[
https://issues.apache.org/jira/browse/HBASE-8762?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13687333#comment-13687333
]
Jason Bray commented on HBASE-8762:
-----------------------------------
Thanks for taking a look [[email protected]].
Regarding the javadoc of HTableInterface: I'm happy to change it, but by
bringing it up in this jira I may have confused the issue. To be clear, the
javadoc seems incorrect with or without the patch I've provided. The
RetriesExhaustedWithDetailsException will be thrown and ascend the stack
preventing the Result[] from being returned at all. The javadoc change would
simply be to have the statement read "If there are any failures even after
retries an exception will be thrown.".
If we're in agreement, I can create a second jira with just that javadoc change
or amend this patch to include both changes - whichever is most appropriate.
> Performance/operational penalty when calling HTable.get with a list of one Get
> ------------------------------------------------------------------------------
>
> Key: HBASE-8762
> URL: https://issues.apache.org/jira/browse/HBASE-8762
> Project: HBase
> Issue Type: Bug
> Components: Client
> Reporter: Jason Bray
> Priority: Minor
> Attachments: HBASE-8762.patch
>
>
> There are two implications to calling HTable.get with a list of one Get.
> 1. The overhead of processBatch is paid unnecessarily, which is not
> insignificant.
> 2. The get requests show up as a 'multi' when reviewing RPC handlers, when
> the request should just be a single Get. It seems likely that there are
> other places in logs/ui it shows up as a multi as well.
> To give some context to the overhead, here are some timings performed by a
> member of our team:
> In a very simple test, of reading the same key 100 times, taking the time it
> took, and then repeating this 10 times (1000 total gets), the times are as
> follows (excluding the actual first iteration as there was considerable HBase
> warm-up times on the JVM for establishing connections):
> ||Iteration||Batch (in ms)||Single (in ms)||
> |1|2255|815|
> |2|1545|823|
> |3|1427|742|
> |4|1451|721|
> |5|1480|775|
> |6|1379|735|
> |7|1657|775|
> |8|1392|804|
> While I can see the argument that callers should use the single Get method
> signature, the cost implications are somewhat surprising and it's very easy
> to be smart in this case. We simply need to have HTable.get(List<Get>)
> delegate to HTable.get(<Get>) if the list has one Get.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
