Re: more uniform task assignment across kafka stream nodes
Awesome! Thanks. Can’t wait to try it out. Ara. On Mar 28, 2017, at 2:13 PM, Matthias J. Sax <matth...@confluent.io<mailto:matth...@confluent.io>> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 28, 2017 at 2:13:51 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> Created a JIRA: https://issues.apache.org/jira/browse/KAFKA-4969 -Matthias On 3/27/17 4:33 PM, Ara Ebrahimi wrote: Well, even with 4-5x better performance thanks to the session window fix, I expect to get ~10x better performance if I throw 10x more nodes at the problem. That won’t be the case due to task assignment unfortunately. I may end up with say 5-6 nodes with aggregation assigned to them and 4-5 nodes sitting there doing nothing. So it is a problem. Ara. On Mar 27, 2017, at 4:15 PM, Matthias J. Sax <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 27, 2017 at 4:15:07 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>> Great! So overall, the issue is not related to task assignment. Also the description below, does not indicate that different task assignment would change anything. -Matthias On 3/27/17 3:08 PM, Ara Ebrahimi wrote: Let me clarify, cause I think we’re using different terminologies: - message key is phone number, reversed - all call records for a phone number land on the same partition - then we apply a session window on them and aggregate+reduce - so we end up with a group of records for a phone number. This group is literally an avro object with an array of records in it for the session. - since records arrive chronologically and since we have a session window, then all call records for a phone number end up in the same partition and session (intended behavior). We can easily have many such phone call record groups with 100s of call records in them. The aggregate object (the avro object with array of records in it) can get updated 100s of times for the same phone number in the course of an hour or so. - we process billions of such call records a day - we can’t expect our clients to install massive clusters of 10s or 100s of nodes. We do need to make sure this processing is as efficient as possible. The session window bug was killing us. Much better with the fix Damian provided! Ara. On Mar 27, 2017, at 2:41 PM, Matthias J. Sax <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io><mailto:matth...@confluent.io>> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. ____ From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io><mailto:matth...@confluent.io>> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 27, 2017 at 2:41:06 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org><mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>> Ara, I assume your performance issue is most likely related to the fix Damian pointed out already. Couple o
Re: more uniform task assignment across kafka stream nodes
Created a JIRA: https://issues.apache.org/jira/browse/KAFKA-4969 -Matthias On 3/27/17 4:33 PM, Ara Ebrahimi wrote: > Well, even with 4-5x better performance thanks to the session window fix, I > expect to get ~10x better performance if I throw 10x more nodes at the > problem. That won’t be the case due to task assignment unfortunately. I may > end up with say 5-6 nodes with aggregation assigned to them and 4-5 nodes > sitting there doing nothing. So it is a problem. > > Ara. > > On Mar 27, 2017, at 4:15 PM, Matthias J. Sax > <matth...@confluent.io<mailto:matth...@confluent.io>> wrote: > > > > > > > This message is for the designated recipient only and may contain privileged, > proprietary, or otherwise confidential information. If you have received it > in error, please notify the sender immediately and delete the original. Any > other use of the e-mail by you is prohibited. Thank you in advance for your > cooperation. > > > > From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>> > Subject: Re: more uniform task assignment across kafka stream nodes > Date: March 27, 2017 at 4:15:07 PM PDT > To: users@kafka.apache.org<mailto:users@kafka.apache.org> > Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> > > > Great! > > So overall, the issue is not related to task assignment. Also the > description below, does not indicate that different task assignment > would change anything. > > > -Matthias > > On 3/27/17 3:08 PM, Ara Ebrahimi wrote: > Let me clarify, cause I think we’re using different terminologies: > > - message key is phone number, reversed > - all call records for a phone number land on the same partition > - then we apply a session window on them and aggregate+reduce > - so we end up with a group of records for a phone number. This group is > literally an avro object with an array of records in it for the session. > - since records arrive chronologically and since we have a session window, > then all call records for a phone number end up in the same partition and > session (intended behavior). We can easily have many such phone call record > groups with 100s of call records in them. The aggregate object (the avro > object with array of records in it) can get updated 100s of times for the > same phone number in the course of an hour or so. > - we process billions of such call records a day > - we can’t expect our clients to install massive clusters of 10s or 100s of > nodes. We do need to make sure this processing is