Status

...

...

...

...

...

...

...

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

Motivation

We propose to add a new type of state that can be used to create simpler and more efficient implementations of various temporal operations, including temporal joins and sorting.

What these operations have in common is the need to store timestamped records for later retrieval, either indexed by, or relative to, a timestamp. This is something RocksDB can do very efficiently, if we let it.

the operators that maintain a MapState<?, List<?>> and/or require range operations. This includes temporal joins, operators that sort their input streams by timestamp, and a few other cases:

• TemporalRowTimeJoinOperator
• RowTimeSortOperator
• CepOperator
• IntervalJoinOperator
• AbstractStreamArrowPythonOverWindowAggregateFunctionOperator
• The various over/rank window operators
• Min/max with retractions
• Custom window implementations using keyed process functions

Several of these operators More concretely, there are many Flink operators that are using MapState<Long, List<T>> to store lists of records associated with timestamps. Updating these those lists is expensive, especially when they are stored in RocksDB, as that requires deserializing and reserializing the lists when appending new list entries.

With this proposal, this MapState with lists as values becomes TemporalListState<T>. We can use the timestamps that were the MapState’s keys as the common case of MapState<Long, List<T>> becomes BinarySortedState<Long, T>. The implementation can exploit namespaces in the state backends for representing the keys (timestamps, in this case), and the MapState’s values can be stored as lists using each state backend’s native ListState implementation.

This means that for the RocksDB state backend, the implementation of TemporalListState can use the optimized append operation from ListState. This avoids having to deserialize and reserialize the list when adding items to the end of the list, which is a very common (and very expensive) operation.

Another opportunity for optimization is to But rather than limiting the scope to using Longs as the keys in BinarySortedState<UK, UV>, we propose instead to support any type for which a lexicographically ordered type serializer can be implemented. This means that when comparing any two keys, a and b, their binary, serialized representations will be ordered in exactly the same way as the original objects, a and b. We can then take advantage of the fact that RocksDB stores its key/value pairs in order, sorted by key. We can use this to reduce the cost of searching for the key closest to a particular timestamp, an operation needed, for example, when computing a temporal jointhe serialized representations of the keys, to implement efficient range scans.

This FLIP will also make makes it possible for users to create custom time windows that are implemented as keyed process functions that achieve performance similar to that of the built-in window operators. This isn’t currently possible because the window operators take advantage of the state backends’ namespaces to optimize state access in a way as outlined above, something that can’t be done with Flink’s public APIs. By exposing these namespaces in this limited, controlled way, users will be able to use temporal state to achieve comparable performance. We have verified this with a POC implementation.

We have also implemented as a POC a revised version of the TemporalRowTimeJoinOperator, and found it to have more than 2x the throughput of the current implementation.

Public Interfaces

Two new keyed state interfaces are introduced – TemporalValueState and TemporalListState – each extending the new TemporalState interface:

An implementation is underway. Benchmarking has shown a ~130% increase in throughput for sorting, and ~60% for temporal joins.

Public Interfaces

The complete proposed interface for BinarySortedState<UK, UV> is shown below. In summary, it includes:

Point operations:

• valuesAt(key)
• add(key, value)
• put(key, values)
• clearEntryAt(key)

Lookups:

• firstEntry(), firstEntry(fromKey)
• lastEntry(), lastEntry(UK toKey)

Range-operations:

• readRange(fromKey, toKey, inclusiveToKey)
• readRangeUntil(toKey, inclusiveToKey)
• readRangeFrom(fromKey)

Range-deletes:

• clearEntryAt(key)
• clearRange(fromKey, toKey, inclusiveToKey)
• clearRangeUntil(toKey, inclusiveToKey)
• clearRangeFrom(fromKey)

Cleanup:

• clear()

@PublicEvolving
public interface TemporalStateBinarySortedState<UK, UV> extends State {

       /** Removes the value mapped for the specified time under the current key. */
    void clear(long time);

