[
https://issues.apache.org/jira/browse/ZOOKEEPER-965?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13042610#comment-13042610
]
Marshall McMullen commented on ZOOKEEPER-965:
---------------------------------------------
So I'm finally getting a chance to circle back around to Benjamin's comments
above about the stale ChangeRecords that get left in PrepRequestProcessor on a
failed multi-op. I want to be sure I understand what needs to be done here, so
Benjamin, feel free to correct me if I'm off track here....
You mentioned:
> we have to update both the list and hashmap of pending operations.
If I'm reading this right, this is all currently updated via addChangeRecord
(in PrepRequestProcess). I think it's adding it to the list via
zks.outstandingChanges.add(c) and adding to the hashmap via
zks.outstandingChangesForPath.put(c.path, c). So, adding in a
removeChangeRecord method here that mirrors addChangeRecord looks like a good
approach. I think I can use the transaction id on the failed multi op to
identify all the failed change requests that need to be removed. Luckily, all
change requests associated with a multiop will have the same txn id, so that
should be super helpful here.
Also, you mentioned :
> we also need to be careful since the pending operations are modified by the
> thread executing Finalrequestprocessor
Can you elaborate on what the concern is there? If I make the
removeChangeRecord synchronized like addChangeRecord is, will that be
sufficient?
> Need a multi-update command to allow multiple znodes to be updated safely
> -------------------------------------------------------------------------
>
> Key: ZOOKEEPER-965
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-965
> Project: ZooKeeper
> Issue Type: Bug
> Affects Versions: 3.3.3
> Reporter: Ted Dunning
> Assignee: Ted Dunning
> Fix For: 3.4.0
>
> Attachments: ZOOKEEPER-965.patch, ZOOKEEPER-965.patch,
> ZOOKEEPER-965.patch, ZOOKEEPER-965.patch, ZOOKEEPER-965.patch,
> ZOOKEEPER-965.patch, ZOOKEEPER-965.patch, ZOOKEEPER-965.patch,
> ZOOKEEPER-965.patch, ZOOKEEPER-965.patch, ZOOKEEPER-965.patch,
> ZOOKEEPER-965.patch, ZOOKEEPER-965.patch, ZOOKEEPER-965.patch,
> ZOOKEEPER-965.patch, ZOOKEEPER-965.patch, ZOOKEEPER-965.patch,
> ZOOKEEPER-965.patch
>
>
> The basic idea is to have a single method called "multi" that will accept a
> list of create, delete, update or check objects each of which has a desired
> version or file state in the case of create. If all of the version and
> existence constraints can be satisfied, then all updates will be done
> atomically.
> Two API styles have been suggested. One has a list as above and the other
> style has a "Transaction" that allows builder-like methods to build a set of
> updates and a commit method to finalize the transaction. This can trivially
> be reduced to the first kind of API so the list based API style should be
> considered the primitive and the builder style should be implemented as
> syntactic sugar.
> The total size of all the data in all updates and creates in a single
> transaction should be limited to 1MB.
> Implementation-wise this capability can be done using standard ZK internals.
> The changes include:
> - update to ZK clients to all the new call
> - additional wire level request
> - on the server, in the code that converts transactions to idempotent form,
> the code should be slightly extended to convert a list of operations to
> idempotent form.
> - on the client, a down-rev server that rejects the multi-update should be
> detected gracefully and an informative exception should be thrown.
> To facilitate shared development, I have established a github repository at
> https://github.com/tdunning/zookeeper and am happy to extend committer
> status to anyone who agrees to donate their code back to Apache. The final
> patch will be attached to this bug as normal.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