as efficient as possible. > The session window bug was killing us. Much better with the fix Damian > provided! > > Ara. > > On Mar 27, 2017, at 2:41 PM, Matthias J. Sax > <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>> > wrote: > > > > > > > This message is for the designated recipient only and may contain privileged, > proprietary, or otherwise confidential information. If you have received it > in error, please notify the sender immediately and delete the original. Any > other use of the e-mail by you is prohibited. Thank you in advance for your > cooperation. > > > > From: "Matthias J. Sax" > <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>> > Subject: Re: more uniform task assignment across kafka stream nodes > Date: March 27, 2017 at 2:41:06 PM PDT > To: > users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org> > Reply-To: > <users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>> > > > Ara, > > I assume your performance issue is most likely related to the fix Damian > pointed out already. > > Couple of follow up comments: > > critical part of our pipeline involves grouping relevant records together > > Can you explain this a little better? The abstraction of a task does > group data together already. Furthermore, if you get multiple tasks, > those are independent units with regard to grouping as consecutive tasks > are "connected" via repartitioning steps. Thus, even if we apply the > task assignment as you ask for, I am not sure if this would change > anything? Maybe you can give a small data example what > behavior/data-co-location you need and what Streams provides for you. > > And for hot keys this can lead to sometimes 100s of records to get grouped > together
Re: more uniform task assignment across kafka stream nodes
Well, even with 4-5x better performance thanks to the session window fix, I expect to get ~10x better performance if I throw 10x more nodes at the problem. That won’t be the case due to task assignment unfortunately. I may end up with say 5-6 nodes with aggregation assigned to them and 4-5 nodes sitting there doing nothing. So it is a problem. Ara. On Mar 27, 2017, at 4:15 PM, Matthias J. Sax <matth...@confluent.io<mailto:matth...@confluent.io>> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 27, 2017 at 4:15:07 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> Great! So overall, the issue is not related to task assignment. Also the description below, does not indicate that different task assignment would change anything. -Matthias On 3/27/17 3:08 PM, Ara Ebrahimi wrote: Let me clarify, cause I think we’re using different terminologies: - message key is phone number, reversed - all call records for a phone number land on the same partition - then we apply a session window on them and aggregate+reduce - so we end up with a group of records for a phone number. This group is literally an avro object with an array of records in it for the session. - since records arrive chronologically and since we have a session window, then all call records for a phone number end up in the same partition and session (intended behavior). We can easily have many such phone call record groups with 100s of call records in them. The aggregate object (the avro object with array of records in it) can get updated 100s of times for the same phone number in the course of an hour or so. - we process billions of such call records a day - we can’t expect our clients to install massive clusters of 10s or 100s of nodes. We do need to make sure this processing is as efficient as possible. The session window bug was killing us. Much better with the fix Damian provided! Ara. On Mar 27, 2017, at 2:41 PM, Matthias J. Sax <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 27, 2017 at 2:41:06 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>> Ara, I assume your performance issue is most likely related to the fix Damian pointed out already. Couple of follow up comments: critical part of our pipeline involves grouping relevant records together Can you explain this a little better? The abstraction of a task does group data together already. Furthermore, if you get multiple tasks, those are independent units with regard to grouping as consecutive tasks are "connected" via repartitioning steps. Thus, even if we apply the task assignment as you ask for, I am not sure if this would change anything? Maybe you can give a small data example what behavior/data-co-location you need and what Streams provides for you. And for hot keys this can lead to sometimes 100s of records to get grouped together This is independent of task placement -- it a partitioning issue. If you want to work on that, you can provide a custom `Partioner` for the used Kafka producers (note, your external Producer writing to your Streams input topic might already generate "hot" keys, so you might need to use a custom partitioner there, too) Also "100s of records" does not sound much to me. Streams can process multiple hundredths of thousandth records per thread. That is the reason, why I think that the fix Damian pointed out will most likely fix your problem. -Matthias On 3/27/17 1:56 PM, Ara Ebrahimi wrote: Thanks for the response Mathias! The reason
Re: more uniform task assignment across kafka stream nodes