    /** Removes all values for all timestamps under the current key. */
    void clearAll();
}

Both the value and list flavors of temporal state support storing data for a particular timestamp, and retrieving the data either at, at-or-before, or at-or-after a particular timestamp:

public interface TemporalValueState<T> extends TemporalState {

    T valueAt(long time) throws IOException;

    TemporalValue<T> valueAtOrBefore(long time) throws IOException;

    TemporalValue<T> valueAtOrAfter(long time) throws IOException;

    void update(long time, T value) throws IOException;
}

public interface TemporalListState<T> extends TemporalState {

    @Nullable
    Iterable<T> getAt(long time) throws Exception;

    @Nullable
    TemporalValue<Iterable<T>> getAtOrBefore(long time) throws Exception;

    @Nullable
    TemporalValue<Iterable<T>> getAtOrAfter(long time) throws Exception;

    void update(long time, List<T> values) throws Exception;

    void addAll(long time, List<T> values) throws Exception;

    void add(long time, T value) throws Exception;
}

Under this proposal, RuntimeContext and KeyedStateStore are extended with methods for registering TemporalValueState and TemporalListState using the existing ValueStateDescriptor and ListStateDescriptor types.

public interface RuntimeContext {

    <T> TemporalValueState<T> getTemporalValueState(ValueStateDescriptor<T> stateProperties);

    <T> TemporalListState<T> getTemporalListState(ListStateDescriptor<T> stateProperties);
}

public interface KeyedStateStore {

    <T> TemporalValueState<T> getTemporalValueState(ValueStateDescriptor<T> stateProperties);

    <T> TemporalListState<T> getTemporalListState(ListStateDescriptor<T> stateProperties);
}

Proposed Changes

There are many places in Flink SQL, for example, where something like this is used:

private transient MapState<Long, List<RowData>> dataState;

Under this proposal, this will come

private transient TemporalListState<RowData> dataState;

Having the getAtOrBefore and getAtOrAfter methods available for these temporal state types makes it much easier to implement operations such as temporal joins and temporal sorting. By providing optimized implementations of these methods for each state backend, the simpler implementations of these temporal operations can also be more performant.

EmbeddedRocksDBStateBackend: Add a LexicographicLongSerializer

In order to have an efficient implementation of the getAtOrBefore and getAtOrAfter methods for the RocksDB state backend, we want to leverage the iterators available in RocksDB. 

For this purpose, we have implemented a new LexicographicLongSerializer that serializes the timestamps, which are longs, with an unsigned big-endian encoding so that their binary sort order matches the numerical sort order. 

HashMapStateBackend

     * Returns the current values associated with the given key.
     *
     * @param key The key of the mapping
     * @return The values of the mapping with the given key (or <tt>null</tt> if there is none)
     * @throws Exception Thrown if the system cannot access the state.
     */
    @Nullable
    Iterable<UV> valuesAt(UK key) throws Exception;

    /**
     * Returns the current values associated with the first key in state, based on the state's
     * ordering function.
     *
     * @return The entry of the first key (or <tt>null</tt> if there is none)
     * @throws Exception Thrown if the system cannot access the state.
     */
    @Nullable
    Map.Entry<UK, Iterable<UV>> firstEntry() throws Exception;

    /**
     * Returns the current values associated with the first key in state at or after {@code
     * fromKey}, based on the state's ordering function.
     *
     * @param fromKey The query key (inclusive)
     * @return The entry of the first key after {@code fromKey} (or <tt>null</tt> if there is none)
     * @throws Exception Thrown if the system cannot access the state.
     */
    @Nullable
    Map.Entry<UK, Iterable<UV>> firstEntry(UK fromKey) throws Exception;

    /**
     * Returns the current values associated with the last key in state, based on the state's
     * ordering function.
     *
     * @return The entry of the last key (or <tt>null</tt> if there is none)
     * @throws Exception Thrown if the system cannot access the state.
     */
    @Nullable
    Map.Entry<UK, Iterable<UV>> lastEntry() throws Exception;

