[
https://issues.apache.org/jira/browse/SPARK-18258?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15637576#comment-15637576
]
Cody Koeninger commented on SPARK-18258:
----------------------------------------
The sink doesn't have to reason about equality of the representations.
It just has to be able to store those representations, in addition the batch id
if necessary, so that the job can be recovered if spark fails in a way that
renders the batch id meaningless or the user wants to switch to a different
streaming system.
> Sinks need access to offset representation
> ------------------------------------------
>
> Key: SPARK-18258
> URL: https://issues.apache.org/jira/browse/SPARK-18258
> Project: Spark
> Issue Type: Improvement
> Components: Structured Streaming
> Reporter: Cody Koeninger
>
> Transactional "exactly-once" semantics for output require storing an offset
> identifier in the same transaction as results.
> The Sink.addBatch method currently only has access to batchId and data, not
> the actual offset representation.
> I want to store the actual offsets, so that they are recoverable as long as
> the results are and I'm not locked in to a particular streaming engine.
> I could see this being accomplished by adding parameters to Sink.addBatch for
> the starting and ending offsets (either the offsets themselves, or the
> SPARK-17829 string/json representation). That would be an API change, but if
> there's another way to map batch ids to offset representations without
> changing the Sink api that would work as well.
> I'm assuming we don't need the same level of access to offsets throughout a
> job as e.g. the Kafka dstream gives, because Sinks are the main place that
> should need them.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
