Data Sources

Data Sources #

This page describes Flink’s Data Source API and the concepts and architecture behind it. Read this, if you are interested in how data sources in Flink work, or if you want to implement a new Data Source.

If you are looking for pre-defined source connectors, please check the Connector Docs.

Data Source Concepts #

Core Components

A Data Source has three core components: Splits, the SplitEnumerator, and the SourceReader.

  • A Split is a portion of data consumed by the source, like a file or a log partition. Splits are the granularity by which the source distributes the work and parallelizes reading data.

  • The SourceReader requests Splits and processes them, for example by reading the file or log partition represented by the Split. The SourceReaders run in parallel on the Task Managers in the SourceOperators and produce the parallel stream of events/records.

  • The SplitEnumerator generates the Splits and assigns them to the SourceReaders. It runs as a single instance on the Job Manager and is responsible for maintaining the backlog of pending Splits and assigning them to the readers in a balanced manner.

The Source class is the API entry point that ties the above three components together.

Illustration of SplitEnumerator and SourceReader interacting

Unified Across Streaming and Batch

The Data Source API supports both unbounded streaming sources and bounded batch sources, in a unified way.

The difference between both cases is minimal: In the bounded/batch case, the enumerator generates a fixed set of splits, and each split is necessarily finite. In the unbounded streaming case, one of the two is not true (splits are not finite, or the enumerator keeps generating new splits).

Examples #

Here are some simplified conceptual examples to illustrate how the data source components interact, in streaming and batch cases.

Note that this does not accurately describe how the Kafka and File source implementations work; parts are simplified, for illustrative purposes.

Bounded File Source

The source has the URI/Path of a directory to read, and a Format that defines how to parse the files.

  • A Split is a file, or a region of a file (if the data format supports splitting the file).
  • The SplitEnumerator lists all files under the given directory path. It assigns Splits to the next reader that requests a Split. Once all Splits are assigned, it responds to requests with NoMoreSplits.
  • The SourceReader requests a Split and reads the assigned Split (file or file region) and parses it using the given Format. If it does not get another Split, but a NoMoreSplits message, it finishes.

Unbounded Streaming File Source

This source works the same way as described above, except that the SplitEnumerator never responds with NoMoreSplits and periodically lists the contents under the given URI/Path to check for new files. Once it finds new files, it generates new Splits for them and can assign them to the available SourceReaders.

Unbounded Streaming Kafka Source

The source has a Kafka Topic (or list of Topics or Topic regex) and a Deserializer to parse the records.

  • A Split is a Kafka Topic Partition.
  • The SplitEnumerator connects to the brokers to list all topic partitions involved in the subscribed topics. The enumerator can optionally repeat this operation to discover newly added topics/partitions.
  • The SourceReader reads the assigned Splits (Topic Partitions) using the KafkaConsumer and deserializes the records using the provided Deserializer. The splits (Topic Partitions) do not have an end, so the reader never reaches the end of the data.

Bounded Kafka Source

Same as above, except that each Split (Topic Partition) has a defined end offset. Once the SourceReader reaches the end offset for a Split, it finishes that Split. Once all assigned Splits are finished, the SourceReader finishes.

The Data Source API #

This section describes the major interfaces of the new Source API introduced in FLIP-27, and provides tips to the developers on the Source development.

Source #

The Source API is a factory style interface to create the following components.

  • Split Enumerator
  • Source Reader
  • Split Serializer
  • Enumerator Checkpoint Serializer

In addition to that, the Source provides the boundedness attribute of the source, so that Flink can choose the appropriate mode to run the Flink jobs.

The Source implementations should be serializable as the Source instances are serialized and uploaded to the Flink cluster at runtime.

SplitEnumerator #

The SplitEnumerator is expected to be the “brain” of the Source. Typical implementations of the SplitEnumerator do the following:

  • SourceReader registration handling
  • SourceReader failure handling
    • The addSplitsBack() method will be invoked when a SourceReader fails. The SplitEnumerator should take back the split assignments that have not been acknowledged by the failed SourceReader.
  • SourceEvent handling
    • SourceEvents are custom events sent between SplitEnumerator and SourceReader. The implementation can leverage this mechanism to perform sophisticated coordination.
  • Split discovery and assignment
    • The SplitEnumerator can assign splits to the SourceReaders in response to various events, including discovery of new splits, new SourceReader registration, SourceReader failure, etc.

