Hey Guozhang, Thanks for the detailed comments. Responses inline:
> 1. I'd like to clarify how we can make "--abort" work with old brokers, since without the additional field "Partitions" the tool needs to set the coordinator epoch correctly instead of "-1"? Arguably that's still doable but would require different call paths, and it's not clear whether that's worth doing for old versions. That's a good question. What I had in mind was to write the marker using the last coordinator epoch that was used by the respective ProducerId. I realized that I left the coordinator epoch out of the `DescribeProducers` response, so I have updated the KIP to include it. It is possible that there is no coordinator epoch associated with a given ProducerId (e.g. if it is the first transaction from that producer), but in this case we can use 0. As for whether this is worth doing, I guess I would be more inclined to leave it out if users had a reasonable alternative today to address this problem. > 2. Why do we have to enforce "DescribeProducers" to be sent to only leaders while ListTransactions can be sent to any brokers? Or is it really "ListTransactions to be sent to coordinators only"? From the workflow you've described, based on the results back from DescribeProducers, we should just immediately send ListTransactions to the corresponding coordinators based on the collected producer ids, instead of trying to send to any brokers right? I'm going to change `DescribeProducers` so that it can be handled by any replica of a topic partition. This was suggested by Lucas in order to allow this API to be used for replica consistency testing. As far as `ListTransactions`, I was treating this similarly to `ListGroups`. Although we know that the coordinators are the leaders of the __transaction_state partitions, this is more of an implementation detail. From an API perspective, we say that any broker could be a transaction coordinator. > 3. One thing I'm a bit hesitant about is that, is `Describe` permission on the associated topic sufficient to allow any users to get all producer information writing to the specific topic-partitions including last timestamp, txn-start-timestamp etc, which may be considered sensitive? Should we require "ClusterAction" to only allow operators only? That's a fair point. Do you think `Read` permission would be reasonable? This is all information that could be obtained by reading the topic. > 4. From the example it seems "TxnStartOffset" should be included in the DescribeTransaction response schema? Otherwise the user would not get it in the following WriteTxnMarker request. The `DescribeTransaction` API is sent to the transaction coordinator, which does not know the start offset of a transaction in each topic partition. That is why we need `DescribeProducers`. > 5. It is a bit easier for readers to highlight the added fields in the existing WriteTxnMarkerRequest (btw I read is that we are only adding "Partitions" with the starting offset, right?). Also as for its response it seems we do not make any schema changes except adding one more potential error code "INVALID_TXN_STATE" to it, right? If that's the case we can just state that explicitly. I highlighted the new field in the request. For the response, the KIP states the following: "There are no changes to the response schema, but it will be bumped. Note that we are also enabling flexible version support." > 6. It is not clear to me for the overloaded function that the following option classes are not specified, what should be the default options? ... I was just trying to stick with existing conventions, but I will add some more detail here. I think we should probably still include `AbortTransactionOptions`. The `Options` classes are how users override timeouts. > 7.1 Is "--broker" a required or optional (in that case I presume we would just query all brokers iteratively) in "--find-hanging"? I think it should be required as a reasonable way to limit the scope of the search. This is meant to be guided by metrics after all. If we do not limit the scope to a single broker, then the behavior might get worse as the cluster grows. I will clarify this. > 7.2 Seems "list-producers" is not exposed as a standalone feature in the cmd but only used in the wrapping "--find-hanging", is that intentional? Personally I feel exposing a "--list-producers" may be useful too: if we believe the user has the right ACL, it is legitimate to return the producer information to her anyways. But that is debatable in the meta point 3) above. Yeah, I was planning to add this to support the use case that Lucas mentioned. There is some awkwardness since it is a little difficult to convey different sources of information through the same command. I guess we can do `--list producers` and `--list transactions` and explain in the documentation. Maybe that is good enough. > 7.3 "Describing Transactions": we should also explain how that would be executed, e.g. at least we should clarify that we would first find the coordinator based on the transactional.id and hence users do not need to specify one. Sure, makes sense. > 7.4. In "Aborting Transactions", should we also specify the "--broker" node as a required option? Otherwise we would not know which broker to send to. The --topic and --partition arguments are required, so the target is always the leader of that partition. Thanks, Jason On Fri, Aug 28, 2020 at 8:13 AM Robert Barrett <bob.barr...