use of org.apache.beam.model.pipeline.v1.RunnerApi.Pipeline in project beam by apache.
the class GreedyStageFuser method anyInputsSideInputs.
private static boolean anyInputsSideInputs(PTransformNode consumer, QueryablePipeline pipeline) {
for (String inputPCollectionId : consumer.getTransform().getInputsMap().values()) {
RunnerApi.PCollection pCollection = pipeline.getComponents().getPcollectionsMap().get(inputPCollectionId);
PCollectionNode pCollectionNode = PipelineNode.pCollection(inputPCollectionId, pCollection);
if (!pipeline.getSingletonConsumers(pCollectionNode).isEmpty()) {
return true;
}
}
return false;
}
use of org.apache.beam.model.pipeline.v1.RunnerApi.Pipeline in project beam by apache.
the class GreedyStageFuser method forGrpcPortRead.
/**
* Returns an {@link ExecutableStage} where the initial {@link PTransformNode PTransform} is a
* Remote gRPC Port Read, reading elements from the materialized {@link PCollectionNode
* PCollection}.
*
* @param initialNodes the initial set of sibling transforms to fuse into this node. All of the
* transforms must consume the {@code inputPCollection} on a per-element basis, and must all
* be mutually compatible.
*/
public static ExecutableStage forGrpcPortRead(QueryablePipeline pipeline, PCollectionNode inputPCollection, Set<PTransformNode> initialNodes) {
checkArgument(!initialNodes.isEmpty(), "%s must contain at least one %s.", GreedyStageFuser.class.getSimpleName(), PTransformNode.class.getSimpleName());
// Choose the environment from an arbitrary node. The initial nodes may not be empty for this
// subgraph to make any sense, there has to be at least one processor node
// (otherwise the stage is gRPC Read -> gRPC Write, which doesn't do anything).
Environment environment = getStageEnvironment(pipeline, initialNodes);
ImmutableSet.Builder<PTransformNode> fusedTransforms = ImmutableSet.builder();
fusedTransforms.addAll(initialNodes);
Set<SideInputReference> sideInputs = new LinkedHashSet<>();
Set<UserStateReference> userStates = new LinkedHashSet<>();
Set<TimerReference> timers = new LinkedHashSet<>();
Set<PCollectionNode> fusedCollections = new LinkedHashSet<>();
Set<PCollectionNode> materializedPCollections = new LinkedHashSet<>();
Queue<PCollectionNode> fusionCandidates = new ArrayDeque<>();
for (PTransformNode initialConsumer : initialNodes) {
fusionCandidates.addAll(pipeline.getOutputPCollections(initialConsumer));
sideInputs.addAll(pipeline.getSideInputs(initialConsumer));
userStates.addAll(pipeline.getUserStates(initialConsumer));
timers.addAll(pipeline.getTimers(initialConsumer));
}
while (!fusionCandidates.isEmpty()) {
PCollectionNode candidate = fusionCandidates.poll();
if (fusedCollections.contains(candidate) || materializedPCollections.contains(candidate)) {
// This should generally mean we get to a Flatten via multiple paths through the graph and
// we've already determined what to do with the output.
LOG.debug("Skipping fusion candidate {} because it is {} in this {}", candidate, fusedCollections.contains(candidate) ? "fused" : "materialized", ExecutableStage.class.getSimpleName());
continue;
}
PCollectionFusibility fusibility = canFuse(pipeline, candidate, environment, fusedCollections);
switch(fusibility) {
case MATERIALIZE:
materializedPCollections.add(candidate);
break;
case FUSE:
// All of the consumers of the candidate PCollection can be fused into this stage. Do so.
fusedCollections.add(candidate);
fusedTransforms.addAll(pipeline.getPerElementConsumers(candidate));
for (PTransformNode consumer : pipeline.getPerElementConsumers(candidate)) {
// The outputs of every transform fused into this stage must be either materialized or
// themselves fused away, so add them to the set of candidates.
fusionCandidates.addAll(pipeline.getOutputPCollections(consumer));
sideInputs.addAll(pipeline.getSideInputs(consumer));
}
break;
default:
throw new IllegalStateException(String.format("Unknown type of %s %s", PCollectionFusibility.class.getSimpleName(), fusibility));
}
}
return ImmutableExecutableStage.ofFullComponents(pipeline.getComponents(), environment, inputPCollection, sideInputs, userStates, timers, fusedTransforms.build(), materializedPCollections, DEFAULT_WIRE_CODER_SETTINGS);
}
use of org.apache.beam.model.pipeline.v1.RunnerApi.Pipeline in project beam by apache.
the class PipelineValidator method validate.
public static void validate(RunnerApi.Pipeline p) {
Components components = p.getComponents();
for (String transformId : p.getRootTransformIdsList()) {
checkArgument(components.containsTransforms(transformId), "Root transform id %s is unknown", transformId);
}
validateComponents("pipeline", components, ImmutableSet.copyOf(p.getRequirementsList()));
}
use of org.apache.beam.model.pipeline.v1.RunnerApi.Pipeline in project beam by apache.
the class GreedyPipelineFuser method fusePipeline.