A SplitEnumerator can accomplish the above work with the help of the SplitEnumeratorContext which is provided to the Source on creation or restore of the SplitEnumerator. The SplitEnumeratorContext allows a SplitEnumerator to retrieve necessary information of the readers and perform coordination actions. The Source implementation is expected to pass the SplitEnumeratorContext to the SplitEnumerator instance.

While a SplitEnumerator implementation can work well in a reactive way by only taking coordination actions when its method is invoked, some SplitEnumerator implementations might want to take actions actively. For example, a SplitEnumerator may want to periodically run split discovery and assign the new splits to the SourceReaders. Such implementations may find that the callAsync() method in the SplitEnumeratorContext is handy. The code snippet below shows how the SplitEnumerator implementation can achieve that without maintaining its own threads.

class MySplitEnumerator implements SplitEnumerator<MySplit, MyCheckpoint> {
    private final long DISCOVER_INTERVAL = 60_000L;

    private final SplitEnumeratorContext<MySplit> enumContext            ;

    /** The Source creates instances of SplitEnumerator and provides the context. */
    MySplitEnumerator(SplitEnumeratorContext<MySplit> enumContext) {
        this.enumContext = enumContext;
    }

    /**
     * A method to discover the splits.
     */
    private List<MySplit> discoverSplits() {...}
    
    @Override
    public void start() {
        ...
        enumContext.callAsync(this::discoverSplits, (splits, thrown) -> {
            Map<Integer, List<MySplit>> assignments = new HashMap<>();
            int parallelism = enumContext.currentParallelism();
            for (MySplit split : splits) {
                int owner = split.splitId().hashCode() % parallelism;
                assignments.computeIfAbsent(owner, s -> new ArrayList<>()).add(split);
            }
            enumContext.assignSplits(new SplitsAssignment<>(assignments));
        }, 0L, DISCOVER_INTERVAL);
        ...
    }
    ...
}

Still not supported in Python API.

SourceReader #

The SourceReader is a component running in the Task Managers to consume the records from the Splits.

The SourceReader exposes a pull-based consumption interface. A Flink task keeps calling pollNext(ReaderOutput) in a loop to poll records from the SourceReader. The return value of the pollNext(ReaderOutput) method indicates the status of the source reader.

  • MORE_AVAILABLE - The SourceReader has more records available immediately.
  • NOTHING_AVAILABLE - The SourceReader does not have more records available at this point, but may have more records in the future.
  • END_OF_INPUT - The SourceReader has exhausted all the records and reached the end of data. This means the SourceReader can be closed.

In the interest of performance, a ReaderOutput is provided to the pollNext(ReaderOutput) method, so a SourceReader can emit multiple records in a single call of pollNext() if it has to. For example, sometimes the external system works at the granularity of blocks. A block may contain multiple records but the source can only checkpoint at the block boundaries. In this case the SourceReader can emit all the records in one block at a time to the ReaderOutput. However, the SourceReader implementation should avoid emitting multiple records in a single pollNext(ReaderOutput) invocation unless necessary. This is because the task thread that is polling from the SourceReader works in an event-loop and cannot block.

All the state of a SourceReader should be maintained inside the SourceSplits which are returned at the snapshotState() invocation. Doing this allows the SourceSplits to be reassigned to other SourceReaders when needed.

A SourceReaderContext is provided to the Source upon a SourceReader creation. It is expected that the Source will pass the context to the SourceReader instance. The SourceReader can send SourceEvent to its SplitEnumerator through the SourceReaderContext. A typical design pattern of the Source is letting the SourceReaders report their local information to the SplitEnumerator who has a global view to make decisions.

The SourceReader API is a low level API that allows users to deal with the splits manually and have their own threading model to fetch and handover the records. To facilitate the SourceReader implementation, Flink has provided a SourceReaderBase class which significantly reduces the amount the work needed to write a SourceReader. It is highly recommended for the connector developers to take advantage of the SourceReaderBase instead of writing the SourceReaders from scratch. For more details please check the Split Reader API section.

