in parallel across the worker pool.
Modified definition of PaneInfo.{EARLY, ON_TIME, LATE} so that panes with only late data are always LATE, and an ON_TIME pane can never cause a later computation to yield a LATE pane.
Modified GroupByKey to drop late data when that late data arrives for a window that has expired. An expired window means the end of the window is passed by more than the allowed lateness.
When using GlobalWindows, you are no longer required to specify withAllowedLateness(), since no data is ever dropped.
Added support for obtaining the default project ID from the default project configuration produced by newer versions of the gcloud utility. If the default project configuration does not exist, Dataflow reverts to using the old project configuration generated by older versions of the gcloud utility.

1.3.0 (December 4, 2015)

Improved IterableLikeCoder to efficiently encode small values. This change is backward compatible; however, if you have a running pipeline that was constructed with SDK version 1.3.0 or later, it may not be possible to "update" that pipeline with a replacement that was constructed using SDK version 1.2.1 or older. Updating a running pipeline with a pipeline constructed using a new SDK version, however, should be successful.
When TextIO.Write or AvroIO.Write outputs to a fixed number of files, added a reshard (shuffle) step immediately prior to the write step. The cost of this reshard is often exceeded by additional parallelism available to the preceding stage.
Added support for RFC 3339 timestamps in PubsubIO. This allows reading from Cloud Pub/Sub topics published by Cloud Logging without losing timestamp information.
Improved memory management to help prevent pipelines in the streaming execution mode from stalling when running with high memory utilization. This particularly benefits pipelines with large GroupByKey results.
Added ability to customize timestamps of emitted windows. Previously, the watermark was held to the earliest timestamp of any buffered input. With this change, you can choose a later time to allow the watermark to progress further. For example, using the end of the window will prevent long-lived sessions from holding up the output. See Window.Bound.withOutputTime().
Added a simplified syntax for early and late firings with an AfterWatermark trigger, as follows: AfterWatermark.pastEndOfWindow().withEarlyFirings(...).withLateFirings(...).

1.2.1 (October 21, 2015)

Fixed a regression in BigQueryIO that unnecessarily printed a lot of messages when executed using DirectPipelineRunner.

1.2.0 (October 5, 2015)

Added Java 8 support. Added new MapElements and FlatMapElements transforms that accept Java 8 lambdas, for those cases when the full power of ParDo is not required. Filter and Partition accept lambdas as well. Java 8 functionality is demonstrated in a new MinimalWordCountJava8 example.
Enabled @DefaultCoder annotations for generic types. Previously, a @DefaultCoder annotation on a generic type was ignored, resulting in diminished functionality and confusing error messages. It now works as expected.
DatastoreIO now supports (parallel) reads within namespaces. Entities can be written to namespaces by setting the namespace in the Entity key.
Limited the slf4j-jdk14 dependency to the test scope. When a Dataflow job is executing, the slf4j-api, slf4j-jdk14, jcl-over-slf4j, log4j-over-slf4j, and log4j-to-slf4j dependencies will be provided by the system.

1.1.0 (September 15, 2015)

Added a coder for type Set<T> to the coder registry, when type T has its own registered coder.
Added NullableCoder, which can be used in conjunction with other coders to encode a PCollection whose elements may possibly contain null values.
Added Filter as a composite PTransform. Deprecated static methods in the old Filter implementation that return ParDo transforms.
Added SourceTestUtils, which is a set of helper functions and test harnesses for testing correctness of Source implementations.

1.0.0 (August 10, 2015)

The initial General Availability (GA) version, open to all developers, and considered stable and fully qualified for production use. It coincides with the General Availability of the Dataflow Service.
Removed the default values for numWorkers, maxNumWorkers, and similar settings. If these are unspecified, the Dataflow Service will pick an appropriate value.
Added checks to DirectPipelineRunner to help ensure that DoFns obey the existing requirement that inputs and outputs must not be modified.
Added support in AvroCoder for @Nullable fields with deterministic encoding.
Added a requirement that anonymous CustomCoder subclasses override getEncodingId method.
Changed Source.Reader, BoundedSource.BoundedReader, UnboundedSource.UnboundedReader to be abstract classes, instead of interfaces. AbstractBoundedReader has been merged into BoundedSource.BoundedReader.
Renamed ByteOffsetBasedSource and ByteOffsetBasedReader to OffsetBasedSource and OffsetBasedReader, introducing getBytesPerOffset as a translation layer.
Changed OffsetBasedReader, such that the subclass now has to override startImpl and advanceImpl, rather than start and advance. The protected variable rangeTracker is now hidden and updated by base class automatically. To indicate split points, use the method isAtSplitPoint.
Removed methods for adjusting watermark triggers.
Removed an unecessary generic parameter from TimeTrigger.
Removed generation of empty panes unless explicitly requested.

