Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

[This FLIP proposal is a joint work between Xuannan Su   and Dong Lin  ]

Motivation

Assuming the user needs to perform a temporal join, where the Probe Side comes from Kafka and the Build Side comes from a MySQL CDC Source (with a snapshot reading phase and a binlog reading phase), and all input records lack event-time information. The user requires each record on the Probe Side to be joined with at least the records from the Build Side's snapshot phase. In other words, the Join operator needs to wait for the Build Side's snapshot phase to finish reading its records before processing the Probe Side's data.

Currently, Flink cannot support the use case described above. Flink SQL does not support the usage of "SYSTEM_TIME AS OF" syntax used in temporal join with the latest version of any view/table. This is because the TemporalProcessTimeJoinOperator supports temporal joins based on processing time but does not support the Probe Side waiting for records from the Build Side. The operator may start processing the Probe Side's data before reading the data from the Build Side's snapshot phase, which can lead to situations where the Probe Side's data cannot be joined with any data, resulting in output that does not meet the user's requirements. For more details, you can refer to FLINK-19830.

This document proposes APIs that allow source operators (e.g., HybridSource, MySQL CDC Source) to send watermark to downstream operator indicating that watermark should start to increase according to the system time. In addition to supporting the processing-time temporal join, this change also allows us to simplify DataStream APIs such as KeyedStream#window(...) so that users would not need to explicitly differentiate between TumblingEventTimeWindows and TumblingProcessingTimeWindows.

Terminology and Background

The FLIP will make changes to Flink's watermark and timestamp concepts. To help understand the intuition behind the design, we recap the relevant concepts in this section.

Probe Side: The left side of the stream in a temporal join, sometimes referred to as the Fact Table. Usually, the data on the Probe Side doesn't need to be retained after processing.

Build Side: The right side of the stream in a temporal join, which can be a versioned table. It is also known as the Dimension Table. Typically, the Build Side has the latest data for each key.

Watermark: Inform operators that no elements with a timestamp older than or equal to the watermark timestamp should arrive at the operator.

TimestampAssigner: Assigns event time timestamps to elements. These timestamps are used by all functions that operate on event time, such as event time windows. Here is the `TimestampAssigner` interface:

public interface TimestampAssigner<T> {
    long extractTimestamp(T element, long recordTimestamp);
}

WatermarkGenerator: Generates watermarks either based on events or periodically. Here is the WatermarkGenerator interface:

@Public
public interface WatermarkGenerator<T> {
    void onEvent(T event, long eventTimestamp, WatermarkOutput output);
    void onPeriodicEmit(WatermarkOutput output);
}

Here is how a DataStream program determines the value of event-time and watermark:

• When creating a source, user provides a `WatermarkStrategy` to `StreamExecutionEnvironment#fromSource`.
• If the the source supports event time natively (e.g., KafkaSource) or the user provides a custom `TimestampAssigner` in the `WatermarkStrategy` to extract the timestamp from the record, Flink will add the timestamp to the record. Otherwise, the timestamp on the record will be `Long.MIN_VALUE`.
• If the user uses `NoWatermarkGenerator` in the `WatermarkStrategy`, the job will not generate watermarks. Otherwise, the job will periodically emit watermarks, and the watermark value depends on event time. The output frequency of watermarks is determined by `pipeline.auto-watermark-interval`, with a default value of 200ms.

Public Interfaces

1) Add an `useProcessingTime` field to `org.apache.flink.api.common.eventtime.Watermark` and `org.apache.flink.streaming.api.watermark.Watermark`.

/**
 * Watermarks are the progress indicators in the data streams. A watermark signifies that no events
 * with a timestamp smaller than or equal to the watermark's time will occur after the watermark.
 *
 * <ul>
 *   <li>A watermark with a timestamp <i>T</i> and useProcessingTime set to false indicates that the
 *       event time of the stream has progressed to time <i>T</i>.
 *   <li>A watermark with a timestamp <i>T</i> and useProcessingTime set to true indicates that the
 *       event time of the stream progresses in synchronization with the system time. The timestamp
 *       <i>T</> must be less than the current system time. Otherwise, an exception will be thrown.
 * </ul>
 *
 * <p>Watermarks are created at the sources and propagate through the streams and operators.
 *
 * <p>In some cases a watermark is only a heuristic, meaning some events with a lower timestamp may
 * still follow. In that case, it is up to the logic of the operators to decide what to do with the
 * "late events". Operators can for example ignore these late events, route them to a different
 * stream, or send update to their previously emitted results.
 *
 * <p>When a source reaches the end of the input, it emits a final watermark with timestamp {@code
 * Long.MAX_VALUE}, indicating the "end of time".
 *
 * <p>Note: A stream's time starts with a watermark of {@code Long.MIN_VALUE}. That means that all
 * records in the stream with a timestamp of {@code Long.MIN_VALUE} are immediately late.
 *
 * <p>Note: After sending a watermark with useProcessingTime set to true, the source should only send
 * subsequent watermarks with useProcessingTime set to true. Sending a watermark with
 * useProcessingTime set to false will result in an exception.
 */
@Public
public final class Watermark implements Serializable {
  	...

