[
https://issues.apache.org/jira/browse/HADOOP-18184?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17754153#comment-17754153
]
ASF GitHub Bot commented on HADOOP-18184:
-----------------------------------------
steveloughran commented on code in PR #5832:
URL: https://github.com/apache/hadoop/pull/5832#discussion_r1293641249
##########
hadoop-tools/hadoop-aws/src/main/java/org/apache/hadoop/fs/s3a/prefetch/S3ACachingInputStream.java:
##########
@@ -76,36 +79,75 @@ public S3ACachingInputStream(
S3AInputStreamStatistics streamStatistics,
Configuration conf,
LocalDirAllocator localDirAllocator) {
- super(context, s3Attributes, client, streamStatistics);
- this.numBlocksToPrefetch = this.getContext().getPrefetchBlockCount();
- int bufferPoolSize = this.numBlocksToPrefetch + 1;
- this.blockManager = this.createBlockManager(
- this.getContext().getFuturePool(),
- this.getReader(),
- this.getBlockData(),
- bufferPoolSize,
- conf,
- localDirAllocator);
+ super(context, s3Attributes, client, streamStatistics);
+ this.conf = conf;
+ this.localDirAllocator = localDirAllocator;
+ this.numBlocksToPrefetch = getContext().getPrefetchBlockCount();
+ demandCreateBlockManager();
int fileSize = (int) s3Attributes.getLen();
LOG.debug("Created caching input stream for {} (size = {})",
this.getName(),
fileSize);
+ streamStatistics.setPrefetchState(numBlocksToPrefetch > 0,
Review Comment:
set to true. i am thinking about a "no prefetch" option where it's only on
demand/vector io; but can tune that if/when implemented
> s3a prefetching stream to support unbuffer()
> --------------------------------------------
>
> Key: HADOOP-18184
> URL: https://issues.apache.org/jira/browse/HADOOP-18184
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/s3
> Affects Versions: 3.4.0
> Reporter: Steve Loughran
> Assignee: Steve Loughran
> Priority: Minor
> Labels: pull-request-available
>
> Apache Impala uses unbuffer() to free up all client side resources held by a
> stream, so allowing it to have a map of available (path -> stream) objects,
> retained across queries.
> This saves on having to reopen the files, with the cost of HEAD checks etc.
> S3AInputStream just closes its http connection. here there is a lot more
> state to discard, but all memory and file storage must be freed.
> until this done, ITestS3AContractUnbuffer must skip when the prefetch stream
> is used.
> its notable that the other tests don't fail, even though the stream doesn't
> implement the interface; the graceful degradation handles that. it should
> fail if the test xml resource says the stream does it, but that the stream
> capabilities say it doesn't.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