/**
* Fuses a {@link Pipeline} into a collection of {@link ExecutableStage}.
*
* <p>The input is the initial collection of siblings sets which will be fused into {@link
* ExecutableStage stages}. A sibling in this context represents a pair of (PCollection,
* PTransform), where the PTransform consumes input elements on a per-element basis from the
* PCollection, represented by a {@link CollectionConsumer}. A sibling set is a collection of
* siblings which can execute within a single {@link ExecutableStage}, determined by {@link
* GreedyPCollectionFusers#isCompatible(PTransformNode, PTransformNode, QueryablePipeline)}.
*
* <p>While a pending sibling set exists:
*
* <ul>
* <li>Retrieve a pending sibling set from the front of the queue.
* <li>If the pending sibling set has already been created, continue. Each materialized {@link
* PTransformNode} can be consumed by any number of {@link ExecutableStage stages}, but each
* {@link PTransformNode} may only be present in a single stage rooted at a single {@link
* PCollectionNode}, otherwise it will process elements of that {@link PCollectionNode}
* multiple times.
* <li>Create a {@link GreedyStageFuser} with those siblings as the initial consuming transforms
* of the stage
* <li>For each materialized {@link PCollectionNode}, find all of the descendant in-environment
* consumers. See {@link #getDescendantConsumers(PCollectionNode)} for details.
* <li>Construct all of the sibling sets from the descendant in-environment consumers, and add
* them to the queue of sibling sets.
* </ul>
*/
private FusedPipeline fusePipeline(Collection<PTransformNode> initialUnfusedTransforms, NavigableSet<NavigableSet<CollectionConsumer>> initialConsumers, Set<String> requirements) {
Map<CollectionConsumer, ExecutableStage> consumedCollectionsAndTransforms = new HashMap<>();
Set<ExecutableStage> stages = new LinkedHashSet<>();
Set<PTransformNode> unfusedTransforms = new LinkedHashSet<>(initialUnfusedTransforms);
Queue<Set<CollectionConsumer>> pendingSiblingSets = new ArrayDeque<>(initialConsumers);
while (!pendingSiblingSets.isEmpty()) {
// Only introduce new PCollection consumers. Not performing this introduces potential
// duplicate paths through the pipeline.
Set<CollectionConsumer> candidateSiblings = pendingSiblingSets.poll();
Set<CollectionConsumer> siblingSet = Sets.difference(candidateSiblings, consumedCollectionsAndTransforms.keySet());
checkState(siblingSet.equals(candidateSiblings) || siblingSet.isEmpty(), "Inconsistent collection of siblings reported for a %s. Initial attempt missed %s", PCollectionNode.class.getSimpleName(), siblingSet);
if (siblingSet.isEmpty()) {
LOG.debug("Filtered out duplicate stage root {}", candidateSiblings);
continue;
}
// Create the stage with these siblings as the initial consuming transforms
ExecutableStage stage = fuseSiblings(siblingSet);
// don't place them in multiple stages.
for (CollectionConsumer sibling : siblingSet) {
consumedCollectionsAndTransforms.put(sibling, stage);
}
stages.add(stage);
for (PCollectionNode materializedOutput : stage.getOutputPCollections()) {
// Get all of the descendant consumers of each materialized PCollection, and add them to the
// queue of pending siblings.
DescendantConsumers descendantConsumers = getDescendantConsumers(materializedOutput);
unfusedTransforms.addAll(descendantConsumers.getUnfusedNodes());
NavigableSet<NavigableSet<CollectionConsumer>> siblings = groupSiblings(descendantConsumers.getFusibleConsumers());
pendingSiblingSets.addAll(siblings);
}
}
// TODO: Figure out where to store this.
DeduplicationResult deduplicated = OutputDeduplicator.ensureSingleProducer(pipeline, stages, unfusedTransforms);
// as can compatible producers/consumers if a PCollection is only materialized once.
return FusedPipeline.of(deduplicated.getDeduplicatedComponents(), stages.stream().map(stage -> deduplicated.getDeduplicatedStages().getOrDefault(stage, stage)).map(GreedyPipelineFuser::sanitizeDanglingPTransformInputs).collect(Collectors.toSet()), Sets.union(deduplicated.getIntroducedTransforms(), unfusedTransforms.stream().map(transform -> deduplicated.getDeduplicatedTransforms().getOrDefault(transform.getId(), transform)).collect(Collectors.toSet())), requirements);
}
use of org.apache.beam.model.pipeline.v1.RunnerApi.Pipeline in project beam by apache.
the class OutputDeduplicatorTest method duplicateOverStages.
@Test
public void duplicateOverStages() {
/* When multiple stages and a runner-executed transform produce a PCollection, all should be
* replaced with synthetic flattens.
