聊聊flink的CheckpointedFunction

序

本文主要研究一下flink的CheckpointedFunction

实例

public class BufferingSink
        implements SinkFunction<Tuple2<String, Integer>>,
                   CheckpointedFunction {

    private final int threshold;

    private transient ListState<Tuple2<String, Integer>> checkpointedState;

    private List<Tuple2<String, Integer>> bufferedElements;

    public BufferingSink(int threshold) {
        this.threshold = threshold;
        this.bufferedElements = new ArrayList<>();
    }

    @Override
    public void invoke(Tuple2<String, Integer> value) throws Exception {
        bufferedElements.add(value);
        if (bufferedElements.size() == threshold) {
            for (Tuple2<String, Integer> element: bufferedElements) {
                // send it to the sink
            }
            bufferedElements.clear();
        }
    }

    @Override
    public void snapshotState(FunctionSnapshotContext context) throws Exception {
        checkpointedState.clear();
        for (Tuple2<String, Integer> element : bufferedElements) {
            checkpointedState.add(element);
        }
    }

    @Override
    public void initializeState(FunctionInitializationContext context) throws Exception {
        ListStateDescriptor<Tuple2<String, Integer>> descriptor =
            new ListStateDescriptor<>(
                "buffered-elements",
                TypeInformation.of(new TypeHint<Tuple2<String, Integer>>() {}));

        checkpointedState = context.getOperatorStateStore().getListState(descriptor);

        if (context.isRestored()) {
            for (Tuple2<String, Integer> element : checkpointedState.get()) {
                bufferedElements.add(element);
            }
        }
    }
}

• 这个BufferingSink实现了CheckpointedFunction接口，它定义了ListState类型的checkpointedState，以及List结构的bufferedElements
• 在invoke方法里头先将value缓存到bufferedElements，缓存个数触发阈值时，执行sink操作，然后清空bufferedElements
• 在snapshotState方法里头对bufferedElements进行snapshot操作，在initializeState先创建ListStateDescriptor，然后通过FunctionInitializationContext.getOperatorStateStore().getListState(descriptor)来获取ListState，之后判断state是否有在前一次execution的snapshot中restored，如果有则将ListState中的数据恢复到bufferedElements

CheckpointedFunction

flink-streaming-java_2.11-1.7.0-sources.jar!/org/apache/flink/streaming/api/checkpoint/CheckpointedFunction.java

@PublicEvolving
@SuppressWarnings("deprecation")
public interface CheckpointedFunction {

    /**
     * This method is called when a snapshot for a checkpoint is requested. This acts as a hook to the function to
     * ensure that all state is exposed by means previously offered through {@link FunctionInitializationContext} when
     * the Function was initialized, or offered now by {@link FunctionSnapshotContext} itself.
     *
     * @param context the context for drawing a snapshot of the operator
     * @throws Exception
     */
    void snapshotState(FunctionSnapshotContext context) throws Exception;

    /**
     * This method is called when the parallel function instance is created during distributed
     * execution. Functions typically set up their state storing data structures in this method.
     *
     * @param context the context for initializing the operator
     * @throws Exception
     */
    void initializeState(FunctionInitializationContext context) throws Exception;

}

• CheckpointedFunction是stateful transformation functions的核心接口，用于跨stream维护state
• snapshotState在checkpoint的时候会被调用，用于snapshot state，通常用于flush、commit、synchronize外部系统
• initializeState在parallel function初始化的时候(第一次初始化或者从前一次checkpoint recover的时候)被调用，通常用来初始化state，以及处理state recovery的逻辑

FunctionSnapshotContext

flink-runtime_2.11-1.7.0-sources.jar!/org/apache/flink/runtime/state/FunctionSnapshotContext.java

/**
 * This interface provides a context in which user functions that use managed state (i.e. state that is managed by state
 * backends) can participate in a snapshot. As snapshots of the backends themselves are taken by the system, this
 * interface mainly provides meta information about the checkpoint.
 */
@PublicEvolving
public interface FunctionSnapshotContext extends ManagedSnapshotContext {
}

• FunctionSnapshotContext继承了ManagedSnapshotContext接口

ManagedSnapshotContext

flink-runtime_2.11-1.7.0-sources.jar!/org/apache/flink/runtime/state/ManagedSnapshotContext.java

/**
 * This interface provides a context in which operators that use managed state (i.e. state that is managed by state
 * backends) can perform a snapshot. As snapshots of the backends themselves are taken by the system, this interface
 * mainly provides meta information about the checkpoint.
 */
@PublicEvolving
public interface ManagedSnapshotContext {

