[
https://issues.apache.org/jira/browse/GEODE-8127?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17110665#comment-17110665
]
ASF subversion and git services commented on GEODE-8127:
--------------------------------------------------------
Commit e084aa4bdaed4d626e559fec4e7e64e1fbf93755 in geode's branch
refs/heads/feature/GEODE-8067 from Darrel Schneider
[ https://gitbox.apache.org/repos/asf?p=geode.git;h=e084aa4 ]
GEODE-8130: use a single region for redis Sets and Hashes (#5120)
* Replaced replicated metaRegion with a partitioned dataRegion.
Currently the dataRegion is used the same way as the metaRegion
except for sets and hashes which store their actual data in it.
* Exception handling now correctly deals with FunctionException
* Disabled a test until GEODE-8127 if fixed.
* Now uses the ByteArrayWrapper as the key on the meta region
and the locks map instead of using a String.
Since a ByteArrayWrapper is used as the key in the data region
this will end up saving memory.
* Found a problem with redis dynamic region management.
Some of the code was executing when we added a new set or hash to
the metaDataRegion. It was only ignoring STRING and HLL.
This caused some extra memory to be used for every redis set/hash.
Now the dynamic region code is only used for lists and sortedSet.
* This commit has some TODO comments of what looks like a bug in
the dynamic region code when a new server is started. It looks
like the new server will not create already existing dynamic regions.
We could test this by starting one server, create a LIST, then
start another server, and then shutdown the first server. Does
the LIST still exist? If we change them not to use dynamic regions
then this issue will go away.
> redis function+delta may not always execute function on primary
> ---------------------------------------------------------------
>
> Key: GEODE-8127
> URL: https://issues.apache.org/jira/browse/GEODE-8127
> Project: Geode
> Issue Type: Bug
> Components: redis
> Reporter: Darrel Schneider
> Assignee: Darrel Schneider
> Priority: Major
>
> The redis use of regions depends on the code that will modify the region that
> is storing redis data, to always execute on the primary. It thought it was
> accomplishing this by marking the function as "optimizeForWrite=true" and by
> routing the function to the node with the bucket using "withFilter(key)".
> This works most of the time. But in some cases the function executes on a
> redundant copy. It looks like what is happening is that at the time the
> function is dispatched it has one idea of who the primary is and sends the
> function to that node. But before it executes the primary moves from this
> node to another that is doing redundancy recovery. Then when our function
> finally does a "put" on the localDataSet it ends up being a remote operation
> that is sent to the other node.
> If our redis function could get a lock that prevents the bucket primary
> status from changing (see BucketRegion doLockForPrimary) and then check to
> see if we are the primary (if not throw an exception that causes the function
> sender to retry (see BucketMovedException) otherwise execute the function
> and at the end release the lock (see BucketRegion doUnlockForPrimary).
> We could enable this with a new method added to Function (much like the
> existing isHA and optimizeForWrite). This new method could be
> executeOnPrimary and default to false (adding a default method to the
> Function interface will not cause backwards compatibility issues unless a
> current class that implements Function already had added a method named
> "executeOnPrimary").
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