Great! So overall, the issue is not related to task assignment. Also the description below, does not indicate that different task assignment would change anything. -Matthias On 3/27/17 3:08 PM, Ara Ebrahimi wrote: > Let me clarify, cause I think we’re using different terminologies: > > - message key is phone number, reversed > - all call records for a phone number land on the same partition > - then we apply a session window on them and aggregate+reduce > - so we end up with a group of records for a phone number. This group is > literally an avro object with an array of records in it for the session. > - since records arrive chronologically and since we have a session window, > then all call records for a phone number end up in the same partition and > session (intended behavior). We can easily have many such phone call record > groups with 100s of call records in them. The aggregate object (the avro > object with array of records in it) can get updated 100s of times for the > same phone number in the course of an hour or so. > - we process billions of such call records a day > - we can’t expect our clients to install massive clusters of 10s or 100s of > nodes. We do need to make sure this processing is as efficient as possible. > The session window bug was killing us. Much better with the fix Damian > provided! > > Ara. > > On Mar 27, 2017, at 2:41 PM, Matthias J. Sax > <matth...@confluent.io<mailto:matth...@confluent.io>> wrote: > > > > > > > This message is for the designated recipient only and may contain privileged, > proprietary, or otherwise confidential information. If you have received it > in error, please notify the sender immediately and delete the original. Any > other use of the e-mail by you is prohibited. Thank you in advance for your > cooperation. > > ________________ > > From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>> > Subject: Re: more uniform task assignment across kafka stream nodes > Date: March 27, 2017 at 2:41:06 PM PDT > To: users@kafka.apache.org<mailto:users@kafka.apache.org> > Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> > > > Ara, > > I assume your performance issue is most likely related to the fix Damian > pointed out already. > > Couple of follow up comments: > > critical part of our pipeline involves grouping relevant records together > > Can you explain this a little better? The abstraction of a task does > group data together already. Furthermore, if you get multiple tasks, > those are independent units with regard to grouping as consecutive tasks > are "connected" via repartitioning steps. Thus, even if we apply the > task assignment as you ask for, I am not sure if this would change > anything? Maybe you can give a small data example what > behavior/data-co-location you need and what Streams provides for you. > > And for hot keys this can lead to sometimes 100s of records to get grouped > together > > This is independent of task placement -- it a partitioning issue. If you > want to work on that, you can provide a custom `Partioner` for the used > Kafka producers (note, your external Producer writing to your Streams > input topic might already generate "hot" keys, so you might need to use > a custom partitioner there, too) > > Also "100s of records" does not sound much to me. Streams can process > multiple hundredths of thousandth records per thread. That is the > reason, why I think that the fix Damian pointed out will most likely fix > your problem. > > > > -Matthias > > > On 3/27/17 1:56 PM, Ara Ebrahimi wrote: > Thanks for the response Mathias! > > The reason we want this exact task assignment to happen is that a critical > part of our pipeline involves grouping relevant records together (that’s what > the aggregate function in the topology is for). And for hot keys this can > lead to sometimes 100s of records to get grouped together. Even worse, these > records are session bound, we use session windows. Hence we see lots of > activity around the store backing the aggregate function and even though we > use SSD drives we’re not seeing the kind of performance we want to see. It > seems like the aggregate function leads to lots of updates to these hot keys > which lead to lots of rocksdb activity. > > Now there are many ways to fix this problem: > - just don’t aggregate, create an algorithm which is not reliant on > grouping/aggregating records. Not what we can do with our tight schedule > right now. > - do grouping/aggregating but employ n
Re: more uniform task assignment across kafka stream nodes
Let me clarify, cause I think we’re using different terminologies: - message key is phone number, reversed - all call records for a phone number land on the same partition - then we apply a session window on them and aggregate+reduce - so we end up with a group of records for a phone number. This group is literally an avro object with an array of records in it for the session. - since records arrive chronologically and since we have a session window, then all call records for a phone number end up in the same partition and session (intended behavior). We can easily have many such phone call record groups with 100s of call records in them. The aggregate object (the avro object with array of records in it) can get updated 100s of times for the same phone number in the course of an hour or so. - we process billions of such call records a day - we can’t expect our clients to install massive clusters of 10s or 100s of nodes. We do need to make sure this processing is as efficient as possible. The session window bug was killing us. Much better with the fix Damian provided! Ara. On Mar 27, 2017, at 2:41 PM, Matthias J. Sax <matth...