    /**
     * If useProcessingTime set to false, this is the time of the watermark in milliseconds. If
     * useProcessingTime set to true, this is the last effective time of the watermark in
     * milliseconds.
     */
    private final long timestamp;

    /**
     * If this is true, this watermark indicates the event time of the stream progresses in
     * synchronization with the system time.
     */
    private final boolean useProcessingTime;

    public Watermark(long timestamp) {
      this(timestamp, false);
    }

    public Watermark(long timestamp, boolean useProcessingTime) {
        this.timestamp = timestamp;
        this.useProcessingTime = useProcessingTime;
    }

    /** Returns whether the time of watermark can be determined by the system time. */
    public boolean useProcessingTime() {
        return useProcessingTime;
    }
}

Please note that the proposed change extends the Watermark class' expressibility without changing its core abstraction. When `useProcessingTime` is false, then the watermark is a static value (i.e. Watermark#timestamp). When `useProcessingTime` is true, then the watermark should be derived dynamically using a function (i.e. System#currentTimeMillis). In both case, the Watermark class is used to tell downstream operators how to determine the watermark value.

2) Update `AbstractStreamOperator/AbstractStreamOperatorV2` to handle Watermark#useProcessingTime properly:

Original Behavior:

• After receiving an instance of Watermark, the operator triggers those event-time timers whose scheduled time <= watermark#getTimestamp

Updated Behavior：

• After receiving a watermark:
  • If watermark#useProcessingTime  == false, the operator triggers those event-time timers whose scheduled time <= watermark#getTimestamp
  • Otherwise, the operator starts a scheduler to dynamically trigger those event-time timers whose scheduled time <= System#currentTimeMillis

3) Update NoWatermarksGenerator#onPeriodicEmit to emit Watermark(timestamp=Long.MIN_VALUE, useProcessingTime=true).

Note that NoWatermarksGenerator is currently only used when every operator in the job uses "processing time mode" , and these operators's behavior will not rely on the watermark value. Therefore the proposed change will not break existing jobs that use NoWatermarksGenerator.

/**
 * An implementation of a {@link WatermarkGenerator} that only
 * generates Watermark(timestamp=Long.MIN_VALUE, useProcessingTime=true). 
 */
@Public
public final class NoWatermarksGenerator<E> implements WatermarkGenerator<E> {
    ...
}

4) Add delegateWatermarkPeriodicEmit() to the SourceReader interface.

@Public
public interface SourceReader<T, SplitT extends SourceSplit> extends AutoCloseable, CheckpointListener {   
    ...

    /**
     * Provide SourceReader with a runnable that can be used to emit watermark. SourceReader
     * should override this method and return true if it wants to own the responsibility
     * of invoking WatermarkGenerator#onPeriodicEmit.
     *
     * @return true iff the caller should avoid triggering WatermarkGenerator#onPeriodicEmit.
     */
    default boolean delegateWatermarkPeriodicEmit(Runnable onWatermarkEmit) {
        return false;
    }
}

Proposed Changes

1) Update SourceOperator behavior

• SourceOperator should invoke SourceReader#delegateWatermarkPeriodicEmit with a runnable which invokes WatermarkGenerator#onPeriodicEmit.
• If delegateWatermarkPeriodicEmit() returns true, SourceOperator should not start any scheduler that invokes WatermarksGenerator#onPeriodicEmit.

• Otherwise, if emitProgressiveWatermarks == false, SourceOperator should invoke WatermarksGenerator#onPeriodicEmit once before it emits the first record.

• Otherwise, SourceOperator will keep its existing behavior (i.e. start a scheduler to invoke WatermarksGenerator#onPeriodicEmit periodically).

2) Update sources with bounded/unbounded phases (e.g. HybridSource, MySQL CDC Source) to override SourceReader#delegateWatermarkPeriodicEmit.

These sources should invoke onWatermarkEmit when they want to notify downstream operators of the watermark value (e.g. at the beginning of MySQL CDC binlog phase).

3) Update TemporalRowTimeJoinOperator to optimize the case where watermarks from both inputs have useProcessingTime == true. 

After the modifications to the `AbstractStreamOperator` based on the FLIP, the `TemporalRowTimeJoinOperator`, as a subclass of `AbstractStreamOperator`, will be able to support temporal joins based on processing time when both build side and probe side send watermark with `useProcessingTime` set to true, e.g. `MySqlSource` and `KafkaSource`.

