[jira] [Commented] (KAFKA-10410) OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared in a useless place
[ https://issues.apache.org/jira/browse/KAFKA-10410?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17186092#comment-17186092 ] Mark Shelton commented on KAFKA-10410: -- Yes using a Singleton would work, like I said in my first comment. But I'm not coding up a singleton unless I can avoid it. The metrics of the restore operation are only available in the StateRestoreCallback. Since StateRestoreCallback no longer receives the start/end/complete notification its the same problem. > OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared > in a useless place > > > Key: KAFKA-10410 > URL: https://issues.apache.org/jira/browse/KAFKA-10410 > Project: Kafka > Issue Type: Bug > Components: streams >Affects Versions: 2.6.0 >Reporter: Mark Shelton >Priority: Blocker > > In version 2.5.0 and earlier there are "onRestoreStart" and "onRestoreEnd" > methods on StateRestoreCallback. > Version 2.6.0 removed these calls and put them into StateRestoreListener and > requires "streaming.setGlobalStateRestoreListener". > This makes it impossible for the actual StateRestoreCallback implementation > to receive the start and end indication and is blocking me from moving to > 2.6.0. > See: > [https://kafka.apache.org/25/javadoc/index.html?org/apache/kafka/streams/processor/AbstractNotifyingRestoreCallback.html] > > Related JIRA: > https://issues.apache.org/jira/browse/KAFKA-4322 -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10410) OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared in a useless place
[ https://issues.apache.org/jira/browse/KAFKA-10410?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17186019#comment-17186019 ] Guozhang Wang commented on KAFKA-10410: --- Thanks Mark for your explanation. I think for metrics recording since the callbacks contain the partition / store-name information it can be done independently from the callback. For transaction / cleanup, it depends on your specific scenarios, but if it has to couple with the callbacks one way like Sophie mentioned is to let the global listener keeps reference of each callback that ever gets created / registered --- admittedly it is a bit awkward, but it should work around it. > OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared > in a useless place > > > Key: KAFKA-10410 > URL: https://issues.apache.org/jira/browse/KAFKA-10410 > Project: Kafka > Issue Type: Bug > Components: streams >Affects Versions: 2.6.0 >Reporter: Mark Shelton >Priority: Blocker > > In version 2.5.0 and earlier there are "onRestoreStart" and "onRestoreEnd" > methods on StateRestoreCallback. > Version 2.6.0 removed these calls and put them into StateRestoreListener and > requires "streaming.setGlobalStateRestoreListener". > This makes it impossible for the actual StateRestoreCallback implementation > to receive the start and end indication and is blocking me from moving to > 2.6.0. > See: > [https://kafka.apache.org/25/javadoc/index.html?org/apache/kafka/streams/processor/AbstractNotifyingRestoreCallback.html] > > Related JIRA: > https://issues.apache.org/jira/browse/KAFKA-4322 -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10410) OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared in a useless place
[ https://issues.apache.org/jira/browse/KAFKA-10410?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17185511#comment-17185511 ] Sophie Blee-Goldman commented on KAFKA-10410: - It's hard to say without knowing specifically what happens in your StateRestoreCallback, but handling transactions with the StateRestoreListener sounds potentially unsafe: even before 2.6, `onRestoreEnd` was never guaranteed to be called when a store stopped restoring, only when the restoration actually completed (onRestoreComplete would have been a better name). So it was always possible for onRestoreStart to be called without ever calling onRestoreEnd, or for onRestoreStart to be called multiple times in a row, etc That said, can't you just register the callback of each store+partition to your global restore listener, then use the handle on the callback to do whatever cleanup you needed for that particular store when onRestoreEnd is called for it? It's certainly a bit roundabout, but definitely still possible > OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared > in a useless place > > > Key: KAFKA-10410 > URL: https://issues.apache.org/jira/browse/KAFKA-10410 > Project: Kafka > Issue Type: Bug > Components: streams >Affects Versions: 2.6.0 >Reporter: Mark Shelton >Priority: Blocker > > In version 2.5.0 and earlier there are "onRestoreStart" and "onRestoreEnd" > methods on StateRestoreCallback. > Version 2.6.0 removed these calls and put them into StateRestoreListener and > requires "streaming.setGlobalStateRestoreListener". > This makes it impossible for the actual StateRestoreCallback implementation > to receive the start and end indication and is blocking me from moving to > 2.6.0. > See: > [https://kafka.apache.org/25/javadoc/index.html?org/apache/kafka/streams/processor/AbstractNotifyingRestoreCallback.html] > > Related JIRA: > https://issues.apache.org/jira/browse/KAFKA-4322 -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10410) OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared in a useless place
[
https://issues.apache.org/jira/browse/KAFKA-10410?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17185481#comment-17185481
]
Mark Shelton commented on KAFKA-10410:
--
My implementation of StateRestoreCallback needs to know when the state restore
starts and when it ends. This is because it needs to handle statistics,
transaction stuff and cleanup. In version 2.5 and earlier this was easy as the
methods were already there on {{AbstractNotifyingRestoreCallback.}}
With version 2.6 only the "StateRestoreListener", which is per streams
instance, receives the "onRestartStart" and "onRestartEnd". Since the
"StateRestoreListener" is per streams instance there is no easy way for it to
notify any StateRestoreCallback instance(s).
In version 2.5 and earlier the following was available and convenient. But
since {{AbstractNotifyingRestoreCallback}} is removed the update to version 2.6
is a show stopper.
{{public class MyStateRestoreCallback extends AbstractNotifyingRestoreCallback
{}}
{{}}
{{ @Override}}
{{ public void onRestoreStart(TopicPartition topicPartition, String
storeName, long startingOffset, long endingOffset) {}}
{{ // called by Kafka Streams}}
{{ }}}
{{ @Override}}
{{ public void onRestoreEnd(TopicPartition topicPartition, String storeName,
long totalRestored) {}}
{{ // called by Kafka Streams}}{{}}
{{ }}}
{{...}}
{{}}}
> OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared
> in a useless place
>
>
> Key: KAFKA-10410
> URL: https://issues.apache.org/jira/browse/KAFKA-10410
> Project: Kafka
> Issue Type: Bug
> Components: streams
>Affects Versions: 2.6.0
>Reporter: Mark Shelton
>Priority: Blocker
>
> In version 2.5.0 and earlier there are "onRestoreStart" and "onRestoreEnd"
> methods on StateRestoreCallback.
> Version 2.6.0 removed these calls and put them into StateRestoreListener and
> requires "streaming.setGlobalStateRestoreListener".
> This makes it impossible for the actual StateRestoreCallback implementation
> to receive the start and end indication and is blocking me from moving to
> 2.6.0.
> See:
> [https://kafka.apache.org/25/javadoc/index.html?org/apache/kafka/streams/processor/AbstractNotifyingRestoreCallback.html]
>
> Related JIRA:
> https://issues.apache.org/jira/browse/KAFKA-4322
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
[jira] [Commented] (KAFKA-10410) OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared in a useless place
[ https://issues.apache.org/jira/browse/KAFKA-10410?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17185367#comment-17185367 ] Guozhang Wang commented on KAFKA-10410: --- Not sure if I fully understand what do you mean by "GlobalStateRestoreListener have a way of obtaining StateRestoreCallback", but just to clarify, the "StateRestoreListener" is used for notifying when a restoration is started / ended etc, and the "StateRestoreCallback" is for the actual implementation of applying the changelog records to the state stores, and they should naturally be in two separate implementations: 1) You specify the per-store "StateRestoreCallback" when you register a store. 