[
https://issues.apache.org/jira/browse/FLINK-2324?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14642013#comment-14642013
]
ASF GitHub Bot commented on FLINK-2324:
---------------------------------------
GitHub user gyfora opened a pull request:
https://github.com/apache/flink/pull/937
[FLINK-2324] [streaming] Partitioned state checkpointing rework + test
update
This commit reworks the way partitioned operator states are checkpointed
eliminating per-key checkpointing to increase performance and also updates the
StreamCheckpointingITCase to use partitioned state to check exactly-once
semantics.
It also eliminates a double checkpoint commit issue at head operators in
chains.
**Rework of partitioned state checkpointing**
Previously each OperatorState method would return a PartitionedStateHandle
object upon snapshotting which would contain a Map of Key -> StateHandle,
storing the partitioned states by key. This had two negative implications:
* At every checkpoint we stored the states for each key independently
(slowing the checkpoints and overloading the storage layer)
* We had to keep the keys in-memory (this might cause memory issues with
large keys)
This has been changed in this commit to store the map of states at each
operator in 1 statehandle, minimizing the time it takes to take a snapshot and
also minimizing the load on the storage layer. This might change slightly when
we start thinking about state repartitioning for operator scaling.
**Update StreamCheckpointingITCase with partitioned states**
In the StreamCheckpointingITCase I replaced the RichReduceFunction (which
was working incorrectly as it is not checkpointed) with a RichFlatMapFunction
implementing the stateful group reduce functionality using partitioned states,
to count the number of strings received for a specific prefix.
Using this we now test that exactly-once processing happens.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/gyfora/flink partitioned-state-cleanup
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/flink/pull/937.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #937
----
commit fd0f77a5523b5819724a7152e27d23dad03d02f3
Author: Gyula Fora <gyf...@apache.org>
Date: 2015-07-26T16:02:56Z
[FLINK-2324] [streaming] Partitioned state checkpointing rework + test
update
----
> Rework partitioned state storage
> --------------------------------
>
> Key: FLINK-2324
> URL: https://issues.apache.org/jira/browse/FLINK-2324
> Project: Flink
> Issue Type: Improvement
> Reporter: Gyula Fora
> Assignee: Gyula Fora
>
> Partitioned states are currently stored per-key in statehandles. This is
> alright for in-memory storage but is very inefficient for HDFS.
> The logic behind the current mechanism is that this approach provides a way
> to repartition a state without fetching the data from the external storage
> and only manipulating handles.
> We should come up with a solution that can achieve both.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
