[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17192885#comment-17192885 ] feyman commented on KAFKA-10284: Just FYI, created a PR: https://github.com/apache/kafka/pull/9270 > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Critical > Labels: help-wanted > Attachments: How to reproduce the issue in KAFKA-10284.md > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17182608#comment-17182608 ] feyman commented on KAFKA-10284: Checked with [~bchen225242] offline, and I just reproduced the issue with the procedure as in the attachment, happy to know if there are more failure scenarios related.[^How to reproduce the issue in KAFKA-10284.md] > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Critical > Labels: help-wanted > Attachments: How to reproduce the issue in KAFKA-10284.md > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17176639#comment-17176639 ] Boyang Chen commented on KAFKA-10284: - I would say for all. > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Priority: Major > Labels: help-wanted > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17175905#comment-17175905 ] Sophie Blee-Goldman commented on KAFKA-10284: - [~bchen225242] [~hachikuji] Is this a correctness problem for all applications or only EOS? > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Priority: Major > Labels: help-wanted > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17175669#comment-17175669 ] Boyang Chen commented on KAFKA-10284: - Resign from this ticket for now, others feel free to pick up. > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Priority: Major > Labels: help-wanted > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17166681#comment-17166681 ] Boyang Chen commented on KAFKA-10284: - That's a good observation, I forgot that for static members they don't actually do another Fetch offset immediately after they rejoin the group. You should be right about this, and I haven't thought about the scenario where EOS and static membership are both turned on. > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17165892#comment-17165892 ] Jason Gustafson commented on KAFKA-10284: - Hmm.. Not totally sure I buy this: > A static member X joins the group and updates member.id to M1, then gets stuck > Another static member Y with the same instance.id joins and updates member.id > to M2, while starts working and commit offsets > The group coordinator migrates, and the member.id for the same static member > rewinds to M1 > The static member X goes back online, and validated. It would try to fetch > from Y's committed offset Why would member X fetch Y's committed offset? If it doesn't know it had been fenced temporarily, it might just commit its latest offsets. This does seem like a correctness problem to me. > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17164531#comment-17164531 ] Sophie Blee-Goldman commented on KAFKA-10284: - Nope, it just spun in a loop where it would poll and then call Consumer#position to try and initialize some metadata. Poll just returned empty I guess and Consumer#position continued to throw TimeoutException. The TimeoutException part seems pretty weird, so maybe it was an unrelated bug (or some other issue like a hung socket being the primary/only guess we had). I just thought of it because it definitely happened immediately after a "id does not match expected member.id" error. I actually still have some of the broker and client side logs from around this incident, if that might help. But again, I'm not really sure if it could be related to this or not > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17164146#comment-17164146 ] Akshay Sharma commented on KAFKA-10284: --- At initial, Group Coordinator will assign member id to the new joined consumer i.e X with generation id 1, when the consumer take off at the same point and try to rejoin there be no re-balance(i.e generation id will be 1), but group coordinator will assign new member id to the consumer i.e Y. Now, when broker restart due to some reason, groupmetadatamanager will reload the metadata and expects that member.id to be present, he is now expecting X member.id. and that's why consumer is fenced. 1) Is this a expected behaviour? 2) when consumer member id changes, groupmetadatamanager not updates new member.id information? 3) consumer is fenced with new instance, even single consumer is running. Why? [~bchen225242] [~guozhang] > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17164063#comment-17164063 ] Guozhang Wang commented on KAFKA-10284: --- What [~ableegoldman] described seems aligned to this, BUT I thought that fenced instance Y should still try to rejoin the group, but in [~ableegoldman] that thread did not try to rejoin at all? > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17163871#comment-17163871 ] Boyang Chen commented on KAFKA-10284: - [~akshaysh] I didn't see any trace that the group coordinator gets migrated in the pasted ticket, so it might be a separate issue. [~ableegoldman] Well, the symptom matches, but I don't know for sure if this is the same cause :) > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17163796#comment-17163796 ] Sophie Blee-Goldman commented on KAFKA-10284: - You know, I think we actually hit this too, but weren't able to recognize the