[
https://issues.apache.org/jira/browse/KAFKA-7882?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16768663#comment-16768663
]
Matthias J. Sax commented on KAFKA-7882:
----------------------------------------
Well. I guess you can do some manual book-keeping for get exactly one output
per window using wall-clock time punctuation – ie, if you evict, you need to
remember that you did evict in case an even later record appears. The question
is, when can you stop to remember? Seems like you would need to keep this
information forever to be 100% sure?
> StateStores are frequently closed during the 'transform' method
> ---------------------------------------------------------------
>
> Key: KAFKA-7882
> URL: https://issues.apache.org/jira/browse/KAFKA-7882
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Affects Versions: 2.0.0
> Reporter: Mateusz Owczarek
> Priority: Major
>
> Hello, I have a problem with the state store being closed frequently while
> transforming upcoming records. To ensure only one record of the same key and
> the window comes to an aggregate I have created a custom Transformer (I know
> something similar is going to be introduced with suppress method on KTable in
> the future, but my implementation is quite simple and imo should work
> correctly) with the following implementation:
> {code:java}
> override def transform(key: Windowed[K], value: V): (Windowed[K], V) = {
> val partition = context.partition()
> if (partition != -1) store.put(key.key(), (value, partition),
> key.window().start())
> else logger.warn(s"-1 partition")
> null //Ensuring no 1:1 forwarding, context.forward and commit logic is in the
> punctuator callback
> }
> {code}
>
> What I do get is the following error:
> {code:java}
> Store MyStore is currently closed{code}
> This problem appears only when the number of streaming threads (or input
> topic partitions) is greater than 1 even if I'm just saving to the store and
> turn off the punctuation.
> If punctuation is present, however, I sometimes get -1 as a partition value
> in the transform method. I'm familiar with the basic docs, however, I haven't
> found anything that could help me here.
> I build my state store like this:
> {code:java}
> val stateStore = Stores.windowStoreBuilder(
> Stores.persistentWindowStore(
> stateStoreName,
> timeWindows.maintainMs() + timeWindows.sizeMs +
> TimeUnit.DAYS.toMillis(1),
> timeWindows.segments,
> timeWindows.sizeMs,
> false
> ),
> serde[K],
> serde[(V, Int)]
> )
> {code}
> and include it in a DSL API like this:
> {code:java}
> builder.addStateStore(stateStore)
> (...).transform(new MyTransformer(...), "MyStore")
> {code}
> INB4: I don't close any state stores manually, I gave them retention time as
> long as possible for the debugging stage, I tried to hotfix that with the
> retry in the transform method and the state stores reopen at the end and the
> `put` method works, but this approach is questionable and I am concerned if
> it actually works.
> Edit:
> May this be because of the fact that the
> {code:java}StreamsConfig.CACHE_MAX_BYTES_BUFFERING_CONFIG{code} is set to low
> value? If I understand correctly, spilling to disk is done therefore more
> frequently, may it, therefore, cause closing the store?
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
