聊聊flink的PartitionableListState

序

本文主要研究一下flink的PartitionableListState

PartitionableListState

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

/**
     * Implementation of operator list state.
     *
     * @param <S> the type of an operator state partition.
     */
    static final class PartitionableListState<S> implements ListState<S> {

        /**
         * Meta information of the state, including state name, assignment mode, and serializer
         */
        private RegisteredOperatorStateBackendMetaInfo<S> stateMetaInfo;

        /**
         * The internal list the holds the elements of the state
         */
        private final ArrayList<S> internalList;

        /**
         * A serializer that allows to perform deep copies of internalList
         */
        private final ArrayListSerializer<S> internalListCopySerializer;

        PartitionableListState(RegisteredOperatorStateBackendMetaInfo<S> stateMetaInfo) {
            this(stateMetaInfo, new ArrayList<S>());
        }

        private PartitionableListState(
                RegisteredOperatorStateBackendMetaInfo<S> stateMetaInfo,
                ArrayList<S> internalList) {

            this.stateMetaInfo = Preconditions.checkNotNull(stateMetaInfo);
            this.internalList = Preconditions.checkNotNull(internalList);
            this.internalListCopySerializer = new ArrayListSerializer<>(stateMetaInfo.getPartitionStateSerializer());
        }

        private PartitionableListState(PartitionableListState<S> toCopy) {

            this(toCopy.stateMetaInfo.deepCopy(), toCopy.internalListCopySerializer.copy(toCopy.internalList));
        }

        public void setStateMetaInfo(RegisteredOperatorStateBackendMetaInfo<S> stateMetaInfo) {
            this.stateMetaInfo = stateMetaInfo;
        }

        public RegisteredOperatorStateBackendMetaInfo<S> getStateMetaInfo() {
            return stateMetaInfo;
        }

        public PartitionableListState<S> deepCopy() {
            return new PartitionableListState<>(this);
        }

        @Override
        public void clear() {
            internalList.clear();
        }

        @Override
        public Iterable<S> get() {
            return internalList;
        }

        @Override
        public void add(S value) {
            Preconditions.checkNotNull(value, "You cannot add null to a ListState.");
            internalList.add(value);
        }

        @Override
        public String toString() {
            return "PartitionableListState{" +
                    "stateMetaInfo=" + stateMetaInfo +
                    ", internalList=" + internalList +
                    '}';
        }

        public long[] write(FSDataOutputStream out) throws IOException {

            long[] partitionOffsets = new long[internalList.size()];

            DataOutputView dov = new DataOutputViewStreamWrapper(out);

            for (int i = 0; i < internalList.size(); ++i) {
                S element = internalList.get(i);
                partitionOffsets[i] = out.getPos();
                getStateMetaInfo().getPartitionStateSerializer().serialize(element, dov);
            }

            return partitionOffsets;
        }

        @Override
        public void update(List<S> values) {
            internalList.clear();

            addAll(values);
        }

        @Override
        public void addAll(List<S> values) {
            if (values != null && !values.isEmpty()) {
                internalList.addAll(values);
            }
        }
    }

• PartitionableListState是DefaultOperatorStateBackend使用的ListState實現，其內部使用的是ArrayList(internalList)來存儲state，而stateMetaInfo使用的是RegisteredOperatorStateBackendMetaInfo；其write方法將internalList的數據序列化到FSDataOutputStream，並返回每一個記錄對應的offset數組(partitionOffsets)

ListState

flink-core-1.7.0-sources.jar!/org/apache/flink/api/common/state/ListState.java

/**
 * {@link State} interface for partitioned list state in Operations.
 * The state is accessed and modified by user functions, and checkpointed consistently
 * by the system as part of the distributed snapshots.
 *
 * <p>The state is only accessible by functions applied on a {@code KeyedStream}. The key is
 * automatically supplied by the system, so the function always sees the value mapped to the
 * key of the current element. That way, the system can handle stream and state partitioning
 * consistently together.
 *
 * @param <T> Type of values that this list state keeps.
 */
@PublicEvolving
public interface ListState<T> extends MergingState<T, Iterable<T>> {

    /**
     * Updates the operator state accessible by {@link #get()} by updating existing values to
     * to the given list of values. The next time {@link #get()} is called (for the same state
     * partition) the returned state will represent the updated list.
     *
     * <p>If null or an empty list is passed in, the state value will be null.
     *
     * @param values The new values for the state.
     *
     * @throws Exception The method may forward exception thrown internally (by I/O or functions).
     */
    void update(List<T> values) throws Exception;