0.4.150727 (July 27, 2015)

Removed the requirement to explicitly set --project if Google Cloud SDK has the default project configuration set.
Added support for creating BigQuery sources from a query.
Added support for custom unbounded sources in the DirectPipelineRunner and DataflowPipelineRunner. See UnboundedSource for details.
Removed unnecessary ExecutionContext argument in BoundedSource.createReader and related methods.
Changed BoundedReader.splitAtFraction to require thread-safety (i.e. safe to call asynchronously with advance or start). Added RangeTracker to help implement thread-safe readers. Users are heavily encouraged to use the class rather than implementing an ad-hoc solution.
Modified Combine transforms by lifting them into (and above) the GroupByKey resulting in better performance.
Modified triggers such that after a GroupByKey, the system will switch to a "Continuation Trigger", which attempts to preserve the original intention regarding handling of speculative and late triggerings instead of returning to the default trigger.
Added WindowFn.getOutputTimestamp and changed GroupByKey behavior to allow incomplete overlapping windows to not hold up progress of earlier, completed windows.
Changed triggering behavior so that empty panes are produced if they are the first pane after the watermark (ON_TIME) or the final pane.
Removed the Window.Trigger intermediate builder class.
Added validation that allowed lateness is specified on the Window PTransform when a trigger is specified.
Re-enabled verification of GroupByKey usage. Specifically, the key must have a deterministic coder and using GroupByKey with an unbounded PCollection requires windowing or triggers.
Changed PTransform names so that they may no longer contain the '=' or ';' characters.

0.4.150710 (July 10, 2015)

Added support for per-window tables to BigQueryIO.
Added support for a custom source implementation for Avro. See AvroSource for more details.
Removed 250GiB Google Cloud Storage file size upload restriction.
Fixed BigQueryIO.Write table creation bug in streaming mode.
Changed Source.createReader() and BoundedSource.createReader() to be abstract.
Moved Source.splitIntoBundles() to BoundedSource.splitIntoBundles()
Added support for reading bounded views of a Pub/Sub stream in PubsubIO for non-streaming DataflowPipelines and DirectPipelines.
Added support for getting a Coder using a Class to the CoderRegistry.
Changed CoderRegistry.registerCoder(Class<T>, Coder<T>) to enforce that the provided coder actually encodes values of the given class, and its use with rawtypes of generic classes is forbidden as it will rarely work correctly.
Migrate to Create.withCoder() and CreateTimestamped.withCoder() instead of calling setCoder() on the outcoming PCollection when the Create PTransform is being applied.
Added three successively more detailed WordCount examples.
Removed PTransform.getDefaultName() which was redundant with PTransform.getKindString().
Added support a unique name check for PTransform's during job creation.
Removed PTransform.withName() and PTransform.setName() The name of a transform is now immutable after construction. Library transforms (like Combine) can provide builder-like methods to change the name. Names can always be overridden at the location where the transform is applied using apply("name", transform).
Added the ability to select the network for worker VMs using DataflowPipelineWorkerPoolOptions.setNetwork(String)

0.4.150602 (June 2, 2015)