@confluent.io> wrote: > Hi Jason, > > Thanks for this KIP, I think this will be a huge operational improvement > and overall it looks great to me. > > I'm not sure how much value the MaxActiveTransactionDuration metric adds, > given that we have the --find-hanging option in the tool. As you mention, > instances of these transactions are expected to be rare, and a > partition-level metric, which can generate a lot of data, seems very > heavyweight for such a rare occurrence. I think "alert on > PartitionsWithLateTransactionsCount" followed by "run kafka-transactions > --find-hanging on the relevant broker" is a reasonable process for cluster > operators to follow. > > Thanks, > Bob > > On Thu, Aug 27, 2020 at 9:23 PM Guozhang Wang <wangg...@gmail.com> wrote: > > > Hi Jason, > > > > Thanks for the written KIP. I think this is going to be a very useful > tool > > for operational improvements since with eos in its current stage, we > cannot > > confidently assert that we are bug-free, and even in the future when we > are > > confident this is still going to be leveraged by older versioned brokers. > > Regarding the solution, I've also debated myself whether Kafka should > > "self-heal" automatically when detected in such situations, or should we > > instead build into ecosystem tooling to let operators do it. And I've > also > > convinced myself that the latter should be a better solution to keep > Kafka > > software itself simpler. > > > > Regarding the KIP itself, I have a few meta comments below: > > > > 1. I'd like to clarify how we can make "--abort" work with old brokers, > > since without the additional field "Partitions" the tool needs to set the > > coordinator epoch correctly instead of "-1"? Arguably that's still doable > > but would require different call paths, and it's not clear whether that's > > worth doing for old versions. > > > > 2. Why do we have to enforce "DescribeProducers" to be sent to only > leaders > > while ListTransactions can be sent to any brokers? Or is it really > > "ListTransactions to be sent to coordinators only"? From the workflow > > you've described, based on the results back from DescribeProducers, we > > should just immediately send ListTransactions to the > > corresponding coordinators based on the collected producer ids, instead > of > > trying to send to any brokers right? > > > > Also I'm a bit concerned if "ListTransactions" could potentially return > too > > much data with "StateFilters" set to all states, including completed > ones. > > Do we expect users ever want to know transactions that are not pending? > On > > the other hand, maybe we can just require users to specify the "pids[]" > in > > this request too to further filter those un-interested transactions. This > > also works well with the workflow: we know exactly from > "DescribeProducers" > > which pids are we diagnosing right now, so in the follow-up > > "ListTransactions" we should also only care for those partitions only. > > > > 3. One thing I'm a bit hesitant about is that, is `Describe` permission > on > > the associated topic sufficient to allow any users to get all producer > > information writing to the specific topic-partitions including last > > timestamp, txn-start-timestamp etc, which may be considered sensitive? > > Should we require "ClusterAction" to only allow operators only? > > > > Below are more detailed comments: > > > > 4. From the example it seems "TxnStartOffset" should be included in the > > DescribeTransaction response schema? Otherwise the user would not get it > in > > the following WriteTxnMarker request. > > > > 5. It is a bit easier for readers to highlight the added fields in the > > existing WriteTxnMarkerRequest (btw I read is that we are only adding > > "Partitions" with the starting offset, right?). Also as for its response > it > > seems we do not make any schema changes except adding one more potential > > error code "INVALID_TXN_STATE" to it, right? If that's the case we can > just > > state that explicitly. > > > > 6. It is not clear to me for the overloaded function that the following > > option classes are not specified, what should be the default options? > > > > * ListTransactionsOptions > > * DescribeTransactionsOptions > > * DescribeProducersOptions > > > > Also, it seems AbortTransactionOptions would just be empty? If yes do we > > really need this option class for now? > > > > 7. A couple questions from the cmd tool examples: > > 7.1 Is "--broker" a required or optional (in that case I presume we would > > just query all brokers iteratively) in "--find-hanging"? > > 7.2 Seems "list-producers" is not exposed as a standalone feature in the > > cmd but only used in the wrapping "--find-hanging", is that intentional? > > Personally I feel exposing a "--list-producers" may be useful too: if we > > believe the user has the right ACL, it is legitimate to return the > producer > > information to her anyways. But that is debatable in the meta point 3) > > above. > > 7.3 "Describing Transactions": we should also explain how that would be > > executed, e.g. at least we should clarify that we would first find the > > coordinator based on the transactional.id and hence users do not need to > > specify one. > > 7.4. In "Aborting Transactions", should we also specify the "--broker" > node > > as a required option? Otherwise we would not know which broker to send > to. > > > > > > Overall, nice written one, thanks Jason. > > > > Guozhang > > > > > > On Thu, Aug 27, 2020 at 11:44 AM Lucas Bradstreet <lu...@confluent.io> > > wrote: > > > > > >> Would it be worth returning transactional.id.expiration.ms in the > > > DescribeProducersResponse? > > > > > > > That's an interesting thought as well. Are you trying to avoid the > need > > > to > > > specify it through the command line? The tool could also query the > value > > > with DescribeConfigs I suppose. > > > > > > Basically. I'm not sure how useful this will be in practice, though it > > > might help when debugging. > > > > > > Lucas > > > > > > On Thu, Aug 27, 2020 at 11:00 AM Jason Gustafson <ja...