[
https://issues.apache.org/jira/browse/BEAM-704?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15555572#comment-15555572
]
Raghu Angadi commented on BEAM-704:
-----------------------------------
.q I'd also suggest we persist the latest offset as part of the CheckpointMark
so that once latest is set, it is remembered until new messages arrive and it
doesn't need to be resolved again (and if there were new messages available
they won't be missed upon failure).
The offset is the primary info stored in CheckpointMark. Can you elaborate on
what is missing now?
.q I think we should consider checking the offsets (could be the same for
"earliest") when running initialSplits (that's how Spark does that as well, one
call from the driver for all topic-partitions).
Can you expand on this with an example? offset is part of checkpoint. So the
reader either resumes from checkpoined offset or the 'default offset' when it
is restarted from scratch. 'Default offset' is 'latest' in default config.
Users can customize this (e.g. by setting a group-id that preserves latest
consumed offset on the Kafka, though this is not the same as Beam checkpoint).
> KafkaIO should handle "latest offset" evenly, and persist it as part of the
> CheckpointMark.
> -------------------------------------------------------------------------------------------
>
> Key: BEAM-704
> URL: https://issues.apache.org/jira/browse/BEAM-704
> Project: Beam
> Issue Type: Improvement
> Components: sdk-java-extensions
> Reporter: Amit Sela
>
> Currently, the KafkaIO (when configured to "latest") will check the latest
> offset on the worker. This means that each worker sees a "different" latest
> for the time it checks for the partitions assigned to it.
> This also means that if a worker fails before starting to read, and new
> messages were added in between, they would be missed.
> I think we should consider checking the offsets (could be the same for
> "earliest") when running initialSplits (that's how Spark does that as well,
> one call from the driver for all topic-partitions).
> I'd also suggest we persist the latest offset as part of the CheckpointMark
> so that once latest is set, it is remembered until new messages arrive and it
> doesn't need to be resolved again (and if there were new messages available
> they won't be missed upon failure).
> For Spark this is even more important as state is passed in-between
> micro-batches and sparse partitions may skip messages until a message finally
> arrives within the read time-frame.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