    /**
     * Returns the ID of the checkpoint for which the snapshot is taken.
     * 
     * <p>The checkpoint ID is guaranteed to be strictly monotonously increasing across checkpoints.
     * For two completed checkpoints <i>A</i> and <i>B</i>, {@code ID_B > ID_A} means that checkpoint
     * <i>B</i> subsumes checkpoint <i>A</i>, i.e., checkpoint <i>B</i> contains a later state
     * than checkpoint <i>A</i>.
     */
    long getCheckpointId();

    /**
     * Returns timestamp (wall clock time) when the master node triggered the checkpoint for which
     * the state snapshot is taken.
     */
    long getCheckpointTimestamp();
}

• ManagedSnapshotContext定义了getCheckpointId、getCheckpointTimestamp方法

FunctionInitializationContext

flink-runtime_2.11-1.7.0-sources.jar!/org/apache/flink/runtime/state/FunctionInitializationContext.java

/**
 * This interface provides a context in which user functions can initialize by registering to managed state (i.e. state
 * that is managed by state backends).
 *
 * <p>
 * Operator state is available to all functions, while keyed state is only available for functions after keyBy.
 *
 * <p>
 * For the purpose of initialization, the context signals if the state is empty or was restored from a previous
 * execution.
 *
 */
@PublicEvolving
public interface FunctionInitializationContext extends ManagedInitializationContext {
}

• FunctionInitializationContext继承了ManagedInitializationContext接口

ManagedInitializationContext

flink-runtime_2.11-1.7.0-sources.jar!/org/apache/flink/runtime/state/ManagedInitializationContext.java

/**
 * This interface provides a context in which operators can initialize by registering to managed state (i.e. state that
 * is managed by state backends).
 *
 * <p>
 * Operator state is available to all operators, while keyed state is only available for operators after keyBy.
 *
 * <p>
 * For the purpose of initialization, the context signals if the state is empty (new operator) or was restored from
 * a previous execution of this operator.
 *
 */
public interface ManagedInitializationContext {

    /**
     * Returns true, if state was restored from the snapshot of a previous execution. This returns always false for
     * stateless tasks.
     */
    boolean isRestored();

    /**
     * Returns an interface that allows for registering operator state with the backend.
     */
    OperatorStateStore getOperatorStateStore();

    /**
     * Returns an interface that allows for registering keyed state with the backend.
     */
    KeyedStateStore getKeyedStateStore();

}

• ManagedInitializationContext接口定义了isRestored、getOperatorStateStore、getKeyedStateStore方法

小结

• flink有两种基本的state，分别是Keyed State以及Operator State(non-keyed state)；其中Keyed State只能在KeyedStream上的functions及operators上使用；每个operator state会跟parallel operator中的一个实例绑定；Operator State支持parallelism变更时进行redistributing
• Keyed State及Operator State都分别有managed及raw两种形式，managed由flink runtime来管理，由runtime负责encode及写入checkpoint；raw形式的state由operators自己管理，flink runtime无法了解该state的数据结构，将其视为raw bytes；所有的datastream function都可以使用managed state，而raw state一般仅限于自己实现operators来使用
• stateful function可以通过CheckpointedFunction接口或者ListCheckpointed接口来使用managed operator state；CheckpointedFunction定义了snapshotState、initializeState两个方法；每当checkpoint执行的时候，snapshotState会被调用；而initializeState方法在每次用户定义的function初始化的时候(第一次初始化或者从前一次checkpoint recover的时候)被调用，该方法不仅可以用来初始化state，还可以用于处理state recovery的逻辑
• 对于manageed operator state，目前仅仅支持list-style的形式，即要求state是serializable objects的List结构，方便在rescale的时候进行redistributed；关于redistribution schemes的模式目前有两种，分别是Even-split redistribution(在restore/redistribution的时候每个operator仅仅得到整个state的sublist)及Union redistribution(在restore/redistribution的时候每个operator得到整个state的完整list)
• FunctionSnapshotContext继承了ManagedSnapshotContext接口，它定义了getCheckpointId、getCheckpointTimestamp方法；FunctionInitializationContext继承了ManagedInitializationContext接口，它定义了isRestored、getOperatorStateStore、getKeyedStateStore方法，可以用来判断是否是在前一次execution的snapshot中restored，以及获取OperatorStateStore、KeyedStateStore对象

doc

• Using Managed Operator State