[
https://issues.apache.org/jira/browse/SPARK-9434?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14646011#comment-14646011
]
Sean Owen commented on SPARK-9434:
----------------------------------
Yes, I mean "direct stream". While I don't think you should open a JIRA to get
attention for your question, I do think it's clearer what this is about now and
it could be a legitimate issue, so let me see if I can move it along one tick:
Per https://spark.apache.org/docs/latest/streaming-kafka-integration.html I
think the answer is "no" you do not manually manage offsets in order to achieve
the exactly-once semantics advertised by the Kafka direct stream. The offsets
are recorded with checkpoints. However from glancing at the code, that also
implies to me that you have to enable checkpointing to get this to work. That
is not shown in the example; I am not sure whether my read is wrong, or whether
it's implied, or just omitted from the example.
If that's correct, then I think that's the solution you want (or else, manage
offsets manually in ZK -- I do that and it works fine) but the example needs to
be updated to show checkpointing.
Otherwise, the question is why you're not getting the expected behavior
(without checkpoints), and a simple reproduction would help.
CC [[email protected]]
> Need how-to for resuming direct Kafka streaming consumers where they had left
> off before getting terminated, OR actual support for that mode in the
> Streaming API
> -----------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: SPARK-9434
> URL: https://issues.apache.org/jira/browse/SPARK-9434
> Project: Spark
> Issue Type: Improvement
> Components: Documentation, Examples, Streaming
> Affects Versions: 1.4.1
> Reporter: Dmitry Goldenberg
>
> We've been getting some mixed information regarding how to cause our direct
> streaming consumers to resume processing from where they left off in terms of
> the Kafka offsets.
> On the one hand side, we're hearing "If you are restarting the streaming app
> with Direct kafka from the checkpoint information (that is, restarting), then
> the last read offsets are automatically recovered, and the data will start
> processing from that offset. All the N records added in T will stay buffered
> in Kafka." (where T is the interval of time during which the consumer was
> down).
> On the other hand, there are tickets such as SPARK-6249 and SPARK-8833 which
> are marked as "won't fix" which seem to ask for the functionality we need,
> with comments like "I don't want to add more config options with confusing
> semantics around what is being used for the system of record for offsets, I'd
> rather make it easy for people to explicitly do what they need."
> The use-case is actually very clear and doesn't ask for confusing semantics.
> An API option to resume reading where you left off, in addition to the
> smallest or greatest auto.offset.reset should be *very* useful, probably for
> quite a few folks.
> We're asking for this as an enhancement request. SPARK-8833 states " I am
> waiting for getting enough usecase to float in before I take a final call."
> We're adding to that.
> In the meantime, can you clarify the confusion? Does direct streaming
> persist the progress information into "DStream checkpoints" or does it not?
> If it does, why is it that we're not seeing that happen? Our consumers start
> with auto.offset.reset=greatest and that causes them to read from the first
> offset of data that is written to Kafka *after* the consumer has been
> restarted, meaning we're missing data that had come in while the consumer was
> down.
> If the progress is stored in "DStream checkpoints", we want to know a) how to
> cause that to work for us and b) where the said checkpointing data is stored
> physically.
> Conversely, if this is not accurate, then is our only choice to manually
> persist the offsets into Zookeeper? If that is the case then a) we'd like a
> clear, more complete code sample to be published, since the one in the Kafka
> streaming guide is incomplete (it lacks the actual lines of code persisting
> the offsets) and b) we'd like to request that SPARK-8833 be revisited as a
> feature worth implementing in the API.
> Thanks.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
