[
https://issues.apache.org/jira/browse/BEAM-2597?focusedWorklogId=99303&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-99303
]
ASF GitHub Bot logged work on BEAM-2597:
----------------------------------------
Author: ASF GitHub Bot
Created on: 08/May/18 00:35
Start Date: 08/May/18 00:35
Worklog Time Spent: 10m
Work Description: bsidhom commented on a change in pull request #5285:
[BEAM-2597] Flink batch ExecutableStage operator
URL: https://github.com/apache/beam/pull/5285#discussion_r186586311
##########
File path:
runners/flink/src/main/java/org/apache/beam/runners/flink/translation/functions/FlinkExecutableStageFunction.java
##########
@@ -38,36 +53,117 @@
public class FlinkExecutableStageFunction<InputT>
extends RichMapPartitionFunction<WindowedValue<InputT>, RawUnionValue> {
+ // Main constructor fields. All must be Serializable because Flink
distributes Functions to
+ // task managers via java serialization.
+
// The executable stage this function will run.
private final RunnerApi.ExecutableStagePayload stagePayload;
// Pipeline options. Used for provisioning api.
- private final Struct pipelineOptions;
+ private final JobInfo jobInfo;
// Map from PCollection id to the union tag used to represent this
PCollection in the output.
private final Map<String, Integer> outputMap;
+ private final SerializableSupplier<FlinkBundleFactory> bundleFactorySupplier;
+ private final DistributedCachePool.Factory cachePoolFactory;
+ private final FlinkStateRequestHandlerFactory stateHandlerFactory;
+
+ // Worker-local fields. These should only be constructed and consumed on
Flink TaskManagers.
+ private transient RuntimeContext runtimeContext;
+ private transient StateRequestHandler stateRequestHandler;
+ private transient StageBundleFactory stageBundleFactory;
+ private transient AutoCloseable distributedCacheCloser;
public FlinkExecutableStageFunction(
RunnerApi.ExecutableStagePayload stagePayload,
- Struct pipelineOptions,
- Map<String, Integer> outputMap) {
+ JobInfo jobInfo,
+ Map<String, Integer> outputMap,
+ SerializableSupplier<FlinkBundleFactory> bundleFactorySupplier,
+ DistributedCachePool.Factory cachePoolFactory,
+ FlinkStateRequestHandlerFactory stateHandlerFactory) {
this.stagePayload = stagePayload;
- this.pipelineOptions = pipelineOptions;
+ this.jobInfo = jobInfo;
this.outputMap = outputMap;
+ this.bundleFactorySupplier = bundleFactorySupplier;
+ this.cachePoolFactory = cachePoolFactory;
+ this.stateHandlerFactory = stateHandlerFactory;
}
@Override
public void open(Configuration parameters) throws Exception {
- throw new UnsupportedOperationException();
+ ExecutableStage executableStage =
ExecutableStage.fromPayload(stagePayload);
+ runtimeContext = getRuntimeContext();
+ // NOTE: It's safe to reuse the state handler between partitions because
each partition uses the
+ // same backing runtime context and broadcast variables. We use checkState
below to catch errors
+ // in backward-incompatible Flink changes.
+ stateRequestHandler = stateHandlerFactory.forStage(executableStage,
runtimeContext);
+ DistributedCachePool cachePool = cachePoolFactory.forJob(jobInfo.jobId());
+ distributedCacheCloser =
cachePool.addCacheToPool(runtimeContext.getDistributedCache());
+ FlinkBundleFactory flinkBundleFactory = bundleFactorySupplier.get();
+ // TODO: Do we really want this layer of indirection when accessing the
stage bundle factory?
Review comment:
OK. I'll leave the TODO in place as this is as-yet unanswered.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
Issue Time Tracking
-------------------
Worklog Id: (was: 99303)
Time Spent: 2h (was: 1h 50m)
> FlinkRunner ExecutableStage batch operator
> ------------------------------------------
>
> Key: BEAM-2597
> URL: https://issues.apache.org/jira/browse/BEAM-2597
> Project: Beam
> Issue Type: Sub-task
> Components: runner-flink
> Reporter: Kenneth Knowles
> Assignee: Ben Sidhom
> Priority: Major
> Labels: portability
> Time Spent: 2h
> Remaining Estimate: 0h
>
> This operator will execute user code in the context of an SDK harness by
> constructing a ProcessBundleDescriptor from an ExecutableStage (physical
> stage plan) and sending instructions/elements over the control and data
> planes.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