@confluent.io> > > > wrote: > > > > > > > Hey Lucas, > > > > > > > > Thanks for the comments. Responses below: > > > > > > > > > Given that it's possible for replica producer states to diverge > from > > > each > > > > other, it would be very useful if DescribeProducers(Request,Response) > > and > > > > tooling is able to query all partition replicas for their producers > > > > > > > > Yes, it makes sense to me to let DescribeProducers work on both > > followers > > > > and leaders. In fact, I'm encouraged that there are use cases for > this > > > work > > > > other than detecting hanging transactions. That was indeed the hope, > > but > > > I > > > > didn't have anything specific in mind. I will update the proposal. > > > > > > > > > Would it be worth returning transactional.id.expiration.ms in the > > > > DescribeProducersResponse? > > > > > > > > That's an interesting thought as well. Are you trying to avoid the > need > > > to > > > > specify it through the command line? The tool could also query the > > value > > > > with DescribeConfigs I suppose. > > > > > > > > Thanks, > > > > Jason > > > > > > > > On Thu, Aug 27, 2020 at 10:48 AM Lucas Bradstreet < > lu...@confluent.io> > > > > wrote: > > > > > > > > > Hi Jason, > > > > > > > > > > This looks like a very useful tool, thanks for writing it up. > > > > > > > > > > Given that it's possible for replica producer states to diverge > from > > > each > > > > > other, it would be very useful if > DescribeProducers(Request,Response) > > > and > > > > > tooling is able to query all partition replicas for their > producers. > > > One > > > > > way I can see this being used immediately is in kafka's system > tests, > > > > > especially the ones that inject failures. At the end of the test we > > can > > > > > query all replicas and make sure that their states have not > > diverged. I > > > > can > > > > > also see it being useful when debugging production clusters too. > > > > > > > > > > Would it be worth returning transactional.id.expiration.ms in the > > > > > DescribeProducersResponse? > > > > > > > > > > Cheers, > > > > > > > > > > Lucas > > > > > > > > > > > > > > > > > > > > On Wed, Aug 26, 2020 at 12:12 PM Ron Dagostino <rndg...@gmail.com> > > > > wrote: > > > > > > > > > > > Yes, that definitely sounds reasonable. Thanks, Jason! > > > > > > > > > > > > Ron > > > > > > > > > > > > On Wed, Aug 26, 2020 at 3:03 PM Jason Gustafson < > > ja...@confluent.io> > > > > > > wrote: > > > > > > > > > > > > > Hey Ron, > > > > > > > > > > > > > > We do not typically backport new APIs to older versions. I > think > > we > > > > can > > > > > > > however make the --abort command compatible with older > versions. > > It > > > > > would > > > > > > > require a user to do some analysis on their own to identify a > > > hanging > > > > > > > transaction, but then they can use the tool from a new release > to > > > > > > recover. > > > > > > > For example, users could detect a hanging transaction through > the > > > > > > existing > > > > > > > "LastStableOffsetLag" metric and then collect the needed > > > information > > > > > > from a > > > > > > > dump of the log (or producer snapshot). It's more work, but at > > > least > > > > > it's > > > > > > > possible. Does that sound fair? > > > > > > > > > > > > > > Thanks, > > > > > > > Jason > > > > > > > > > > > > > > On Wed, Aug 26, 2020 at 11:51 AM Ron Dagostino < > > rndg...@gmail.com> > > > > > > wrote: > > > > > > > > > > > > > > > Hi Jason. Thanks for the excellently-written KIP. > > > > > > > > > > > > > > > > Will the implementation be backported to prior Kafka > versions? > > > The > > > > > > > reason > > > > > > > > I ask is because if it is not backported and similar > > > functionality > > > > is > > > > > > not > > > > > > > > otherwise made available for older versions, then the only > > > recourse > > > > > > > (aside > > > > > > > > from deleting and recreating the topic as you pointed out) > may > > be > > > > to > > > > > > > > upgrade to 2.7 (or whatever version ends up getting this > > > > > > functionality). > > > > > > > > Such an upgrade may not be desirable, especially if the > number > > of > > > > > > > > intermediate versions is considerable. I understand the > mantra > > of > > > > > > "never > > > > > > > > fall too many versions behind" but the reality of it is that > it > > > > isn't > > > > > > > > always the case. Even if the version is relatively recent, > an > > > > > upgrade > > > > > > > may > > > > > > > > still not be possible for some time, and a quicker resolution > > may > > > > be > > > > > > > > necessary. > > > > > > > > > > > > > > > > Ron > > > > > > > > > > > > > > > > On Wed, Aug 26, 2020 at 2:33 PM Jason Gustafson < > > > > ja...@confluent.io> > > > > > > > > wrote: > > > > > > > > > > > > > > > > > Hi All, > > > > > > > > > > > > > > > > > > I've added a proposal to handle the problem of hanging > > > > > transactions: > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-664%3A+Provide+tooling+to+detect+and+abort+hanging+transactions > > > > > > > > > . > > > > > > > > > In theory, this should never happen. In practice, we have > hit > > > one > > > > > bug > > > > > > > > where > > > > > > > > > it was possible and there are few good options today to > > > recover. > > > > > > Take a > > > > > > > > > look and let me know what you think. > > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > Jason > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > -- Guozhang > > >