    /**
     * Returns the current values associated with the last key in state at or before {@code toKey},
     * based on the state's ordering function.
     *
     * @param toKey The query key (inclusive)
     * @return The entry of the last key at or before {@code toKey} (or <tt>null</tt> if there is
     *     none)
     * @throws Exception Thrown if the system cannot access the state.
     */
    @Nullable
    Map.Entry<UK, Iterable<UV>> lastEntry(UK toKey) throws Exception;

    /**
     * Associates a new set of values with the given key (overwrites the previous mapping).
     *
     * @param key The key of the mapping
     * @param values The new values of the mapping
     * @throws Exception Thrown if the system cannot access the state.
     */
    void put(UK key, Iterable<UV> values) throws Exception;

    /**
     * Adds a new value to the mapping with the given key.
     *
     * @param key The key of the mapping
     * @param value The new value to add to the mapping
     * @throws Exception Thrown if the system cannot access the state.
     */
    void add(UK key, UV value) throws Exception;

    /**
     * Returns all the mappings in the state for the given range of keys.
     *
     * @param fromKey The first key of the mappings to return (inclusive)
     * @param toKey The last key of the mappings to return
     * @param inclusiveToKey Whether {@code toKey} should be inclusive or not
     * @return An iterable view of all the state's key-values pairs in the given range.
     * @throws Exception Thrown if the system cannot access the state.
     */
    Iterable<Map.Entry<UK, Iterable<UV>>> readRange(UK fromKey, UK toKey, boolean inclusiveToKey)
            throws Exception;

    /**
     * Returns all the mappings in the state for the range of the smallest key until the given key.
     *
     * @param toKey The last key of the mappings to return
     * @param inclusiveToKey Whether {@code toKey} should be inclusive or not
     * @return An iterable view of all the state's key-values pairs in the given range.
     * @throws Exception Thrown if the system cannot access the state.
     */
    Iterable<Map.Entry<UK, Iterable<UV>>> readRangeUntil(UK toKey, boolean inclusiveToKey)
            throws Exception;

    /**
     * Returns all the mappings in the state for the range of the given key (inclusive) until the
     * largest key.
     *
     * @param fromKey The first key of the mappings to return
     * @return An iterable view of all the state's key-values pairs in the given range.
     * @throws Exception Thrown if the system cannot access the state.
     */
    Iterable<Map.Entry<UK, Iterable<UV>>> readRangeFrom(UK fromKey) throws Exception;

    /**
     * Deletes all values for the mapping with the given key.
     *
     * @param key The key of the mapping
     * @throws Exception Thrown if the system cannot access the state.
     */
    void clearEntryAt(UK key) throws Exception;

    /**
     * Deletes all values in the given range of keys.
     *
     * @param fromKey The first key of the mappings to remove (inclusive)
     * @param toKey The last key of the mappings to remove
     * @param inclusiveToKey Whether {@code toKey} should be inclusive or not
     * @throws Exception Thrown if the system cannot access the state.
     */
    void clearRange(UK fromKey, UK toKey, boolean inclusiveToKey) throws Exception;

    /**
     * Deletes all values in the range of the smallest key until the given key.
     *
     * @param toKey The last key of the mappings to remove
     * @param inclusiveToKey Whether {@code toKey} should be inclusive or not
     * @throws Exception Thrown if the system cannot access the state.
     */
    void clearRangeUntil(UK toKey, boolean inclusiveToKey) throws Exception;

    /**
     * Deletes all values in the range of the given key (inclusive) until the largest key.
     *
     * @param fromKey The first key of the mappings to remove
     * @throws Exception Thrown if the system cannot access the state.
     */
    void clearRangeFrom(UK fromKey) throws Exception;

    /** Deletes all values from the state. */
    void clear();
}

Under this proposal, a new state descriptor type is added:

@PublicEvolving
public class BinarySortedStateDescriptor<UK, UV> extends ListStateDescriptor<UV> {

