[
https://issues.apache.org/jira/browse/BOOKKEEPER-795?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ivan Kelly updated BOOKKEEPER-795:
----------------------------------
Attachment: TEST-org.apache.bookkeeper.client.LedgerCloseTest.xml
0001-Demonstrate-race-condition.patch
> Race condition causes writes to hang if ledger is fences
> --------------------------------------------------------
>
> Key: BOOKKEEPER-795
> URL: https://issues.apache.org/jira/browse/BOOKKEEPER-795
> Project: Bookkeeper
> Issue Type: Bug
> Reporter: Ivan Kelly
> Priority: Blocker
> Fix For: 4.4.0
>
> Attachments: 0001-Demonstrate-race-condition.patch,
> TEST-org.apache.bookkeeper.client.LedgerCloseTest.xml
>
>
> If a ledger is fenced while the write is still writing to it, some of the
> writes will fail to ever complete.
> I've attached the log of this happening along with a test case that will
> trigger the behaviour.
> What appears to be happening is that when the fence occurs, the first write
> after the fence gets an unrecoverable error, so tries to close the ledger.
> Closing the ledger sets the closed flag on the ledger metadata, and tries to
> write it, which fails as the metadata in zookeeper was modified by the
> fencing operation, so the close op fails, resets the closed status for a
> moment, a write operation gets through, which then fails with a fencing
> error, so we try to close the ledger, but the other close operation has since
> closed the ledger in our metadata, so nothing happens, and the write hangs
> forever.
> There's a number of issues here, but foremost, the ledger metadata that the
> handle is using should only ever represent what is actually in zookeeper.
> Having various parts of the code flipping bits just explodes the state space.
> The LedgerMetadata object itself should be immutable, and should only be
> modified, as a local variable, using a builder, before writing to zookeeper.
> Only when the zookeeper operation succeeds should we update the reference
> which LedgerHandle has access to.
> There's also a problem in how we handle pendingaddops when we close. Really
> it shouldn't be possible for a write op to get through after a closure, but
> we should be defensive here and error out anything that has gotten through,
> adding a big old log message to alert us that this cases that shouldn't
> happen, is happening.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
