Status

Current state: Adopted

Discussion thread: here

Vote thread: here

JIRA: Unable to render Jira issues macro, execution error.

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

Motivation

Fetching range of records from Kafka Streams state stores comes with an iterator to traverse elements from oldest to newest, e.g ReadOnlyWindowStore#fetch(K key, long fromTime, long toTime) mentions:

For each key, the iterator guarantees ordering of windows, starting from the oldest/earliest”

Similar guarantees are provided on other fetch and range operations. But in the case of key ranges, there are some nuances regarding order:

The returned iterator must be safe from {@link java.util.ConcurrentModificationException}s and must not return null values. No ordering guarantees are provided.

Ordering is not guaranteed as backing structure is based on maps keyed by o.a.k.common.utils.Bytes. Though, Bytes support Lexicographic byte array comparison, which defines ordering in-memory and RocksDB stores.

These APIs constraint the usage of local state store for some use-cases:

When storing records on time windows, or records by key; and an operation wants to return the last N values inserted withing a time range containing M records: Currently there is no alternative other than iterating records from oldest to newest–traversing M records, where M » N.

If a backward read direction option becomes available, then we could start from the latest record within a time range and go backwards, returning the first N value more efficiently.

At Zipkin Kafka-based storage, we are planning to use this feature to replace two KeyValueStores (one for traces indexed by id, and another with trace_ids indexed by timestamp) for one WindowStore. A backward read direction will allow to support queries like: “within this time range, find the last traces that match this criteria”, and return latest values quickly.

Internally, both implementations: persistent (RocksDB), and in-memory (TreeMap) support reverse/descending iteration:



final RocksIterator iter = db.newIterator();
iter.seekToFirst();
iter.next();
final RocksIterator reverse = db.newIterator();
reverse.seekToLast();
reverse.prev();

final TreeMap<String, String> map = new TreeMap<>();
final NavigableSet<String> nav = map.navigableKeySet();
final NavigableSet<String> rev = map.descendingKeySet();


Reference issues

Proposed Changes

There are 2 important ranges in Kafka Streams Stores:

  • Key Range
  • Time Range


For SessionStore/ReadOnlySessionStore: findSessions and findSession operations will be moved from SessionStore to ReadOnlySessionStore to align with how other stores are design.

Reverse Key Ranges

Extend existing interface for reverse KeyValueStore

public interface ReadOnlyKeyValueStore<K, V> {
    default KeyValueIterator<K, V> reverseRange(K from, K to) { 
        throw new UnsupportedOperationException(); 
    }
    default KeyValueIterator<K, V> reverseAll() {
        throw new UnsupportedOperationException(); 
    }
}

Backward Time Ranges

Window and Session stores are based on a set of KeyValue Stores (Segments) organized by a time-based index. Therefore, for these stores time-range is more important than key-range to lookup for values.

Existing stores will be extended with backward methods:

public interface ReadOnlyWindowStore<K, V> {
    default WindowStoreIterator<V> backwardFetch(K key, Instant from, Instant to) throws IllegalArgumentException {
        throw new UnsupportedOperationException(); 
    }

    default KeyValueIterator<Windowed<K>, V> backwardFetch(K from, K to, Instant fromTime, Instant toTime) throws IllegalArgumentException {
        throw new UnsupportedOperationException(); 
    }
    
    default KeyValueIterator<Windowed<K>, V> backwardAll() {
        throw new UnsupportedOperationException(); 
    }
    
    default KeyValueIterator<Windowed<K>, V> backwardFetchAll(Instant from, Instant to) throws IllegalArgumentException {
        throw new UnsupportedOperationException(); 
    }
}

public interface ReadOnlySessionStore<K, AGG> {
    // Moving read functions from SessionStore to ReadOnlySessionStore
    default KeyValueIterator<Windowed<K>, AGG> findSessions(final K key, final long earliestSessionEndTime, final long latestSessionStartTime) {
        throw new UnsupportedOperationException("Moved from SessionStore"); 
    }

    default KeyValueIterator<Windowed<K>, AGG> findSessions(final K keyFrom, final K keyTo, final long earliestSessionEndTime, final long latestSessionStartTime) {
        throw new UnsupportedOperationException("Moved from SessionStore"); 
    }

    default AGG fetchSession(final K key, final long startTime, final long endTime) {
        throw new UnsupportedOperationException("Moved from SessionStore"); 
    }

    // New 
    default KeyValueIterator<Windowed<K>, AGG> backwardFindSessions(final K key, final long earliestSessionEndTime, final long latestSessionStartTime) {
        throw new UnsupportedOperationException(); 
    }

    default KeyValueIterator<Windowed<K>, AGG> backwardFindSessions(final K keyFrom, final K keyTo, final long earliestSessionEndTime, final long latestSessionStartTime) {
        throw new UnsupportedOperationException(); 
    }

    default KeyValueIterator<Windowed<K>, AGG> backwardFetch(final K key) {
        throw new UnsupportedOperationException(); 
    }
    default KeyValueIterator<Windowed<K>, AGG> backwardFetch(final K from, final K to) {
        throw new UnsupportedOperationException(); 
    }
}


Compatibility, Deprecation, and Migration Plan

New methods will have default implementations to avoid affecting current implementations.

Rejected Alternatives

  • Create a parallel hierarchy of interfaces for backward operation. Even though this option seems like the best way to extend functionality, it was proved to not work in practice in KIP-614 discussion as interfaces get wrapped in different layers (Metered, Caching, Logging) so all the current hierarchy to create stores with Kafka Streams DSL will have to be duplicated.
  • Initially it was considered to have additional parameter on all readOnlyStore methods e.g. Store#fetch(keyFrom, keyTo, timeFrom, timeTo, ReadDirection.FORWARD|BACKWARD), but has been declines as passing arguments in inverse is more intuitive. As this could cause unexpected effects in future versions, a flag has been added to overcome this.
  • Implicit ordering by flipping from and to variables has been discouraged in favor of a more explicit approach based on new interfaces that make explicit the availability of reverse and backward fetch operations.
  • No labels