Ron Cai created BEAM-6809:
-----------------------------
Summary: spark.streaming.backpressure not work properly
Key: BEAM-6809
URL: https://issues.apache.org/jira/browse/BEAM-6809
Project: Beam
Issue Type: Bug
Components: runner-spark
Reporter: Ron Cai
When enable configuration of "spark.streaming.backpressure.enabled", the rate
controller will work and return the maxNumRecords in SourceDStream.
computeReadMaxRecords() when create MicrobatchSource instances for each batch
Refer to source file:
beam\runners\spark\src\main\java\org\apache\beam\runners\spark\io\SourceDStream.java
{code:java}
private long computeReadMaxRecords() {
if (boundMaxRecords > 0) {
LOG.info(
"Max records per batch has been set to {}, as configured in the
PipelineOptions.",
boundMaxRecords);
return boundMaxRecords;
} else {
final scala.Option<Long> rateControlledMax = rateControlledMaxRecords();
if (rateControlledMax.isDefined()) {
LOG.info(
"Max records per batch has been set to {}, as advised by the rate controller.",
rateControlledMax.get());
return rateControlledMax.get();
} else {
LOG.info(
"Max records per batch has not been limited by neither configuration "
+ "nor the rate controller, and will remain unlimited for the current batch "
+ "({}).",
Long.MAX_VALUE);
return Long.MAX_VALUE;
}
}
}
{code}
But in MicrobatchSource, there is a cache of reader, the key of the cache is
based on MicrobatchSource
{code:java}
private static volatile Cache<MicrobatchSource<?, ?>, Source.Reader<?>>
readerCache;
@SuppressWarnings("unchecked")
public Source.Reader<T> getOrCreateReader(
final PipelineOptions options, final CheckpointMarkT checkpointMark) throws
IOException {
try {
initReaderCache((long) readerCacheInterval);
return (Source.Reader<T>) readerCache.get(this, new ReaderLoader(options,
checkpointMark));
} catch (final ExecutionException e) {
throw new RuntimeException("Failed to get or create reader", e);
}
} {code}
When comparing two instances of MicrobatchSource, it only check the sourceId
and splitId.
{code:java}
@Override
public boolean equals(Object o) {
if (this == o) {
return true;
}
if (!(o instanceof MicrobatchSource)) {
return false;
}
MicrobatchSource<?, ?> that = (MicrobatchSource<?, ?>) o;
if (sourceId != that.sourceId) {
return false;
}
return splitId == that.splitId;
}
{code}
The reader cache expiration is defined in SourceDStream.java
{code:java}
// Reader cache expiration interval. 50% of batch interval is added to
accommodate latency.
this.readerCacheInterval=1.5*sparkOptions.getBatchIntervalMillis(); {code}
It means the reader cache only expire when the processing time of batch longer
than 1.5 batch interval for the two instances of MicrobatchSource are equal
even the value of maxNumRecords changed. It would make the backpress behavior
work strange, it only update reader when the processing time of batch > 1.5
batchInterval.
I suggest to add the maxNumRecords in the equals() of MicrobatchSource.
{code:java}
@Override
public boolean equals(Object o) {
if (this == o) {
return true;
}
if (!(o instanceof MicrobatchSource)) {
return false;
}
MicrobatchSource<?, ?> that = (MicrobatchSource<?, ?>) o;
if (sourceId != that.sourceId) {
return false;
}
if (maxNumRecords != that.maxNumRecords) {
return false;
}
return splitId == that.splitId;
}
{code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