    /**
     * Create a new {@code BinarySortedStateDescriptor} with the given name and the given type
     * serializers.
     *
     * @param name The name of the {@code MapStateDescriptor}.
     * @param keySerializer The type serializer for the keys in the state.
     * @param valueSerializer The type serializer for the values in the state.
     */
    public BinarySortedStateDescriptor(
            String name,
            LexicographicalTypeSerializer<UK> keySerializer,
            TypeSerializer<UV> valueSerializer);

    /**
     * Create a new {@code BinarySortedStateDescriptor} with the given name and the given type
     * information.
     *
     * @param name The name of the {@code MapStateDescriptor}.
     * @param keySerializer The type serializer for the keys in the state.
     * @param valueTypeInfo The type information for the values in the state.
     */
    public BinarySortedStateDescriptor(
            String name,
            LexicographicalTypeSerializer<UK> keySerializer,
            TypeInformation<UV> valueTypeInfo);

    /**
     * Create a new {@code BinarySortedStateDescriptor} with the given name and the given type
     * information.
     *
     * <p>If this constructor fails (because it is not possible to describe the type via a class),
     * consider using the {@link #BinarySortedStateDescriptor(String, LexicographicalTypeSerializer,
     * TypeInformation)} constructor.
     *
     * @param name The name of the {@code MapStateDescriptor}.
     * @param keySerializer The type serializer for the keys in the state.
     * @param valueType The class of the type of values in the state.
     */
    public BinarySortedStateDescriptor(
            String name, LexicographicalTypeSerializer<UK> keySerializer, Class<UV> valueType);

    /**
     * Gets the serializer for the keys in the state.
     *
     * @return The serializer for the keys in the state.
     */
    public LexicographicalTypeSerializer<UK> getKeySerializer();
}

Here the LexicographicalTypeSerializer<T> is a special TypeSerializer that provides a well-defined binary sort order that we will use in the operations above (ideally the same that you would expect from the original Java objects). For better performance with heap-based state-backends, however, you can also provide an efficient Java object comparator based on the data type's binary sort order. This way, heap-based backends don't have to serialize keys in order to sort them. The proposal is to initially provide only an internal implementation of a LexicographicLongSerializer in order to support BinarySortedState<Long, T> out of the box. Users will be able to extend LexicographicalTypeSerializer if they want to use other types as the keys for ordering the state.

@PublicEvolving
public abstract class LexicographicalTypeSerializer<T> extends TypeSerializer<T> {

    /**
     * Provides an efficient Java object comparator based on this data type's binary sort order.
     *
     * @return If empty, implementations can compare based on serialized form and using {@link
     *     BinarySortOrderComparator#INSTANCE}, otherwise use the provided (efficient) comparator.
     */
    public Optional<Comparator<T>> findComparator() {}

    public static final class BinarySortOrderComparator implements Comparator<byte[]> {

        public static final BinarySortOrderComparator INSTANCE = new BinarySortOrderComparator();

        @Override
        public int compare(byte[] first, byte[] second) {}

        /**
         * Variant of {@link #compare(byte[], byte[])} that starts the comparison from a given
         * position. This can be used if you know that the first <code>from</code> bytes are equal.
         *
         * @param first the first object to be compared.
         * @param second the second object to be compared.
         * @param from array position to start the comparison at
         * @return a negative integer, zero, or a positive integer as the first argument is less
         *     than, equal to, or greater than the second.
         * @see #compare(byte[], byte[])
         */
        public int compare(byte[] first, byte[] second, int from) {}
    }
}

Here's an example of how these pieces fit together:

private transient BinarySortedState<Long, Long> events;

BinarySortedStateDescriptor<Long, Long> desc =
    new BinarySortedStateDescriptor<>(
        "Events",
        LexicographicLongSerializer.INSTANCE,
        Types.LONG);

events = getRuntimeContext()
    .getBinarySortedState(desc);

As illustrated above, RuntimeContext (and KeyedStateStore) need to be extended with a method for registering BinarySortedState:

    @PublicEvolving
    <UK, UV> BinarySortedState<UK, UV> getBinarySortedState(
        BinarySortedStateDescriptor<UK, UV> stateProperties);

Proposed Changes

EmbeddedRocksDBStateBackend

The RocksDB state backend is extended to use the state descriptor's LexicographicalTypeSerializer as the namespace serializer.