Added a dependency on the gcloud core component version 2015.02.05 or newer. Update to the latest version of gcloud by running gcloud components update. See Application Default Credentials for more details on how credentials can be specified.
Removed previously deprecated Flatten.create(). Use Flatten.pCollections() instead.
Removed previously deprecated Coder.isDeterministic(). Implement Coder.verifyDeterministic() instead.
Replaced DoFn.Context#createAggregator with DoFn#createAggregator.
Added support for querying the current value of an Aggregator. See PipelineResult for more information.
Added experimental DoFnWithContext to simplify accessing additional information from a DoFn.
Removed experimental RequiresKeyedState.
Added CannotProvideCoderException to indicate inability to infer a coder, instead of returning null in such cases.
Added CoderProperties for assembling test suites for user-defined coders.
Replaced a constructor of PDone with a static factory PDone.in(Pipeline).
Updated string formatting of the TIMESTAMP values returned by the BigQuery source, when using DirectPipelineRunner or when BigQuery data is used as a side input, which aligns it with the case when BigQuery data is used as a main input.
Added a requirement that the value returned by Source.Reader.getCurrent() must be immutable and remain valid indefinitely.
Replaced some usage of Source with BoundedSource. For example, Read.from() transform can now only be applied to BoundedSource objects.
Moved experimental late-data handling, i.e., the data that arrives to the streaming pipeline after the watermark has passed it, from PubSubIO to Window. Late data will default to being dropped at the first GroupByKey following a Read operation. To allow late data through use Window.Bound#withAllowedLateness.
Added experimental support for accumulating elements within a window across panes.

0.4.150414 (April 14, 2015)

Initial Beta release of the Dataflow SDK for Java.
Improved execution performance in many areas of the system.
Added support for progress estimation and dynamic work rebalancing for user-defined sources.
Added support for user-defined sources to provide the timestamp of the values read via Reader.getCurrentTimestamp().
Added support for user-defined sinks.
Added support for custom types in PubsubIO.
Added support for reading and writing XML files. See XmlSource and XmlSink.
Renamed DatastoreIO.Write.to to DatastoreIO.writeTo. In addition, entities written to Cloud Datastore must have complete keys.
Renamed ReadSource transform into Read.
Replaced Source.createBasicReader with Source.createReader.
Added support for triggers, which allows getting early or partial results for a window, and specifying when to process late data. See Window.into.triggering.
Reduced visibility of PTransform's getInput(), getOutput(), getPipeline(), and getCoderRegistry(). These methods will soon be deleted.
Renamed DoFn.ProcessContext#windows to DoFn.ProcessContext#window. In order for a DoFn to call DoFn.ProcessContext#window, it must implement RequiresWindowAccess.
Added DoFn.ProcessContext#windowingInternals to enable windowing on third-party runners.
Added support for side inputs when running streaming pipelines on the [Blocking]DataflowPipelineRunner.
Changed [Keyed]CombineFn.addInput() to return the new accumulator value. Renamed Combine.perElement().withHotKeys() to Combine.perElement().withHotKeyFanout().
Renamed First.of to Sample.any and RateLimiting to IntraBundleParallelization to better represent its functionality.

0.3.150326 (March 26, 2015)

Added support for accessing PipelineOptions in the Dataflow worker.
Removed one of the type parameters in PCollectionView, which may require simple changes to user's code that uses PCollectionView.
Changed side input API to apply per window. Calls to sideInput() now return values only in the specific window corresponding to the window of the main input element, and not the whole side input PCollectionView. Consequently, sideInput() can no longer be called from startBundle and finishBundle of a DoFn.
Added support for viewing a PCollection as a Map when used as a side input. See View.asMap().
Renamed custom source API to use term "bundle" instead of "shard" in all names. Additionally, term "fork" is replaced with "dynamic split".
Custom source Reader now requires implementing new method start(). Existing code can be fixed by simply adding this method that just calls advance() and returns its value. Additionally, code that uses the Reader should be updated to use both start() and advance(), instead of advance() only.

0.3.150227 (February 27, 2015)

Initial Alpha version of the Dataflow SDK for Java with support for streaming pipelines.
Added determinism checker in AvroCoder to make it easier to interoperate with GroupByKey.
Added support for accessing PipelineOptions in the worker.
Added support for compressed sources.

0.3.150211 (February 11, 2015)

Removed the dependency on the gcloud core component version 2015.02.05 or newer.

0.3.150210 (February 11, 2015)

Caution: depends on the gcloud core component version 2015.02.05 or newer.

Included streaming pipeline runner, which, for now, requires additional whitelisting.
Renamed several windowing-related APIs in a non-backward-compatible way.
Added support for custom sources, which you can use to read from your own input formats.
Introduced worker parallelism: one task per processor.

0.3.150109 (January 10, 2015)

Fixed several platform-specific issues for Microsoft Windows.
Fixed several Java 8-specific issues.
Added a few new examples.

0.3.141216 (December 16, 2014)

Initial Alpha version of the Dataflow SDK for Java.