[
https://issues.apache.org/jira/browse/FLINK-25360?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17462480#comment-17462480
]
刘方奇 commented on FLINK-25360:
-----------------------------
[~yunta] , Actually, we can restore StateMetaInfoSnapshot in the state file,
but in different mode of state snapshot, there are some diff :
#
org.apache.flink.contrib.streaming.state.snapshot.RocksIncrementalSnapshotStrategy:
we use a different output stream from state output stream to write meta.
# org.apache.flink.contrib.streaming.state.snapshot.RocksFullSnapshotStrategy
/ org.apache.flink.runtime.state.DefaultOperatorStateBackendSnapshotStrategy :
we use one same output stream to write both state data and meta data which
means that we will load the StateMetaInfoSnapshot hardly if the state is so big.
So can we just use different output stream to write the meta data and store the
StreamHandle in a different way like RocksIncrementalSnapshotStrategy. I think
it can improve the performance of restoring state.
> Add State Desc to CheckpointMetadata
> ------------------------------------
>
> Key: FLINK-25360
> URL: https://issues.apache.org/jira/browse/FLINK-25360
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Checkpointing, Runtime / State Backends
> Reporter: 刘方奇
> Priority: Major
> Attachments: image-2021-12-17-20-01-42-423.png
>
>
> Now we can't get the State Descriptor info in the checkpoint meta. Like the
> case if we use state-processor-api to load state then rewrite state, we can't
> flexible use the state.
> Maybe there are other cases we need the State Descriptor, so can we add this
> info?
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