In order to optimize the performance, when both the probe side and build side receive a watermark with `useProcessingTime` set to true, the operator can process the data in pure processing time mode without timer. The probe side data can directly join with the build side, and the build side only needs to keep the latest record without triggering clean-up based on event time.

TODO: check whether it offers the best performance.

4) Update Flink SQL Planner to support processing-time temporal join and remove TemporalProcessTimeJoinOperator

We will update the Flink SQL planner to select the appropriate temporal join operator based on the time attribute and the `WatermarkCharacteristic` of the source.

Original behavior:

• If the temporal join is based on row time (event time), use `TemporalRowTimeJoinOperator`.

• If the temporal join is based on processing time

  • If use the `TemporalFunctionJoin` syntax, use `TemporalProcessTimeJoinOperator`.

  • If not use the `TemporalFunctionJoin` syntax, an exception is thrown.

Updated behavior：

• Use TemporalRowTimeJoinOperator for temporal join.

TODO: check whether it is correct when TemporalJoinFunction syntax is used.

5) Update WatermarkToDataOutput to handle Watermark#useProcessingTime.

`WatermarkToDataOutput` is responsible for sending the watermark from the `SourceReader` downstream and ensuring that the watermark never decreases. The proposed change will keep this semantics.

Original Behavior:

• When the `WatermarkToDataOutput` receive a watermark, it checks the timestamp of the watermark. It only send the watermark downstream if the timestamp of the watermark is greater than the timestamp of the most recently sent watermark.

Updated Behavior：

• When the `WatermarkToDataOutput` receive a watermark, it checks the timestamp and `useProcessingTime` field of the watermark. It only send the watermark downstream if the timestamp of the watermark is greater than the timestamp of the most recently sent watermark or if the `useProcessingTime` field is set to true.

• If the watermark to be sent has `useProcessingTime` set to true, and the current system time is lesser than the timestamp of the most recently sent watermark, an exception is thrown.

• If a watermark with `useProcessingTime` set to true has been previously sent, and the watermark to be sent has `useProcessingTime` set to false, an exception is thrown.

The updated `WatermarkToDataOutput` ensures that when `useProcessingTime` is false, the watermark#timestamp never decreases. It also ensures that information with `useProcessingTime` set to true is sent downstream and that the `useProcessingTime` will not set back to false once it is set to true.

6) Update StatusWatermarkValve to handle Watermark#useProcessingTime.

The `StatusWatermarkValve` is used to calculate the current watermark for an input that has multiple input channels and invoke the `processWatermark` method of the operator. `StatusWatermarkValve` ensures that the watermark passed to the `processWatermark` method is the minimum among all the input channels of that input, and it should never decrease. The proposed change will keep this semantics.

Original behavior:

• Each input channel can be either active or idle. `StatusWatermarkValve` updates the status of each input channel when it receive `WatermarkStatus` from the upstream.

• Each input channel can be either aligned or unaligned. An input channel is consider aligned iff it is active and its watermark is greater than or equals to the last watermark timestamp of the input.

• It gets the minimum watermark timestamp among all the aligned input channels and uses it as the watermark for the input. If the new watermark timestamp is greater than the previous watermark timestamp, it invokes the `processWatermark` method.

Updated Behavior：

• Each input channel can be either active or idle. `StatusWatermarkValve` updates the status of each input channel when it receive `WatermarkStatus` from the upstream. If the input channel has `useProcessingTime` set to true, then it is considered active.

• Each input channel can be either aligned or unaligned. If the `useProcessingTime` of the last watermark of the input is set to false, an input channel is consider aligned iff it is active and its watermark is greater than or equals to the last watermark timestamp of the input. If the `useProcessingTime` of the last watermark of the input is set to true, an input channel is consider aligned iff it is active and its `useProcessingTime` is set to true.

• If there exists any input channel with useProcessingTime==false, the watermark of the input is the minimum watermark timestamp among all the aligned input channels whose useProcessingTime == false. Otherwise, the watermark of the input is Watermark(timestamp=Long.MIN_VALUE, useProcessingTime=true).

After the update, we can still ensure that the effective watermark never decrease. The chart below shows the effective watermark of the input given the watermark of the two input channels.

Note that if any operator emits Watermark(timestamp=t, useProcessingTime=true), then it is required that t <= System#currentTimeMillis. This effectively makes Watermark(timestamp=t, useProcessingTime=true) equivalent to Watermark(timestamp=Long.MIN_VALUE, useProcessingTime=true)

7) Update IndexedCombinedWatermarkStatus to handle Watermark#useProcessingTime.

The `IndexedCombinedWatermarkStatus` represents combined value and status of a watermark for a set number of input partial watermarks. Operator advances the event time of `timeServiceManager` base on the combined watermark. `IndexedCombinedWatermarkStatus` ensures that the watermark timestamp of the operator that has multiple inputs is the minimum timestamp among all the inputs and it should never decrease. The proposed change will keep this semantics.