2) You specify the global "StateRestoreListener" when you instantiate a streams instance. We do not recommend having just one impl class that extends both interfaces moving forward. If you can share your current code maybe that can help me better understanding your pattern. > OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared > in a useless place > > > Key: KAFKA-10410 > URL: https://issues.apache.org/jira/browse/KAFKA-10410 > Project: Kafka > Issue Type: Bug > Components: streams >Affects Versions: 2.6.0 >Reporter: Mark Shelton >Priority: Blocker > > In version 2.5.0 and earlier there are "onRestoreStart" and "onRestoreEnd" > methods on StateRestoreCallback. > Version 2.6.0 removed these calls and put them into StateRestoreListener and > requires "streaming.setGlobalStateRestoreListener". > This makes it impossible for the actual StateRestoreCallback implementation > to receive the start and end indication and is blocking me from moving to > 2.6.0. > See: > [https://kafka.apache.org/25/javadoc/index.html?org/apache/kafka/streams/processor/AbstractNotifyingRestoreCallback.html] > > Related JIRA: > https://issues.apache.org/jira/browse/KAFKA-4322 -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10410) OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared in a useless place
[ https://issues.apache.org/jira/browse/KAFKA-10410?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17182163#comment-17182163 ] Mark Shelton commented on KAFKA-10410: -- The problem is that since 2.6 the StateRestoreCallback has no way of getting called upon "onRestoreStart" and "onRestoreEnd". Setting a GlobalStateRestoreListener does not really solve it since the GlobalStateRestoreListener does not have a good way of obtaining the StateRestoreCallback. One could put a singleton in between but that seems excessively awkward. > OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared > in a useless place > > > Key: KAFKA-10410 > URL: https://issues.apache.org/jira/browse/KAFKA-10410 > Project: Kafka > Issue Type: Bug > Components: streams >Affects Versions: 2.6.0 >Reporter: Mark Shelton >Priority: Blocker > > In version 2.5.0 and earlier there are "onRestoreStart" and "onRestoreEnd" > methods on StateRestoreCallback. > Version 2.6.0 removed these calls and put them into StateRestoreListener and > requires "streaming.setGlobalStateRestoreListener". > This makes it impossible for the actual StateRestoreCallback implementation > to receive the start and end indication and is blocking me from moving to > 2.6.0. > See: > [https://kafka.apache.org/25/javadoc/index.html?org/apache/kafka/streams/processor/AbstractNotifyingRestoreCallback.html] > > Related JIRA: > https://issues.apache.org/jira/browse/KAFKA-4322 -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10410) OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared in a useless place
[ https://issues.apache.org/jira/browse/KAFKA-10410?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17180811#comment-17180811 ] Guozhang Wang commented on KAFKA-10410: --- Hello [~markshelton], thanks for reporting this. Yes this is a by-design fix to decouple the stateRestoreCallback (which is used for just applying records read from changelogs to state stores, per store) from stateRestoreListener (used for indicating the beginning / end / etc during restoration, global) since originally they was merged in a hacky way to just enable certain rocksDB optimization which turns out to have other side effects. Could you let me know why you cannot set the state restore listener through `KafkaStreams#setGlobalStateRestoreListener`? Maybe I can try to figure out something to unblock you. > OnRestoreStart disappeared from StateRestoreCallback in 2.6.0 and reappeared > in a useless place > > > Key: KAFKA-10410 > URL: https://issues.apache.org/jira/browse/KAFKA-10410 > Project: Kafka > Issue Type: Bug > Components: streams >Affects Versions: 2.6.0 >Reporter: Mark Shelton >Priority: Blocker > > In version 2.5.0 and earlier there are "onRestoreStart" and "onRestoreEnd" > methods on StateRestoreCallback. > Version 2.6.0 removed these calls and put them into StateRestoreListener and > requires "streaming.setGlobalStateRestoreListener". > This makes it impossible for the actual StateRestoreCallback implementation > to receive the start and end indication and is blocking me from moving to > 2.6.0. > See: > [https://kafka.apache.org/25/javadoc/index.html?org/apache/kafka/streams/processor/AbstractNotifyingRestoreCallback.html] > > Related JIRA: > https://issues.apache.org/jira/browse/KAFKA-4322 -- This message was sent by Atlassian Jira (v8.3.4#803005)
