Note: This describes the new Data Source API, introduced in Flink 1.11 as part of FLIP-27. This new API is currently in BETA status.
Most of the existing source connectors are not yet (as of Flink 1.11) implemented using this new API, but using the previous API, based on SourceFunction.
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.
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 granularity by which the source distributes the work and parallelizes the data reading.
The SourceReader requests Splits and processes them, for example by reading the file or log partition represented by the Split. The SourceReader run in parallel on the Task Managers in the
SourceOperators and produces the parallel stream of events/records.
The SplitEnumerator generates the Splits and assignes 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 API entry point that ties the above three components together.
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 fix 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 keep generating new splits).
Here are some simplified conceptual examples to illustrate how the data source components interact, in streaming and batch cases.
Note that this does not the 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.
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.
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.
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.
The Source API is a factory style interface to create the following components.
In addition to that, the Source provides the boundedness attribute of the source, so that Flink can choose 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.
The SplitEnumerator is expected to be the “brain” of the Source. Typical implementations of the
SplitEnumerator do the following:
addSplitsBack()method will be invoked when a
SourceReaderfails. The SplitEnumerator should take back the split assignments that have not been acknowledged by the failed
SourceEvents are custom events sent between
SourceReader. The implementation can leverage this mechanism to perform sophisticated coordination.
SplitEnumeratorcan assign splits to the
SourceReaders in response to various events, including discovery of new splits, new
SplitEnumerator can accomplish the above work with the help of the SplitEnumeratorContext which is provided to the
Source on creation or restore of the
SplitEnumeratorContext allows a
SplitEnumerator to retrieve necessary information of the readers and perform coordination actions.
Source implementation is expected to pass the
SplitEnumeratorContext to the
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
Such implementations may find that the
SplitEnumeratorContext is handy. The code snippet below shows how the
SplitEnumerator implementation can achieve that without maintaining its own threads.
The SourceReader is a component running in the Task Managers to consume the records from the Splits.
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 intermediately.
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 one block at a time to the
SourceReader implementation should avoid emitting multiple records in a single
pollNext(ReaderOutput) invocation unless necessary.
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.
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. A typical design pattern for the
Source is letting the
SourceReaders report their local information to the
SplitEnumerator who has a global view to make decisions.
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 reduce the amount the work needed to write a
It is highly recommended for the connector developers to take advantage of the
SourceReaderBase instead of writing the
SourceReaders by themselves from scratch. For more details please check the Split Reader API section.
In order to create a
DataStream from a
Source, one needs to pass the
Source to a
StreamExecutionEnvironment. For example,
The core SourceReader API is fully asynchronous and requires implementations to manage asynchronous split reading manually.
However, in practice, most sources use 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 API only has three methods:
SplitReader only focuses on reading the records from the external system, therefore is much simpler compared with
Please check the Java doc of the class for more details.
It is quite common that a
SourceReader implementation does the following:
In order to reduce the work of writing a new
SourceReader, Flink provides a SourceReaderBase class to serve as a base implementation of the
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.
SourceReaderBase supports a few threading models out of the box, depending on the behavior of the SplitFetcherManager it works with.
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
The following code snippet implements the this threading model.
SourceReader using this threading model can be created like following:
SourceReader implementations can also implement their own threading model easily on top of the
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.
Important 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.
WatermarkGenerator run transparently as part of the
SourceOutput) so source implementors do not have to implement any timestamp extraction and watermark generation code.
Event timestamps are assigned in two steps:
The SourceReader may attach the source record timestamp to the event, by calling
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.
TimestampAssigner, which is configured by the application, assigns the final timestamp.
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 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
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.
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 ouput 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
releaseOutputForSplit(splitId) methods. Please refer to the JavaDocs of the class and methods for details.