Original behavior:

• Each channel can be either active or idle. It updates the status of the input when it receive `WatermarkStatus` from that input.

• It gets the minimum watermark timestamp among all the active input and uses it as the watermark of the operator. If the new watermark timestamp is greater than the previous watermark timestamp, it advanced the event time of `timeServiceManager`.

Updated Behavior：

• Each channel can be either active or idle. It updates the status of the input when it receive `WatermarkStatus` from that input. If `useProcessingTime` of an input is set to true, it is always active.

• If there exists any input with useProcessingTime==false, the watermark timestamp of the operator is the minimum watermark timestamp among all the active inputs. Otherwise, the watermark of the operator is Watermark(timestamp=Long.MIN_VALUE, useProcessingTime=true).

Example Usage

Here is the Flink SQL example that demonstrates how to perform processing time temporal join after the FLIP.

-- Create mysql cdc source table (dimension table)
CREATE TEMPORARY TABLE user_info (
    user_id INTEGER PRIMARY KEY NOT ENFORCED, 
    gender STRING
) WITH (
    'connector' = 'mysql-cdc',
    'database-name' = 'example_database',
    'hostname' = 'localhost',
    'username' = 'root',
    'password' = 'root',
    'table-name' = 'user_info'
);

-- Create datagen source table (fact table)
CREATE TEMPORARY TABLE click_event (
    user_id INTEGER,
    item_id INTEGER,
    proctime AS PROCTIME()
) WITH (
    'connector' = 'datagen',
    'rows-per-second' = '1',
    'fields.user_id.min' = '0',
    'fields.user_id.max' = '9'
);

-- Create a print sink table
CREATE TEMPORARY TABLE print_sink (
    user_id INTEGER,
    item_id INTEGER,
    gender STRING
) WITH (
    'connector' = 'print'
);

-- Processing time temporal join
INSERT INTO print_sink
SELECT click_event.user_id AS user_id, item_id, gender FROM click_event 
LEFT JOIN user_info FOR SYSTEM_TIME AS OF click_event.proctime
ON click_event.user_id = user_info.user_id;

Compatibility, Deprecation, and Migration Plan

The change might be backward incompatible in the following case:

• User writes a Flink SQL program with temporal join in processing time, which do not have TemporalFunctionJoin syntax.
• The source on the Build Side has bounded/unbounded phase (e.g. HybridSource, MySQL CDC) and has not been updated to overwrite SourceReader#delegateWatermarkPeriodicEmit as specified in this FLIP.
• User prefers to have the job fail fast rather than producing results where the probe-side records fail to join with the records from the bounded phase of the build-side source.

We believe the impact of the above scenario is out-weighted by the benefits of this FLIP. In the common case, users can notice the data quality issue and be able to upgrade the source library version to resolve the issue. We should mention this change in the Flink release notice.

Other than the issue mentioned above, the changes made in this FLIP is backward compatible, for the following reasons:

• We only add new APIs, which do not cause existing code handling Watermarks to fail. Additionally, the default setting for the `useProcessingTime` parameter in Watermark instances is false, preserving the existing semantics.

• With the updates to `AbstractStreamOperator/AbstractStreamOperatorV2` based on the FLIP, all operators can support Watermarks with `useProcessingTime`. And correctly trigger the operator's timer based on event time or the system time. For sources that haven't been updated, the Watermarks they send always have `useProcessingTime` set to false. In this case, the behavior of the operators remains unchanged, ensuring compatibility with existing jobs.

Test Plan

The change will be covered with unit and integration tests.

Rejected Alternatives

Not yet.

Future Work

After this FLIP, we can unify the API for processing time and event time. The following are some examples of the APIs pairs that distinguish between event time and processing time. Currently, in the DataStream API, users need to explicitly differentiate between Processing Time and Event Time in several places when expressing job logic.

When invoking methods like `DataStream.windowAll` or `KeyedDataStream.window`, users need to select the appropriate `WindowAssigner` based on processing time or event time.

• org.apache.flink.streaming.api.windowing.assigners.WindowAssigner

  • TumblingEventTimeWindows vs TumblingProcessingTimeWindows

  • SlidingEventTimeWindows vs SlidingProcessingTimeWindows

  • DynamicEventTimeSessionWindows vs DynamicProcessingTimeSessionWindows

  • EventTimeSessionWindows vs ProcessingTimeSessionWindows

When implementing custom `ProcessFunction` or `KeyedProcessFunction`, users need to differentiate between registering a timer for processing time or event time using the `TimerService`.

• org.apache.flink.streaming.api.TimerService

  • registerProcessingTimeTimer vs registerEventTimeTimer