[
https://issues.apache.org/jira/browse/HBASE-8942?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13813157#comment-13813157
]
Amitanand Aiyer commented on HBASE-8942:
----------------------------------------
Seems like the testcase is doing some append operations. This is not available
on 0.89, so unable to port the test back.
Will probably just focus on the open source trunk failures.
> DFS errors during a read operation (get/scan), may cause write outliers
> -----------------------------------------------------------------------
>
> Key: HBASE-8942
> URL: https://issues.apache.org/jira/browse/HBASE-8942
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.89-fb, 0.95.2
> Reporter: Amitanand Aiyer
> Assignee: Amitanand Aiyer
> Priority: Minor
> Fix For: 0.89-fb, 0.98.0, 0.96.1, 0.94.14
>
> Attachments: 8942.096.txt, HBase-8942.txt
>
>
> This is a similar issue as discussed in HBASE-8228
> 1) A scanner holds the Store.ReadLock() while opening the store files ...
> encounters errors. Thus, takes a long time to finish.
> 2) A flush is completed, in the mean while. It needs the write lock to
> commit(), and update scanners. Hence ends up waiting.
> 3+) All Puts (and also Gets) to the CF, which will need a read lock, will
> have to wait for 1) and 2) to complete. Thus blocking updates to the system
> for the DFS timeout.
> Fix:
> Open Store files outside the read lock. getScanners() already tries to do
> this optimisation. However, Store.getScanner() which calls this functions
> through the StoreScanner constructor, redundantly tries to grab the readLock.
> Causing the readLock to be held while the storeFiles are being opened, and
> seeked.
> We should get rid of the readLock() in Store.getScanner(). This is not
> required. The constructor for StoreScanner calls getScanners(xxx, xxx, xxx).
> This has the required locking already.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
