[
https://issues.apache.org/jira/browse/CASSANDRA-7000?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Benedict updated CASSANDRA-7000:
--------------------------------
Attachment: 7000.supplement.txt
Not sure if we want to tack this onto the same ticket, but the same basic
problem seems to occur elsewhere as well: some referrers of sstablereader do
not acquire a reference prior to accessing resources that will be closed when
ref count hits zero. I found specifically CFS.estimatedKeysForRange,
CFS.keySamples, couldn't see any others.
This latter is the cause of the currently failing unit test for 2.1
> Assertion in SSTableReader during repair.
> -----------------------------------------
>
> Key: CASSANDRA-7000
> URL: https://issues.apache.org/jira/browse/CASSANDRA-7000
> Project: Cassandra
> Issue Type: Bug
> Reporter: Ben Chan
> Assignee: Ben Chan
> Attachments: 7000-2.1-v2.txt, 7000.supplement.txt,
> sstablereader-assertion-bisect-helper,
> sstablereader-assertion-bisect-helper-v2, sstablereader-assertion.patch
>
>
> I ran a {{git bisect run}} using the attached bisect script. Repro code:
> {noformat}
> # 5dfe241: trunk as of my git bisect run
> # 345772d: empirically determined "good" commit.
> git bisect start 5dfe241 345772d
> git bisect run ./sstablereader-assertion-bisect-helper-v2
> {noformat}
> The first failing commit is 5ebadc1 (first parent of {{refs/bisect/bad}}).
> Prior to 5ebadc1, SSTableReader#close() never checked reference count. After
> 5ebadc1, there was an assertion for {{references.get() == 0}}. However, since
> the reference count is initialized to 1, a SSTableReader#close() was always
> guaranteed to either throw an AssertionError or to be a second call to
> SSTableReader#tidy() on the same SSTableReader.
> The attached patch chooses an in-between behavior. It requires the reference
> count to match the initialization value of 1 for SSTableReader#close(), and
> the same behavior as 5ebadc1 otherwise.
> This allows repair to finish successfully, but I'm not 100% certain what the
> desired behavior is for SSTableReader#close(). Should it close without regard
> to reference count, as it did pre-5ebadc1?
> Edit: accidentally uploaded a flawed version of
> {{sstablereader-assertion-bisect-helper}} (doesn't work out-of-the-box with
> {{git bisect}}).
--
This message was sent by Atlassian JIRA
(v6.2#6252)
