[
https://issues.apache.org/jira/browse/FLUME-936?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13198388#comment-13198388
]
[email protected] commented on FLUME-936:
-----------------------------------------------------
bq. On 2012-02-01 18:09:27, Prasad Mujumdar wrote:
bq. > The changes overall look fine. A couple of comments,
bq. >
bq. > 1) It should be okay to bump up the queue capacity during before
rollback start, to ensure sufficient space to return the events.
bq. >
bq. > 2) The commit() after put() and rollback() after take() are both
bq. > adding elements to the channel's queue. But they are under different
locks. That means concurrent source and sink could cause queue to reach
capacity in the middle of commit/rollback and one of them would be partially
complete. I guess its better to use same lock for both operations.
bq. >
bq. >
bq. >
Totally right about 2) I've fixed that and will put the patch for that up now
As far as 1) resizing the queue is pretty awkward. Anything we resize up for
rollback would have to be resized back down at some time. If we want to
guarantee that rollbacks succeed, I think it would be better just to reserve
space in the main queue for the contents of every threads take queue, by
maintaining queue.remainingCapacity() >= sumallthreads(takeList.size())
perhaps this should be a separate issue?
- Juhani
-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/3704/#review4750
-----------------------------------------------------------
On 2012-02-01 09:55:19, Juhani Connolly wrote:
bq.
bq. -----------------------------------------------------------
bq. This is an automatically generated e-mail. To reply, visit:
bq. https://reviews.apache.org/r/3704/
bq. -----------------------------------------------------------
bq.
bq. (Updated 2012-02-01 09:55:19)
bq.
bq.
bq. Review request for Flume.
bq.
bq.
bq. Summary
bq. -------
bq.
bq. This is an initial go at fixing the threading issues with memory channel.
bq.
bq. It uses the preliminary work on FLUME-935 and I have included the code
from that.
bq.
bq. The tagging of the events became unnecessary so I dropped that. One thing
that concerns me slightly is how to deal with not having enough space in the
queue to rollback failed takes. One method would be to keep a minimum buffer of
transactionCapacity. Another would be to implement the queue of queues as
suggested in FLUME-889
bq.
bq. Anyway, just putting up this early version to see what people think
bq.
bq.
bq. This addresses bug FLUME-936.
bq. https://issues.apache.org/jira/browse/FLUME-936
bq.
bq.
bq. Diffs
bq. -----
bq.
bq.
flume-ng-core/src/main/java/org/apache/flume/channel/BasicChannelSemantics.java
PRE-CREATION
bq.
flume-ng-core/src/main/java/org/apache/flume/channel/BasicTransactionSemantics.java
PRE-CREATION
bq. flume-ng-core/src/main/java/org/apache/flume/channel/ChannelUtils.java
PRE-CREATION
bq. flume-ng-core/src/main/java/org/apache/flume/channel/MemoryChannel.java
d379b64
bq.
flume-ng-core/src/test/java/org/apache/flume/channel/TestFanoutChannel.java
ada9a72
bq.
flume-ng-core/src/test/java/org/apache/flume/channel/TestMemoryChannel.java
b44030e
bq.
flume-ng-core/src/test/java/org/apache/flume/channel/TestMemoryChannelConcurrency.java
PRE-CREATION
bq.
flume-ng-core/src/test/java/org/apache/flume/channel/TestMemoryChannelTransaction.java
d18045b
bq. flume-ng-core/src/test/java/org/apache/flume/source/TestExecSource.java
6acbbd5
bq.
bq. Diff: https://reviews.apache.org/r/3704/diff
bq.
bq.
bq. Testing
bq. -------
bq.
bq. The original tests pass, though I had to take out the state checks because
of the changes to semantics from the flume-935 code. I also had to add a
transaction.close statement where semantics were not properly being followed
bq. I have to retrofit my new concurrency test since without the tagged events
it cannot fail without checking that the content is correct. I'll put that up
asap, just wanted to get some eyes on this before I head out.
bq.
bq.
bq. Thanks,
bq.
bq. Juhani
bq.
bq.
> MemoryChannel is not thread safe
> --------------------------------
>
> Key: FLUME-936
> URL: https://issues.apache.org/jira/browse/FLUME-936
> Project: Flume
> Issue Type: Bug
> Components: Channel
> Affects Versions: NG alpha 2
> Reporter: Juhani Connolly
> Assignee: Juhani Connolly
> Fix For: v1.1.0
>
> Attachments: FLUME-936-unittest.patch
>
>
> The memory channel isn't thread safe as a couple of parallel transactions can
> commit/rollback each others entries if called in the wrong order.
> I'm attaching a unit test I made that demonstrates it using a cyclicbarrier
> to force the event order that causes the precondition to fail.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
