Examples with SnapshotRecord com.hazelcast.jet.impl.execution.SnapshotRecord used on opensource projects

Example 1 with SnapshotRecord

use of com.hazelcast.jet.impl.execution.SnapshotRecord in project hazelcast-jet by hazelcast.

the class SnapshotRepository method registerSnapshot.

/**
 * Registers a new snapshot. Returns the ID for the registered snapshot
 */
long registerSnapshot(long jobId, Collection<String> vertexNames) {
    IMapJet<Long, Object> snapshots = getSnapshotMap(jobId);
    SnapshotRecord record;
    do {
        long nextSnapshotId = generateNextSnapshotId(snapshots);
        record = new SnapshotRecord(jobId, nextSnapshotId, vertexNames);
    } while (snapshots.putIfAbsent(record.snapshotId(), record) != null);
    return record.snapshotId();
}

Example 2 with SnapshotRecord

use of com.hazelcast.jet.impl.execution.SnapshotRecord in project hazelcast-jet by hazelcast.

the class SnapshotRepository method deleteSingleSnapshot.

/**
 * Delete a single snapshot for a given job if it exists
 */
void deleteSingleSnapshot(long jobId, Long snapshotId) {
    final IMapJet<Long, SnapshotRecord> snapshotMap = getSnapshotMap(jobId);
    SnapshotRecord record = snapshotMap.get(snapshotId);
    if (record != null) {
        deleteSnapshot(snapshotMap, record);
    }
}

Example 3 with SnapshotRecord

use of com.hazelcast.jet.impl.execution.SnapshotRecord in project hazelcast-jet by hazelcast.

the class SnapshotRepository method deleteAllSnapshots.

/**
 * Delete all snapshots for a given job
 */
void deleteAllSnapshots(long jobId) {
    final IMapJet<Long, SnapshotRecord> snapshotMap = getSnapshotMap(jobId);
    Predicate predicate = e -> !e.getKey().equals(LATEST_STARTED_SNAPSHOT_ID_KEY);
    for (Entry<Long, SnapshotRecord> entry : snapshotMap.entrySet(predicate)) {
        deleteSnapshotData(entry.getValue());
    }
    logFine(logger, "Deleted all snapshots for job %s", idToString(jobId));
    snapshotMap.destroy();
}

Example 4 with SnapshotRecord

use of com.hazelcast.jet.impl.execution.SnapshotRecord in project hazelcast-jet by hazelcast.

the class JobRestartWithSnapshotTest method when_snapshotStartedBeforeExecution_then_firstSnapshotIsSuccessful.

@Test
public void when_snapshotStartedBeforeExecution_then_firstSnapshotIsSuccessful() throws Exception {
    // instance1 is always coordinator
    // delay ExecuteOperation so that snapshot is started before execution is started on the worker member
    delayOperationsFrom(hz(instance1), JetInitDataSerializerHook.FACTORY_ID, singletonList(JetInitDataSerializerHook.START_EXECUTION_OP));
    DAG dag = new DAG();
    dag.newVertex("p", FirstSnapshotProcessor::new).localParallelism(1);
    JobConfig config = new JobConfig();
    config.setProcessingGuarantee(ProcessingGuarantee.EXACTLY_ONCE);
    config.setSnapshotIntervalMillis(0);
    Job job = instance1.newJob(dag, config);
    // the first snapshot should succeed
    assertTrueEventually(() -> {
        IMapJet<Long, SnapshotRecord> records = getSnapshotsMap(job);
        SnapshotRecord record = records.get(0L);
        assertNotNull("no record found for snapshot 0", record);
        assertTrue("snapshot was not successful", record.isSuccessful());
    }, 30);
}

Example 5 with SnapshotRecord

use of com.hazelcast.jet.impl.execution.SnapshotRecord in project hazelcast-jet by hazelcast.

the class JobRestartWithSnapshotTest method when_nodeDown_then_jobRestartsFromSnapshot.

public void when_nodeDown_then_jobRestartsFromSnapshot(boolean twoStage) throws Exception {
    /* Design of this test:

        It uses random partitioned generator of source events. The events are Map.Entry(partitionId, timestamp).
        For each partition timestamps from 0..elementsInPartition are generated.

        We start the test with two nodes and localParallelism(1) for source. Source instances generate items at
        the same rate of 10 per second: this causes one instance to be twice as fast as the other in terms of
        timestamp. The source processor saves partition offsets similarly to how streamKafka() and streamMap()
        do.

        After some time we shut down one instance. The job restarts from snapshot and all partitions are restored
        to single source processor instance. Partition offsets are very different, so the source is written in a way
        that it emits from the most-behind partition in order to not emit late events from more ahead partitions.

        Local parallelism of InsertWatermarkP is also 1 to avoid the edge case when different instances of
        InsertWatermarkP might initialize with first event in different frame and make them start the no-gap
        emission from different WM, which might cause the SlidingWindowP downstream to miss some of the
        first windows.

        The sink writes to an IMap which is an idempotent sink.