A RocksDBBinarySortedNavigator is added to support the various lookup and range operations. It takes advantage of RocksDB's native iterators.

HashMapStateBackend

A HeapBinarySortedNavigator is needed to provide the same lookup and range operations for the heap-based state backend. The proposed implementation uses a

HashMap<FK, TreeSet<UKI>> to keep track of the namespaces that need to be iterated overOur intended implementation is to lazily create a sorted list of the timestamps for the current key in response to the first call to getAtOrBefore or getAtOrBefore, and then reuse that list until it becomes invalid.

Operators that can potentially be improved by using

...

BinarySortedState

As outlined in the Motivation section, many of Flink's operators can benefit by being updated to leverage this new state type. 

However, in a first release, we plan to limit the scope, and only include the new state type

...

• TemporalRowTimeJoinOperator
• RowTimeSortOperator
• CepOperator
• IntervalJoinOperator
• AbstractStreamArrowPythonOverWindowAggregateFunctionOperator

• The various over/rank window operators

New Examples

We have implemented two new examples: a temporal join, and a temporal sort.

New Documentation

In addition to javadocs, state#using-keyed-state should needs to be updated to include TemporalValueState and TemporalListState.

Updated Training

docs/learn-flink/event_driven/#example should be updatedBinarySortedState.

Compatibility, Deprecation, and Migration Plan

Deprecation

While this FLIP does open the door for eventually deprecating some of the existing (and often abused) mechanisms for customizing DataStream windows, we are not proposing to deprecate anything as part of this FLIP.

Compatibility/Migration

Any operators that are reworked to use temporal binary sorted state will thereafter use state in an incompatible way, compared to their current implementations.

To support upgrades to any SQL operator that is reworked to use temporal binary sorted state, the mechanisms of FLIP-190 for handling operator migrations need to be followed. This involves annotating the ExecNode of the operator with an updated version number, and then modifying the operator so that it works regardless of which version of the state it finds available. 

These operators can then either migrate the state, or directly support both the old and new versions of the state. In the context of this FLIP, migrating from any MapState<Long, List<T>> that is found to TemporalListState<T> BinarySortedState<Long, T> will be simpler than continuing to maintain the old state as MapState, so that is the preferred approach.

Test Plan

We have implemented will include unit tests for the new state types that are tested against both state backends.

We should verify that any operators that are updated to use temporal state have sufficient test coverage. Tests will need to be added for state migration.

Rejected Alternatives

Directly expose the state backend namespaces

By itself, this wouldn’t accomplish much, and the interfaces proposed here satisfy the objectives without exposing this internal API.

Provide iterators over the per-key timestamps

Many use cases only need to fetch a single value, and don’t need iteration. So while providing users with iterators might offer a slight performance increase in some cases (by keeping low-level/native pointers into the iteration), the added complexity doesn’t seem to be justified.

Add support for TemporalMapState

We haven’t found a motivating use case.

SortedMapState

Limit the scope to only offering TemporalState

An earlier draft of this FLIP proposed adding TemporalValueState and TemporalListState types, roughly equivalent to BinarySortedState<Long, ?>. After some discussion on the mailing list, this was determined to be too limiting. The idea here would be to support MapState where the key type could be any type that implements Comparable. This would be a generalization of TemporalState, which is similar, but limited to only supporting Longs as the keys. This was rejected as being overly difficult to implement in a way that would cleanly leverage RocksDB’s iterators, and we didn’t have a motivating use case that would benefit from this generalization.