[
https://issues.apache.org/jira/browse/BEAM-9346?focusedWorklogId=406671&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-406671
]
ASF GitHub Bot logged work on BEAM-9346:
----------------------------------------
Author: ASF GitHub Bot
Created on: 20/Mar/20 00:31
Start Date: 20/Mar/20 00:31
Worklog Time Spent: 10m
Work Description: chamikaramj commented on pull request #11122:
[BEAM-9346] Improve the efficiency of TFRecordIO
URL: https://github.com/apache/beam/pull/11122#discussion_r395391171
##########
File path: sdks/java/core/src/main/java/org/apache/beam/sdk/io/WriteFiles.java
##########
@@ -410,13 +412,44 @@ private GatherResults(Coder<ResultT> resultCoder) {
} else {
// Pass results via a side input rather than reshuffle, because we
need to get an empty
// iterable to finalize if there are no results.
- return input
- .getPipeline()
- .apply(Reify.viewInGlobalWindow(input.apply(View.asList()),
ListCoder.of(resultCoder)));
+ return input.apply("ToList", Combine.globally(new
ToListCombineFn<>()));
Review comment:
Seems like this was applied to all file-based IO not just TFRecordIO as
mentioned in the PR description. Have we done enough experiments to make sure
this won't have unintended adverse performance consequences in batch/streaming
? For example, what if the list does not fit in memory ?
Apologies if this was already discussed in the dev list.
cc: @robertwb @lukecwik @iemejia
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to 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: 406671)
Time Spent: 2h 40m (was: 2.5h)
> TFRecordIO inefficient read from sideinput causing pipeline to be slow
> ----------------------------------------------------------------------
>
> Key: BEAM-9346
> URL: https://issues.apache.org/jira/browse/BEAM-9346
> Project: Beam
> Issue Type: Improvement
> Components: sdk-java-core
> Reporter: Ban Piao
> Assignee: Piotr Szuberski
> Priority: Major
> Labels: dataflow, easyfix, performance
> Fix For: Not applicable
>
> Time Spent: 2h 40m
> Remaining Estimate: 0h
>
> In TFRecordIO, Reify.viewInGlobalWindow(input.apply(View.asList()),
> ListCoder.of(resultCoder)) is an inefficient way of reading large set of side
> input.
> Pipeline can be sped up significantly by combinging the PCollection<ResultT>
> to a single element PCollection<List<ResultT>>.
> Sample code:
>
> https://github.com/apache/beam/blob/master/sdks/java/core/src/main/java/org/apache/beam/sdk/io/WriteFiles.java#L412
> from
> ```
> return input
> .getPipeline()
> .apply(Reify.viewInGlobalWindow(input.apply(View.asList()),
> ListCoder.of(resultCoder)));
> ```
> to
> ```
> return input.apply("ToList", Combine.globally(new ToListCombineFn<>()));
> ```
> where ToListCombineFn is defined as
> ```
> public static class ToListCombineFn<ResultT> extends CombineFn<ResultT,
> List<ResultT>, List<ResultT>> {
> @Override
> public List<ResultT> createAccumulator() {
> return new ArrayList<>();
> }
> @Override
> public List<ResultT> addInput(List<ResultT> mutableAccumulator, ResultT
> input) {
> mutableAccumulator.add(input);
> return mutableAccumulator;
> }
> @Override
> public List<ResultT> mergeAccumulators(Iterable<List<ResultT>>
> accumulators) {
> Iterator<List<ResultT>> iter = accumulators.iterator();
> if (!iter.hasNext()) {
> return new ArrayList<>();
> }
> List<ResultT> merged = iter.next();
> while (iter.hasNext()) {
> merged.addAll(iter.next());
> }
> return merged;
> }
> @Override
> public List<ResultT> extractOutput(List<ResultT> accumulator) {
> return accumulator;
> }
> }
> ```
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
