[ 
https://issues.apache.org/jira/browse/FLINK-6034?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15940129#comment-15940129
 ] 

ASF GitHub Bot commented on FLINK-6034:
---------------------------------------

Github user StefanRRichter commented on a diff in the pull request:

    https://github.com/apache/flink/pull/3531#discussion_r107874353
  
    --- Diff: 
flink-contrib/flink-statebackend-rocksdb/src/main/java/org/apache/flink/contrib/streaming/state/RocksDBKeyedStateBackend.java
 ---
    @@ -761,6 +769,13 @@ private void restoreKVStateMetaData() throws 
IOException, ClassNotFoundException
                private void restoreKVStateData() throws IOException, 
RocksDBException {
                        //for all key-groups in the current state handle...
                        for (Tuple2<Integer, Long> keyGroupOffset : 
currentKeyGroupsStateHandle.getGroupRangeOffsets()) {
    +                           int keyGroup = keyGroupOffset.f0;
    +
    +                           // Skip those key groups that do not belong to 
the backend
    +                           if 
(!rocksDBKeyedStateBackend.getKeyGroupRange().contains(keyGroup)) {
    --- End diff --
    
    The same comment I had on the `HeapKeyedStateBackend` also applies here: I 
think the post-filter is no longer required after the change i suggested for 
`StateAssignmentOperation`.


> Add KeyedStateHandle for the snapshots in keyed streams
> -------------------------------------------------------
>
>                 Key: FLINK-6034
>                 URL: https://issues.apache.org/jira/browse/FLINK-6034
>             Project: Flink
>          Issue Type: Sub-task
>          Components: State Backends, Checkpointing
>            Reporter: Xiaogang Shi
>            Assignee: Xiaogang Shi
>
> Currently, the only type of the snapshots in keyed streams is 
> {{KeyGroupsStateHandle}} which is full and store the states one group after 
> another. With the introduction of incremental checkpointing, we need a higher 
> level abstraction of keyed snapshots to allow flexible snapshot formats. 
> The implementation of {{KeyedStateHandle}} s may vary a lot in different 
> backends. The only information needed in {{KeyedStateHandle}} s is their key 
> group range. When recovering the job with a different degree of parallelism, 
> {{KeyedStateHandle}} s will be assigned to those subtasks whose key group 
> ranges overlap with their ranges.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Reply via email to