    /**
     * Updates the operator state accessible by {@link #get()} by adding the given values
     * to existing list of values. The next time {@link #get()} is called (for the same state
     * partition) the returned state will represent the updated list.
     *
     * <p>If null or an empty list is passed in, the state value remains unchanged.
     *
     * @param values The new values to be added to the state.
     *
     * @throws Exception The method may forward exception thrown internally (by I/O or functions).
     */
    void addAll(List<T> values) throws Exception;
}

• ListState主要用於operation存儲partitioned list state，它繼承了MergingState接口(指定OUT的泛型爲Iterable<T>)，同時聲明瞭兩個方法；其中update用於全量更新state，若是參數爲null或者empty，那麼state會被清空；addAll方法用於增量更新，若是參數爲null或者empty，則保持不變，不然則新增給定的values

MergingState

flink-core-1.7.0-sources.jar!/org/apache/flink/api/common/state/MergingState.java

/**
 * Extension of {@link AppendingState} that allows merging of state. That is, two instances
 * of {@link MergingState} can be combined into a single instance that contains all the
 * information of the two merged states.
 *
 * @param <IN> Type of the value that can be added to the state.
 * @param <OUT> Type of the value that can be retrieved from the state.
 */
@PublicEvolving
public interface MergingState<IN, OUT> extends AppendingState<IN, OUT> { }

• MergingState接口僅僅是繼承了AppendingState接口，用接口命名錶示該state支持state合併

AppendingState

flink-core-1.7.0-sources.jar!/org/apache/flink/api/common/state/AppendingState.java

/**
 * Base interface for partitioned state that supports adding elements and inspecting the current
 * state. Elements can either be kept in a buffer (list-like) or aggregated into one value.
 *
 * <p>The state is accessed and modified by user functions, and checkpointed consistently
 * by the system as part of the distributed snapshots.
 *
 * <p>The state is only accessible by functions applied on a {@code KeyedStream}. The key is
 * automatically supplied by the system, so the function always sees the value mapped to the
 * key of the current element. That way, the system can handle stream and state partitioning
 * consistently together.
 *
 * @param <IN> Type of the value that can be added to the state.
 * @param <OUT> Type of the value that can be retrieved from the state.
 */
@PublicEvolving
public interface AppendingState<IN, OUT> extends State {

    /**
     * Returns the current value for the state. When the state is not
     * partitioned the returned value is the same for all inputs in a given
     * operator instance. If state partitioning is applied, the value returned
     * depends on the current operator input, as the operator maintains an
     * independent state for each partition.
     *
     * <p><b>NOTE TO IMPLEMENTERS:</b> if the state is empty, then this method
     * should return {@code null}.
     *
     * @return The operator state value corresponding to the current input or {@code null}
     * if the state is empty.
     *
     * @throws Exception Thrown if the system cannot access the state.
     */
    OUT get() throws Exception;

    /**
     * Updates the operator state accessible by {@link #get()} by adding the given value
     * to the list of values. The next time {@link #get()} is called (for the same state
     * partition) the returned state will represent the updated list.
     *
     * <p>If null is passed in, the state value will remain unchanged.
     *
     * @param value The new value for the state.
     *
     * @throws Exception Thrown if the system cannot access the state.
     */
    void add(IN value) throws Exception;

}

• AppendingState是partitioned state的基本接口，它繼承了State接口，同時聲明瞭get、add兩個方法；get方法用於返回當前state的值，若是爲空則返回null；add方法用於給state添加值

State

flink-core-1.7.0-sources.jar!/org/apache/flink/api/common/state/State.java

/**
 * Interface that different types of partitioned state must implement.
 *
 * <p>The state is only accessible by functions applied on a {@code KeyedStream}. The key is
 * automatically supplied by the system, so the function always sees the value mapped to the
 * key of the current element. That way, the system can handle stream and state partitioning
 * consistently together.
 */
@PublicEvolving
public interface State {

    /**
     * Removes the value mapped under the current key.
     */
    void clear();
}

• State接口定義了全部不一樣partitioned state實現必須實現的方法，這裏定義了clear方法用於清空當前state的全部值

RegisteredOperatorStateBackendMetaInfo

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

/**
 * Compound meta information for a registered state in an operator state backend.
 * This contains the state name, assignment mode, and state partition serializer.
 *
 * @param <S> Type of the state.
 */
public class RegisteredOperatorStateBackendMetaInfo<S> extends RegisteredStateMetaInfoBase {