problem at the time. A few weeks ago one of our StreamThreads/Consumers seemed to "take off" from the group at some point, as evidenced by the steadily increasing last-rebalance-seconds-ago metric (whereas the other members had rebalanced multiple times since then). Right before this occurred we saw that same error message in the logs: {code:java} ERROR given member.id X is identified as a known static member 1,but not matching the expected member.id Y (kafka.coordinator.group.GroupMetadata) {code} Unfortunately we killed the client trying to remotely debug it so we couldn't get any more useful information. Would you say that this was mysterious encounter was likely due to the bug reported here? [~guozhang] [~bchen225242] > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17163248#comment-17163248 ] Akshay Sharma commented on KAFKA-10284: --- Hi [~bchen225242]/[~guozhang], I've raised the bug related to the same issue. Please look at it once. https://issues.apache.org/jira/browse/KAFKA-10285 Analysis, `when i've not restarted the broker and restarted consumer, I could see below logs` ``` [2020-07-16 13:56:17,189] INFO [GroupCoordinator 1001]: Preparing to rebalance group 0 in state PreparingRebalance with old generation 0 (__consumer_offsets-48) (reason: Adding new member 1-159490144 with group instanceid Some(1)) (kafka.coordinator.group.GroupCoordinator) [2020-07-16 13:56:17,236] INFO [GroupCoordinator 1001]: Stabilized group 0 generation 1 (__consumer_offsets-48) (kafka.coordinator.group.GroupCoordinator) [2020-07-16 13:56:17,282] INFO [GroupCoordinator 1001]: Assignment received from leader for group 0 for generation 1 (kafka.coordinator.group.GroupCoordinator) [2020-07-16 13:59:33,613] INFO [GroupCoordinator 1001]: Static member Some(1) with unknown member id rejoins, assigning new member id 1-159490335, while old member 1-159490144 will be removed. (kafka.coordinator.group.GroupCoordinator) [2020-07-16 13:59:33,635] INFO [GroupCoordinator 1001]: Static member joins during Stable stage will not trigger rebalance. (kafka.coordinator.group.GroupCoordinator) ``` when restarted the broker, I could see broker is expecting some other member.id(old member.id of consumer) ``` 2020-07-16 14:04:04,953] ERROR given member.id 1-159490335 is identified as a known static member 1,but not matching the expected member.id 1-159490144 (kafka.coordinator.group.GroupMetadata) ``` > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17160225#comment-17160225 ] Guozhang Wang commented on KAFKA-10284: --- Fair enough. So it seems not a correctness breaking issue at the moment. > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17160216#comment-17160216 ] Boyang Chen commented on KAFKA-10284: - If we have a combined scenario like below: # A static member X joins the group and updates member.id to M1, then gets stuck # Another static member Y with the same instance.id joins and updates member.id to M2, while starts working and commit offsets # The group coordinator migrates, and the member.id for the same static member rewinds to M1 # The static member X goes back online, and validated. It would try to fetch from Y's committed offset In this flow, I don't think we are violating the offset committing policy here. The only downside I could think of is that there is only one member Y who will get fenced by itself after the immigration as stated in the KIP. [~guozhang] > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. > The bug find credit goes to [~hachikuji] -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17160152#comment-17160152 ] Guozhang Wang commented on KAFKA-10284: --- [~bchen225242] Thanks for the find. Just trying to clarify that "bring up the zombie member ID" would break correctness, since if the zombie member still exists, it may be able to commit offsets which would rewind the position. Is that right? > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. -- This message was sent by Atlassian Jira (v8.3.4#803005)
[jira] [Commented] (KAFKA-10284) Group membership update due to static member rejoin should be persisted
[ https://issues.apache.org/jira/browse/KAFKA-10284?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel=17159616#comment-17159616 ] Boyang Chen commented on KAFKA-10284: - [https://github.com/apache/kafka/blob/trunk/core/src/main/scala/kafka/coordinator/group/GroupCoordinator.scala#L1042] > Group membership update due to static member rejoin should be persisted > --- > > Key: KAFKA-10284 > URL: https://issues.apache.org/jira/browse/KAFKA-10284 > Project: Kafka > Issue Type: Bug > Components: consumer >Affects Versions: 2.3.0, 2.4.0, 2.5.0, 2.6.0 >Reporter: Boyang Chen >Assignee: Boyang Chen >Priority: Major > Fix For: 2.6.1 > > > For known static members rejoin, we would update its corresponding member.id > without triggering a new rebalance. This serves the purpose for avoiding > unnecessary rebalance for static membership, as well as fencing purpose if > some still uses the old member.id. > The bug is that we don't actually persist the membership update, so if no > upcoming rebalance gets triggered, this new member.id information will get > lost during group coordinator immigration, thus bringing up the zombie member > identity. -- This message was sent by Atlassian Jira (v8.3.4#803005)