@confluent.io<mailto:matth...@confluent.io>> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 27, 2017 at 2:41:06 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> Ara, I assume your performance issue is most likely related to the fix Damian pointed out already. Couple of follow up comments: critical part of our pipeline involves grouping relevant records together Can you explain this a little better? The abstraction of a task does group data together already. Furthermore, if you get multiple tasks, those are independent units with regard to grouping as consecutive tasks are "connected" via repartitioning steps. Thus, even if we apply the task assignment as you ask for, I am not sure if this would change anything? Maybe you can give a small data example what behavior/data-co-location you need and what Streams provides for you. And for hot keys this can lead to sometimes 100s of records to get grouped together This is independent of task placement -- it a partitioning issue. If you want to work on that, you can provide a custom `Partioner` for the used Kafka producers (note, your external Producer writing to your Streams input topic might already generate "hot" keys, so you might need to use a custom partitioner there, too) Also "100s of records" does not sound much to me. Streams can process multiple hundredths of thousandth records per thread. That is the reason, why I think that the fix Damian pointed out will most likely fix your problem. -Matthias On 3/27/17 1:56 PM, Ara Ebrahimi wrote: Thanks for the response Mathias! The reason we want this exact task assignment to happen is that a critical part of our pipeline involves grouping relevant records together (that’s what the aggregate function in the topology is for). And for hot keys this can lead to sometimes 100s of records to get grouped together. Even worse, these records are session bound, we use session windows. Hence we see lots of activity around the store backing the aggregate function and even though we use SSD drives we’re not seeing the kind of performance we want to see. It seems like the aggregate function leads to lots of updates to these hot keys which lead to lots of rocksdb activity. Now there are many ways to fix this problem: - just don’t aggregate, create an algorithm which is not reliant on grouping/aggregating records. Not what we can do with our tight schedule right now. - do grouping/aggregating but employ n instances and rely on uniform distribution of these tasks. This is the easiest solution and what we expected to work but didn’t work as you can tell from this thread. We threw 4 instances at it but only 2 got used. - tune rocksdb? I tried this actually but it didn’t really help us much, aside from the fact that tuning rocksdb is very tricky. - use in-memory store instead? Unfortunately we have to use session windows for this aggregate function and apparently there’s no in-memory session store impl? I tried to create one but soon realized it’s too much work :) I looked at default PartitionAssigner code too, but that ain’t trivial either. So I’m a bit hopeless :( Ara. On Mar 27, 2017
Re: more uniform task assignment across kafka stream nodes
Holy s...! This lead to ~4-5x better performance. I’m gonna try this with more nodes and if performance improves almost linearly then we are good for now. Thanks! Ara. > On Mar 27, 2017, at 2:10 PM, Damian Guy <damian@gmail.com> wrote: > > Hi Ara, > > There is a performance issue in the 0.10.2 release of session windows. It > is fixed with this PR: https://github.com/apache/kafka/pull/2645 > You can work around this on 0.10.2 by calling the aggregate(..), reduce(..) > etc methods and supplying StateStoreSupplier with caching > disabled, i.e, by doing something like: > > final StateStoreSupplier sessionStore = > Stores.create(*"session-store-name"*) >.withKeys(Serdes.String()) >.withValues(Serdes.String()) >.persistent() >.sessionWindowed(TimeUnit.MINUTES.toMillis(7)) >.build(); > > > The fix has also been cherry-picked to the 0.10.2 branch, so you could > build from source and not have to create the StateStoreSupplier. > > Thanks, > Damian > > On Mon, 27 Mar 2017 at 21:56 Ara Ebrahimi <ara.ebrah...@argyledata.com> > wrote: > > Thanks for the response Mathias! > > The reason we want this exact task assignment to happen is that a critical > part of our pipeline involves grouping relevant records together (that’s > what the aggregate function in the topology is for). And for hot keys this > can lead to sometimes 100s of records to get grouped together. Even worse, > these records are session bound, we use session windows. Hence we see lots > of activity around the store backing the aggregate function and even though > we use SSD drives we’re not seeing the kind of performance we want to see. > It seems like the aggregate function leads to lots of updates to these hot > keys which lead to lots of rocksdb activity. > > Now there are many ways to fix this problem: > - just don’t aggregate, create an algorithm which is not reliant on > grouping/aggregating records. Not what we can do with our tight schedule > right now. > - do grouping/aggregating but employ n instances and rely on uniform > distribution of these tasks. This is the easiest solution and what we > expected to work but didn’t work as you can tell from this thread. We threw > 4 instances at it but only 2 got used. > - tune rocksdb? I tried this actually but it didn’t really help us much, > aside from the fact that tuning rocksdb is very tricky. > - use in-memory store instead? Unfortunately we have to use session windows > for this aggregate function and apparently there’s no in-memory session > store impl? I tried to create one but soon realized it’s too much work :) I > looked at default PartitionAssigner code too, but that ain’t trivial either. > > So I’m a bit hopeless :( > > Ara. > > On Mar 27, 2017, at 1:35 PM, Matthias J. Sax <matth...