        The resulting contents of the sink map are compared to expected value.
         */
    DAG dag = new DAG();
    SlidingWindowPolicy wDef = SlidingWindowPolicy.tumblingWinPolicy(3);
    AggregateOperation1<Object, LongAccumulator, Long> aggrOp = counting();
    IMap<List<Long>, Long> result = instance1.getMap("result");
    result.clear();
    SequencesInPartitionsMetaSupplier sup = new SequencesInPartitionsMetaSupplier(3, 180);
    Vertex generator = dag.newVertex("generator", throttle(sup, 30)).localParallelism(1);
    Vertex insWm = dag.newVertex("insWm", insertWatermarksP(wmGenParams(entry -> ((Entry<Integer, Integer>) entry).getValue(), limitingLag(0), emitByFrame(wDef), -1))).localParallelism(1);
    Vertex map = dag.newVertex("map", mapP((TimestampedEntry e) -> entry(asList(e.getTimestamp(), (long) (int) e.getKey()), e.getValue())));
    Vertex writeMap = dag.newVertex("writeMap", SinkProcessors.writeMapP("result"));
    if (twoStage) {
        Vertex aggregateStage1 = dag.newVertex("aggregateStage1", Processors.accumulateByFrameP(singletonList((DistributedFunction<? super Object, ?>) t -> ((Entry<Integer, Integer>) t).getKey()), singletonList(t1 -> ((Entry<Integer, Integer>) t1).getValue()), TimestampKind.EVENT, wDef, aggrOp.withFinishFn(identity())));
        Vertex aggregateStage2 = dag.newVertex("aggregateStage2", combineToSlidingWindowP(wDef, aggrOp, TimestampedEntry::new));
        dag.edge(between(insWm, aggregateStage1).partitioned(entryKey())).edge(between(aggregateStage1, aggregateStage2).distributed().partitioned(entryKey())).edge(between(aggregateStage2, map));
    } else {
        Vertex aggregate = dag.newVertex("aggregate", Processors.aggregateToSlidingWindowP(singletonList((DistributedFunction<Object, Integer>) t -> ((Entry<Integer, Integer>) t).getKey()), singletonList(t1 -> ((Entry<Integer, Integer>) t1).getValue()), TimestampKind.EVENT, wDef, aggrOp, TimestampedEntry::new));
        dag.edge(between(insWm, aggregate).distributed().partitioned(entryKey())).edge(between(aggregate, map));
    }
    dag.edge(between(generator, insWm)).edge(between(map, writeMap));
    JobConfig config = new JobConfig();
    config.setProcessingGuarantee(ProcessingGuarantee.EXACTLY_ONCE);
    config.setSnapshotIntervalMillis(1200);
    Job job = instance1.newJob(dag, config);
    SnapshotRepository snapshotRepository = new SnapshotRepository(instance1);
    int timeout = (int) (MILLISECONDS.toSeconds(config.getSnapshotIntervalMillis()) + 2);
    // wait until we have at least one snapshot
    IMapJet<Long, Object> snapshotsMap = snapshotRepository.getSnapshotMap(job.getId());
    assertTrueEventually(() -> assertTrue("No snapshot produced", snapshotsMap.entrySet().stream().anyMatch(en -> en.getValue() instanceof SnapshotRecord && ((SnapshotRecord) en.getValue()).isSuccessful())), timeout);
    waitForNextSnapshot(snapshotsMap, timeout);
    // wait a little more to emit something, so that it will be overwritten in the sink map
    Thread.sleep(300);
    instance2.shutdown();
    // Now the job should detect member shutdown and restart from snapshot.
    // Let's wait until the next snapshot appears.
    waitForNextSnapshot(snapshotsMap, (int) (MILLISECONDS.toSeconds(config.getSnapshotIntervalMillis()) + 10));
    waitForNextSnapshot(snapshotsMap, timeout);
    job.join();
    // compute expected result
    Map<List<Long>, Long> expectedMap = new HashMap<>();
    for (long partition = 0; partition < sup.numPartitions; partition++) {
        long cnt = 0;
        for (long value = 1; value <= sup.elementsInPartition; value++) {
            cnt++;
            if (value % wDef.frameSize() == 0) {
                expectedMap.put(asList(value, partition), cnt);
                cnt = 0;
            }
        }
        if (cnt > 0) {
            expectedMap.put(asList(wDef.higherFrameTs(sup.elementsInPartition - 1), partition), cnt);
        }
    }
    // check expected result
    if (!expectedMap.equals(result)) {
        System.out.println("All expected entries: " + expectedMap.entrySet().stream().map(Object::toString).collect(joining(", ")));
        System.out.println("All actual entries: " + result.entrySet().stream().map(Object::toString).collect(joining(", ")));
        System.out.println("Non-received expected items: " + expectedMap.keySet().stream().filter(key -> !result.containsKey(key)).map(Object::toString).collect(joining(", ")));
        System.out.println("Received non-expected items: " + result.entrySet().stream().filter(entry -> !expectedMap.containsKey(entry.getKey())).map(Object::toString).collect(joining(", ")));
        System.out.println("Different keys: ");
        for (Entry<List<Long>, Long> rEntry : result.entrySet()) {
            Long expectedValue = expectedMap.get(rEntry.getKey());
            if (expectedValue != null && !expectedValue.equals(rEntry.getValue())) {
                System.out.println("key: " + rEntry.getKey() + ", expected value: " + expectedValue + ", actual value: " + rEntry.getValue());
            }
        }
        System.out.println("-- end of different keys");
        assertEquals(expectedMap, new HashMap<>(result));
    }
    assertTrue("Snapshots map not empty after job finished", snapshotsMap.isEmpty());
}