Use the Source #

In order to create a DataStream from a Source, one needs to pass the Source to a StreamExecutionEnvironment. For example,

final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();

Source mySource = new MySource(...);

DataStream<Integer> stream = env.fromSource(
        mySource,
        WatermarkStrategy.noWatermarks(),
        "MySourceName");
...

val env = StreamExecutionEnvironment.getExecutionEnvironment()

val mySource = new MySource(...)

val stream = env.fromSource(
      mySource,
      WatermarkStrategy.noWatermarks(),
      "MySourceName")
...

env = StreamExecutionEnvironment.get_execution_environment()

my_source = ...

env.from_source(
    my_source,
    WatermarkStrategy.no_watermarks(),
    "my_source_name")

The Split Reader API #

The core SourceReader API is fully asynchronous and requires implementations to manually manage reading splits asynchronously. However, in practice, most sources perform blocking operations, like blocking poll() calls on clients (for example the KafkaConsumer), or blocking I/O operations on distributed file systems (HDFS, S3, …). To make this compatible with the asynchronous Source API, these blocking (synchronous) operations need to happen in separate threads, which hand over the data to the asynchronous part of the reader.

The SplitReader is the high-level API for simple synchronous reading/polling-based source implementations, like file reading, Kafka, etc.

The core is the SourceReaderBase class, which takes a SplitReader and creates fetcher threads running the SplitReader, supporting different consumption threading models.

SplitReader #

The SplitReader API only has three methods:

  • A blocking fetch method to return a RecordsWithSplitIds
  • A non-blocking method to handle split changes.
  • A non-blocking wake up method to wake up the blocking fetch operation.

The SplitReader only focuses on reading the records from the external system, therefore is much simpler compared with SourceReader. Please check the Java doc of the class for more details.

SourceReaderBase #

It is quite common that a SourceReader implementation does the following:

  • Have a pool of threads fetching from splits of the external system in a blocking way.
  • Handle the synchronization between the internal fetching threads and other methods invocations such as pollNext(ReaderOutput).
  • Maintain the per split watermark for watermark alignment.
  • Maintain the state of each split for checkpoint.

In order to reduce the work of writing a new SourceReader, Flink provides a SourceReaderBase class to serve as a base implementation of the SourceReader. SourceReaderBase has all the above work done out of the box. To write a new SourceReader, one can just let the SourceReader implementation inherit from the SourceReaderBase, fill in a few methods and implement a high level SplitReader .

SplitFetcherManager #

The SourceReaderBase supports a few threading models out of the box, depending on the behavior of the SplitFetcherManager it works with. The SplitFetcherManager helps create and maintain a pool of SplitFetchers each fetching with a SplitReader. It also determines how to assign splits to each split fetcher.

As an example, as illustrated below, a SplitFetcherManager may have a fixed number of threads, each fetching from some splits assigned to the SourceReader.

One fetcher per split threading model.

The following code snippet implements this threading model.

/**
 * A SplitFetcherManager that has a fixed size of split fetchers and assign splits 
 * to the split fetchers based on the hash code of split IDs.
 */
public class FixedSizeSplitFetcherManager<E, SplitT extends SourceSplit> 
        extends SplitFetcherManager<E, SplitT> {
    private final int numFetchers;

    public FixedSizeSplitFetcherManager(
            int numFetchers,
            FutureCompletingBlockingQueue<RecordsWithSplitIds<E>> elementsQueue,
            Supplier<SplitReader<E, SplitT>> splitReaderSupplier) {
        super(elementsQueue, splitReaderSupplier);
        this.numFetchers = numFetchers;
        // Create numFetchers split fetchers.
        for (int i = 0; i < numFetchers; i++) {
            startFetcher(createSplitFetcher());
        }
    }

    @Override
    public void addSplits(List<SplitT> splitsToAdd) {
        // Group splits by their owner fetchers.
        Map<Integer, List<SplitT>> splitsByFetcherIndex = new HashMap<>();
        splitsToAdd.forEach(split -> {
            int ownerFetcherIndex = split.hashCode() % numFetchers;
            splitsByFetcherIndex
                    .computeIfAbsent(ownerFetcherIndex, s -> new ArrayList<>())
                    .add(split);
        });
        // Assign the splits to their owner fetcher.
        splitsByFetcherIndex.forEach((fetcherIndex, splitsForFetcher) -> {
            fetchers.get(fetcherIndex).addSplits(splitsForFetcher);
        });
    }
}

Still not supported in Python API.