    /**
     * The mode how elements in this state are assigned to tasks during restore
     */
    @Nonnull
    private final OperatorStateHandle.Mode assignmentMode;

    /**
     * The type serializer for the elements in the state list
     */
    @Nonnull
    private final TypeSerializer<S> partitionStateSerializer;

    public RegisteredOperatorStateBackendMetaInfo(
            @Nonnull String name,
            @Nonnull TypeSerializer<S> partitionStateSerializer,
            @Nonnull OperatorStateHandle.Mode assignmentMode) {
        super(name);
        this.partitionStateSerializer = partitionStateSerializer;
        this.assignmentMode = assignmentMode;
    }

    private RegisteredOperatorStateBackendMetaInfo(@Nonnull RegisteredOperatorStateBackendMetaInfo<S> copy) {
        this(
            Preconditions.checkNotNull(copy).name,
            copy.partitionStateSerializer.duplicate(),
            copy.assignmentMode);
    }

    @SuppressWarnings("unchecked")
    public RegisteredOperatorStateBackendMetaInfo(@Nonnull StateMetaInfoSnapshot snapshot) {
        this(
            snapshot.getName(),
            (TypeSerializer<S>) Preconditions.checkNotNull(
                snapshot.restoreTypeSerializer(StateMetaInfoSnapshot.CommonSerializerKeys.VALUE_SERIALIZER)),
            OperatorStateHandle.Mode.valueOf(
                snapshot.getOption(StateMetaInfoSnapshot.CommonOptionsKeys.OPERATOR_STATE_DISTRIBUTION_MODE)));
        Preconditions.checkState(StateMetaInfoSnapshot.BackendStateType.OPERATOR == snapshot.getBackendStateType());
    }

    /**
     * Creates a deep copy of the itself.
     */
    @Nonnull
    public RegisteredOperatorStateBackendMetaInfo<S> deepCopy() {
        return new RegisteredOperatorStateBackendMetaInfo<>(this);
    }

    @Nonnull
    @Override
    public StateMetaInfoSnapshot snapshot() {
        return computeSnapshot();
    }

    //......

    @Nonnull
    private StateMetaInfoSnapshot computeSnapshot() {
        Map<String, String> optionsMap = Collections.singletonMap(
            StateMetaInfoSnapshot.CommonOptionsKeys.OPERATOR_STATE_DISTRIBUTION_MODE.toString(),
            assignmentMode.toString());
        String valueSerializerKey = StateMetaInfoSnapshot.CommonSerializerKeys.VALUE_SERIALIZER.toString();
        Map<String, TypeSerializer<?>> serializerMap =
            Collections.singletonMap(valueSerializerKey, partitionStateSerializer.duplicate());
        Map<String, TypeSerializerSnapshot<?>> serializerConfigSnapshotsMap =
            Collections.singletonMap(valueSerializerKey, partitionStateSerializer.snapshotConfiguration());

        return new StateMetaInfoSnapshot(
            name,
            StateMetaInfoSnapshot.BackendStateType.OPERATOR,
            optionsMap,
            serializerConfigSnapshotsMap,
            serializerMap);
    }
}

• RegisteredOperatorStateBackendMetaInfo繼承了抽象類RegisteredStateMetaInfoBase，實現了snapshot的抽象方法，這裏是經過computeSnapshot方法來實現；computeSnapshot方法主要是構造StateMetaInfoSnapshot所需的optionsMap、serializerConfigSnapshotsMap、serializerMap

小結

• flink的manageed operator state僅僅支持ListState，DefaultOperatorStateBackend使用的ListState實現是PartitionableListState，其內部使用的是ArrayList(internalList)來存儲state，而stateMetaInfo使用的是RegisteredOperatorStateBackendMetaInfo
• PartitionableListState實現了ListState接口(update、addAll方法)；而ListState接口繼承了MergingState接口(指定OUT的泛型爲Iterable<T>)；MergingState接口沒有聲明其餘方法，它繼承了AppendingState接口；AppendingState接口繼承了State接口，同時聲明瞭get、add方法；State接口則定義了clear方法
• RegisteredOperatorStateBackendMetaInfo繼承了抽象類RegisteredStateMetaInfoBase，實現了snapshot的抽象方法，這裏是經過computeSnapshot方法來實現；computeSnapshot方法主要是構造StateMetaInfoSnapshot所需的optionsMap、serializerConfigSnapshotsMap、serializerMap

doc

• ListState
• flink state package summary
• Using Managed Operator State