You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@flink.apache.org by "Greg Hogan (JIRA)" <ji...@apache.org> on 2016/02/10 20:26:18 UTC
[jira] [Commented] (FLINK-3385) Fix outer join skipping unprobed
partitions
[ https://issues.apache.org/jira/browse/FLINK-3385?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15141490#comment-15141490 ]
Greg Hogan commented on FLINK-3385:
-----------------------------------
Here is the simple function I was using to isolate this case, and constraining available memory with -Xmx200m. What is the best way to force spilling in an automated test?
{code}
public void testBrokenOuterJoin() throws Exception {
final int NUMBER_OF_ELEMENTS = 2 * 1000 * 1000;
final ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(1);
DataSet<Long> ds1 = env
.fromParallelCollection(new NumberSequenceIterator(0, NUMBER_OF_ELEMENTS - 1), Long.class);
DataSet<Long> ds2 = ds1
.filter(new FilterFunction<Long>() {
@Override
public boolean filter(Long value) throws Exception {
return false;
}
});
DataSet<Long> joinDs = ds1
.leftOuterJoin(ds2, JoinHint.REPARTITION_HASH_FIRST)
.where("*")
.equalTo("*")
.with(new JoinFunction<Long, Long, Long>() {
@Override
public Long join(Long first, Long second) throws Exception {
return (first == null) ? second : first;
}
});
// List<Long> result = joinDs
// .collect();
// Collections.sort(result);
long count = joinDs
.count();
assertEquals(NUMBER_OF_ELEMENTS, count);
}
{code}
> Fix outer join skipping unprobed partitions
> -------------------------------------------
>
> Key: FLINK-3385
> URL: https://issues.apache.org/jira/browse/FLINK-3385
> Project: Flink
> Issue Type: Bug
> Components: Distributed Runtime
> Reporter: Greg Hogan
> Priority: Critical
> Fix For: 1.0.0
>
>
> {{MutableHashTable.nextRecord}} performs three steps for a build-side outer join:
> {code}
> public boolean nextRecord() throws IOException {
> if (buildSideOuterJoin) {
> return processProbeIter() || processUnmatchedBuildIter() || prepareNextPartition();
> } else {
> return processProbeIter() || prepareNextPartition();
> }
> }
> {code}
> {{MutableHashTable.processUnmatchedBuildIter}} eventually calls through to {{MutableHashTable.moveToNextBucket}} which is unable to process spilled partitions:
> {code}
> if (p.isInMemory()) {
> ...
> } else {
> return false;
> }
> {code}
> {{MutableHashTable.prepareNextPartition}} calls {{HashPartition.finalizeProbePhase}} which only spills the partition (to be read and processed in the next instantiation of {{MutableHashTable}}) if probe-side records were spilled. In an equi-join this is fine but with an outer join the unmatched build-side records must still be retained (though no further probing is necessary, so could this be short-circuited when loaded by the next {{MutableHashTable}}?).
> {code}
> if (isInMemory()) {
> ...
> }
> else if (this.probeSideRecordCounter == 0) {
> // partition is empty, no spilled buffers
> // return the memory buffer
> freeMemory.add(this.probeSideBuffer.getCurrentSegment());
> // delete the spill files
> this.probeSideChannel.close();
> this.buildSideChannel.deleteChannel();
> this.probeSideChannel.deleteChannel();
> return 0;
> }
> else {
> // flush the last probe side buffer and register this partition as pending
> this.probeSideBuffer.close();
> this.probeSideChannel.close();
> spilledPartitions.add(this);
> return 1;
> }
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)