* original graph:
* --> one -> .out \
* red -> .out | -> shared -> .out -> blue -> .out
* --> two -> .out /
*
* fused graph:
* --> [one -> .out -> shared ->] .out
* red -> .out | (shared.out) -> blue -> .out
* --> [two -> .out -> shared ->] .out
*
* deduplicated graph:
* --> [one -> .out -> shared ->] .out:0 \
* red -> .out | -> shared -> .out -> blue ->.out
* --> [two -> .out -> shared ->] .out:1 /
*/
PCollection redOut = PCollection.newBuilder().setUniqueName("red.out").build();
PTransform red = PTransform.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(PTransformTranslation.PAR_DO_TRANSFORM_URN).build()).putOutputs("out", redOut.getUniqueName()).build();
PCollection oneOut = PCollection.newBuilder().setUniqueName("one.out").build();
PTransform one = PTransform.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(PTransformTranslation.PAR_DO_TRANSFORM_URN).build()).putInputs("in", redOut.getUniqueName()).putOutputs("out", oneOut.getUniqueName()).build();
PCollection twoOut = PCollection.newBuilder().setUniqueName("two.out").build();
PTransform two = PTransform.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(PTransformTranslation.PAR_DO_TRANSFORM_URN).build()).putInputs("in", redOut.getUniqueName()).putOutputs("out", twoOut.getUniqueName()).build();
PCollection sharedOut = PCollection.newBuilder().setUniqueName("shared.out").build();
PTransform shared = PTransform.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(PTransformTranslation.PAR_DO_TRANSFORM_URN).build()).putInputs("one", oneOut.getUniqueName()).putInputs("two", twoOut.getUniqueName()).putOutputs("shared", sharedOut.getUniqueName()).build();
PCollection blueOut = PCollection.newBuilder().setUniqueName("blue.out").build();
PTransform blue = PTransform.newBuilder().setSpec(FunctionSpec.newBuilder().setUrn(PTransformTranslation.PAR_DO_TRANSFORM_URN).build()).putInputs("in", sharedOut.getUniqueName()).putOutputs("out", blueOut.getUniqueName()).build();
RunnerApi.Components components = Components.newBuilder().putTransforms("one", one).putPcollections(oneOut.getUniqueName(), oneOut).putTransforms("two", two).putPcollections(twoOut.getUniqueName(), twoOut).putTransforms("shared", shared).putPcollections(sharedOut.getUniqueName(), sharedOut).putTransforms("red", red).putPcollections(redOut.getUniqueName(), redOut).putTransforms("blue", blue).putPcollections(blueOut.getUniqueName(), blueOut).build();
ExecutableStage oneStage = ImmutableExecutableStage.of(components, Environment.getDefaultInstance(), PipelineNode.pCollection(redOut.getUniqueName(), redOut), ImmutableList.of(), ImmutableList.of(), ImmutableList.of(), ImmutableList.of(PipelineNode.pTransform("one", one), PipelineNode.pTransform("shared", shared)), ImmutableList.of(PipelineNode.pCollection(sharedOut.getUniqueName(), sharedOut)), DEFAULT_WIRE_CODER_SETTINGS);
ExecutableStage twoStage = ImmutableExecutableStage.of(components, Environment.getDefaultInstance(), PipelineNode.pCollection(redOut.getUniqueName(), redOut), ImmutableList.of(), ImmutableList.of(), ImmutableList.of(), ImmutableList.of(PipelineNode.pTransform("two", two), PipelineNode.pTransform("shared", shared)), ImmutableList.of(PipelineNode.pCollection(sharedOut.getUniqueName(), sharedOut)), DEFAULT_WIRE_CODER_SETTINGS);
PTransformNode redTransform = PipelineNode.pTransform("red", red);
PTransformNode blueTransform = PipelineNode.pTransform("blue", blue);
QueryablePipeline pipeline = QueryablePipeline.forPrimitivesIn(components);
DeduplicationResult result = OutputDeduplicator.ensureSingleProducer(pipeline, ImmutableList.of(oneStage, twoStage), ImmutableList.of(redTransform, blueTransform));
assertThat(result.getIntroducedTransforms(), hasSize(1));
PTransformNode introduced = getOnlyElement(result.getIntroducedTransforms());
assertThat(introduced.getTransform().getOutputsMap().size(), equalTo(1));
assertThat(getOnlyElement(introduced.getTransform().getOutputsMap().values()), equalTo(sharedOut.getUniqueName()));
assertThat(result.getDeduplicatedComponents().getPcollectionsMap().keySet(), hasItems(introduced.getTransform().getInputsMap().values().toArray(new String[0])));
assertThat(result.getDeduplicatedStages().keySet(), hasSize(2));
List<String> stageOutputs = result.getDeduplicatedStages().values().stream().flatMap(stage -> stage.getOutputPCollections().stream().map(PCollectionNode::getId)).collect(Collectors.toList());
assertThat(stageOutputs, containsInAnyOrder(introduced.getTransform().getInputsMap().values().toArray()));
assertThat(result.getDeduplicatedTransforms().keySet(), empty());
assertThat(result.getDeduplicatedComponents().getPcollectionsMap().keySet(), hasItems(stageOutputs.toArray(new String[0])));
assertThat(result.getDeduplicatedComponents().getTransformsMap(), hasEntry(introduced.getId(), introduced.getTransform()));
}
Aggregations