@confluent.io matth...@confluent.io>> wrote: > > > > > > > This message is for the designated recipient only and may contain > privileged, proprietary, or otherwise confidential information. If you have > received it in error, please notify the sender immediately and delete the > original. Any other use of the e-mail by you is prohibited. Thank you in > advance for your cooperation. > > > > From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io >>> > Subject: Re: more uniform task assignment across kafka stream nodes > Date: March 27, 2017 at 1:35:30 PM PDT > To: users@kafka.apache.org<mailto:users@kafka.apache.org> > Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> > > > Ara, > > thanks for the detailed information. > > If I parse this correctly, both instances run the same number of tasks > (12 each). That is all Streams promises. > > To come back to your initial question: > > Is there a way to tell kafka streams to uniformly assign partitions across > instances? If I have n kafka streams instances running, I want each to > handle EXACTLY 1/nth number of partitions. No dynamic task assignment > logic. Just dumb 1/n assignment. > > That is exactly what you get: each of you two instances get 24/2 = 12 > tasks assigned. That is dump 1/n assignment, isn't it? So my original > response was correct. > > However, I now understand better what you are actually meaning by your > question. Note that Streams does not distinguish "type" of tasks -- it > only sees 24 tasks and assigns those in a balanced way. > > Thus, currently there is no easy way to get the assignment you want to > have, except, you implement you own `PartitionAssignor`. > > This is the current implementation for 0.10.2 >
Re: more uniform task assignment across kafka stream nodes
Ara, I assume your performance issue is most likely related to the fix Damian pointed out already. Couple of follow up comments: > critical part of our pipeline involves grouping relevant records together Can you explain this a little better? The abstraction of a task does group data together already. Furthermore, if you get multiple tasks, those are independent units with regard to grouping as consecutive tasks are "connected" via repartitioning steps. Thus, even if we apply the task assignment as you ask for, I am not sure if this would change anything? Maybe you can give a small data example what behavior/data-co-location you need and what Streams provides for you. > And for hot keys this can lead to sometimes 100s of records to get grouped > together This is independent of task placement -- it a partitioning issue. If you want to work on that, you can provide a custom `Partioner` for the used Kafka producers (note, your external Producer writing to your Streams input topic might already generate "hot" keys, so you might need to use a custom partitioner there, too) Also "100s of records" does not sound much to me. Streams can process multiple hundredths of thousandth records per thread. That is the reason, why I think that the fix Damian pointed out will most likely fix your problem. -Matthias On 3/27/17 1:56 PM, Ara Ebrahimi wrote: > Thanks for the response Mathias! > > The reason we want this exact task assignment to happen is that a critical > part of our pipeline involves grouping relevant records together (that’s what > the aggregate function in the topology is for). And for hot keys this can > lead to sometimes 100s of records to get grouped together. Even worse, these > records are session bound, we use session windows. Hence we see lots of > activity around the store backing the aggregate function and even though we > use SSD drives we’re not seeing the kind of performance we want to see. It > seems like the aggregate function leads to lots of updates to these hot keys > which lead to lots of rocksdb activity. > > Now there are many ways to fix this problem: > - just don’t aggregate, create an algorithm which is not reliant on > grouping/aggregating records. Not what we can do with our tight schedule > right now. > - do grouping/aggregating but employ n instances and rely on uniform > distribution of these tasks. This is the easiest solution and what we > expected to work but didn’t work as you can tell from this thread. We threw 4 > instances at it but only 2 got used. > - tune rocksdb? I tried this actually but it didn’t really help us much, > aside from the fact that tuning rocksdb is very tricky. > - use in-memory store instead? Unfortunately we have to use session windows > for this aggregate function and apparently there’s no in-memory session store > impl? I tried to create one but soon realized it’s too much work :) I looked > at default PartitionAssigner code too, but that ain’t trivial either. > > So I’m a bit hopeless :( > > Ara. > > On Mar 27, 2017, at 1:35 PM, Matthias J. Sax > <matth...