And a SourceReader using this threading model can be created like following:

public class FixedFetcherSizeSourceReader<E, T, SplitT extends SourceSplit, SplitStateT>
        extends SourceReaderBase<E, T, SplitT, SplitStateT> {

    public FixedFetcherSizeSourceReader(
            FutureCompletingBlockingQueue<RecordsWithSplitIds<E>> elementsQueue,
            Supplier<SplitReader<E, SplitT>> splitFetcherSupplier,
            RecordEmitter<E, T, SplitStateT> recordEmitter,
            Configuration config,
            SourceReaderContext context) {
        super(
                elementsQueue,
                new FixedSizeSplitFetcherManager<>(
                        config.getInteger(SourceConfig.NUM_FETCHERS),
                        elementsQueue,
                        splitFetcherSupplier),
                recordEmitter,
                config,
                context);
    }

    @Override
    protected void onSplitFinished(Map<String, SplitStateT> finishedSplitIds) {
        // Do something in the callback for the finished splits.
    }

    @Override
    protected SplitStateT initializedState(SplitT split) {
        ...
    }

    @Override
    protected SplitT toSplitType(String splitId, SplitStateT splitState) {
        ...
    }
}

Still not supported in Python API.

The SourceReader implementations can also implement their own threading model easily on top of the SplitFetcherManager and SourceReaderBase.

Event Time and Watermarks #

Event Time assignment and Watermark Generation happen as part of the data sources. The event streams leaving the Source Readers have event timestamps and (during streaming execution) contain watermarks. See Timely Stream Processing for an introduction to Event Time and Watermarks.

Applications based on the legacy SourceFunction typically generate timestamps and watermarks in a separate later step via stream.assignTimestampsAndWatermarks(WatermarkStrategy). This function should not be used with the new sources, because timestamps will be already assigned, and it will override the previous split-aware watermarks.

API #

The WatermarkStrategy is passed to the Source during creation in the DataStream API and creates both the TimestampAssigner and WatermarkGenerator .

environment.fromSource(
    Source<OUT, ?, ?> source,
    WatermarkStrategy<OUT> timestampsAndWatermarks,
    String sourceName);

environment.from_source(
    source: Source,
    watermark_strategy: WatermarkStrategy,
    source_name: str,
    type_info: TypeInformation = None)

The TimestampAssigner and WatermarkGenerator run transparently as part of the ReaderOutput(or SourceOutput) so source implementors do not have to implement any timestamp extraction and watermark generation code.

Event Timestamps #

Event timestamps are assigned in two steps:

  1. The SourceReader may attach the source record timestamp to the event, by calling SourceOutput.collect(event, timestamp). This is relevant only for data sources that are record-based and have timestamps, such as Kafka, Kinesis, Pulsar, or Pravega. Sources that are not based on records with timestamps (like files) do not have a source record timestamp. This step is part of the source connector implementation and not parameterized by the application that uses the source.

  2. The TimestampAssigner, which is configured by the application, assigns the final timestamp. The TimestampAssigner sees the original source record timestamp and the event. The assigner can use the source record timestamp or access a field of the event to obtain the final event timestamp.

This two-step approach allows users to reference both timestamps from the source systems and timestamps in the event’s data as the event timestamp.

Note: When using a data source without source record timestamps (like files) and selecting the source record timestamp as the final event timestamp, events will get a default timestamp equal to LONG_MIN (=-9,223,372,036,854,775,808).

Watermark Generation #

Watermark Generators are only active during streaming execution. Batch execution deactivates Watermark Generators; all related operations described below become effectively no-ops.

The data source API supports running watermark generators individually per split. That allows Flink to observe the event time progress per split individually, which is important to handle event time skew properly and prevent idle partitions from holding back the event time progress of the entire application.

Watermark Generation in a Source with two Splits.

When implementing a source connector using the Split Reader API, this is automatically handled. All implementations based on the Split Reader API have split-aware watermarks out-of-the-box.

For an implementation of the lower level SourceReader API to use split-aware watermark generation, the implementation must output events from different splits to different outputs: the Split-local SourceOutputs. Split-local outputs can be created and released on the main ReaderOutput via the createOutputForSplit(splitId) and releaseOutputForSplit(splitId) methods. Please refer to the JavaDocs of the class and methods for details.

Split Level Watermark Alignment #

Although source operator watermark alignment is handled by Flink runtime, the source needs to additionally implement SourceReader#pauseOrResumeSplits and SplitReader#pauseOrResumeSplits to achieve split level watermark alignment. Split level watermark alignment is useful for when there are multiple splits assigned to a source reader. By default, these implementations will throw an UnsupportedOperationException, pipeline.watermark-alignment.allow-unaligned-source-splits is set to false, when there is more than one split assigned, and the split exceeds the watermark alignment threshold configured by the WatermarkStrategy. SourceReaderBase contains an implementation for SourceReader#pauseOrResumeSplits so that inheriting sources only need to implement SplitReader#pauseOrResumeSplits. See the javadocs for more implementation hints.