[
https://issues.apache.org/jira/browse/FLINK-3231?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15101902#comment-15101902
]
Stephan Ewen commented on FLINK-3231:
-------------------------------------
What we could start adding to Flink is a kind of state that is globally merged.
For example:
- Source subtask 1 checkpoints state (shard1 - offset 56)
- Source subtask 2 checkpoints state (shard2- offset 42)
- Source subtask 3 checkpoints state (shard3 - offset 17)
The checkpoint coordinator makes one state out of that: [ (shard1 - offset 56)
, (shard2 - offset 42) , (shard3 - offset 17) ].
On restore, all tasks get the full state.
Let's say we restore the job and the assignment changed such that source
subtask 1 now gets shard1 and 2. It has all required offsets to start working
from that union of shards and not introduce duplicates.
> Handle Kinesis-side resharding in Kinesis streaming consumer
> ------------------------------------------------------------
>
> Key: FLINK-3231
> URL: https://issues.apache.org/jira/browse/FLINK-3231
> Project: Flink
> Issue Type: Sub-task
> Components: Streaming Connectors
> Affects Versions: 1.0.0
> Reporter: Tzu-Li (Gordon) Tai
>
> A big difference between Kinesis shards and Kafka partitions is that Kinesis
> users can choose to "merge" and "split" shards at any time for adjustable
> stream throughput capacity. This article explains this quite clearly:
> https://brandur.org/kinesis-by-example.
> This will break the static shard-to-task mapping implemented in the basic
> version of the Kinesis consumer
> (https://issues.apache.org/jira/browse/FLINK-3229). The static shard-to-task
> mapping is done in a simple round-robin-like distribution which can be
> locally determined at each Flink consumer task (Flink Kafka consumer does
> this too).
> To handle Kinesis resharding, we will need some way to let the Flink consumer
> tasks coordinate which shards they are currently handling, and allow the
> tasks to ask the coordinator for a shards reassignment when the task finds
> out it has found a closed shard at runtime (shards will be closed by Kinesis
> when it is merged and split).
> We need a centralized coordinator state store which is visible to all Flink
> consumer tasks. Tasks can use this state store to locally determine what
> shards it can be reassigned. Amazon KCL uses a DynamoDB table for the
> coordination, but as described in
> https://issues.apache.org/jira/browse/FLINK-3211, we unfortunately can't use
> KCL for the implementation of the consumer if we want to leverage Flink's
> checkpointing mechanics. For our own implementation, Zookeeper can be used
> for this state store, but that means it would require the user to set up ZK
> to work.
> Since this feature introduces extensive work, it is opened as a separate
> sub-task from the basic implementation
> https://issues.apache.org/jira/browse/FLINK-3229.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