@confluent.io<mailto:matth...@confluent.io>> wrote: > > > > > > > This message is for the designated recipient only and may contain privileged, > proprietary, or otherwise confidential information. If you have received it > in error, please notify the sender immediately and delete the original. Any > other use of the e-mail by you is prohibited. Thank you in advance for your > cooperation. > > > > From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>> > Subject: Re: more uniform task assignment across kafka stream nodes > Date: March 27, 2017 at 1:35:30 PM PDT > To: users@kafka.apache.org<mailto:users@kafka.apache.org> > Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> > > > Ara, > > thanks for the detailed information. > > If I parse this correctly, both instances run the same number of tasks > (12 each). That is all Streams promises. > > To come back to your initial question: > > Is there a way to tell kafka streams to uniformly assign partitions across > instances? If I have n kafka streams instances running, I want each to handle > EXACTLY 1/nth number of partitions. No dynamic task assignment logic. Just > dumb 1/n assignment. > > That is exactly what you get: each of you two instances get 24/2 = 12 > tasks assigned. That is dump 1/n assignment, isn't it? So my original > response was correct. > > However, I now understand better what you are actually meaning by your > question. Note that Streams does not distinguish "
Re: more uniform task assignment across kafka stream nodes
Hi Ara, There is a performance issue in the 0.10.2 release of session windows. It is fixed with this PR: https://github.com/apache/kafka/pull/2645 You can work around this on 0.10.2 by calling the aggregate(..), reduce(..) etc methods and supplying StateStoreSupplier with caching disabled, i.e, by doing something like: final StateStoreSupplier sessionStore = Stores.create(*"session-store-name"*) .withKeys(Serdes.String()) .withValues(Serdes.String()) .persistent() .sessionWindowed(TimeUnit.MINUTES.toMillis(7)) .build(); The fix has also been cherry-picked to the 0.10.2 branch, so you could build from source and not have to create the StateStoreSupplier. Thanks, Damian On Mon, 27 Mar 2017 at 21:56 Ara Ebrahimi <ara.ebrah...@argyledata.com> wrote: Thanks for the response Mathias! The reason we want this exact task assignment to happen is that a critical part of our pipeline involves grouping relevant records together (that’s what the aggregate function in the topology is for). And for hot keys this can lead to sometimes 100s of records to get grouped together. Even worse, these records are session bound, we use session windows. Hence we see lots of activity around the store backing the aggregate function and even though we use SSD drives we’re not seeing the kind of performance we want to see. It seems like the aggregate function leads to lots of updates to these hot keys which lead to lots of rocksdb activity. Now there are many ways to fix this problem: - just don’t aggregate, create an algorithm which is not reliant on grouping/aggregating records. Not what we can do with our tight schedule right now. - do grouping/aggregating but employ n instances and rely on uniform distribution of these tasks. This is the easiest solution and what we expected to work but didn’t work as you can tell from this thread. We threw 4 instances at it but only 2 got used. - tune rocksdb? I tried this actually but it didn’t really help us much, aside from the fact that tuning rocksdb is very tricky. - use in-memory store instead? Unfortunately we have to use session windows for this aggregate function and apparently there’s no in-memory session store impl? I tried to create one but soon realized it’s too much work :) I looked at default PartitionAssigner code too, but that ain’t trivial either. So I’m a bit hopeless :( Ara. On Mar 27, 2017, at 1:35 PM, Matthias J. Sax <matth...@confluent.io> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io >> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 27, 2017 at 1:35:30 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> Ara, thanks for the detailed information. If I parse this correctly, both instances run the same number of tasks (12 each). That is all Streams promises. To come back to your initial question: Is there a way to tell kafka streams to uniformly assign partitions across instances? If I have n kafka streams instances running, I want each to handle EXACTLY 1/nth number of partitions. No dynamic task assignment logic. Just dumb 1/n assignment. That is exactly what you get: each of you two instances get 24/2 = 12 tasks assigned. That is dump 1/n assignment, isn't it? So my original response was correct. However, I now understand better what you are actually meaning by your question. Note that Streams does not distinguish "type" of tasks -- it only sees 24 tasks and assigns those in a balanced way. Thus, currently there is no easy way to get the assignment you want to have, except, you implement you own `PartitionAssignor`. This is the current implementation for 0.10.2 https://github.com/apache/kafka/blob/0.10.2/streams/src/main/java/org/apache/kafka/streams/processor/internals/StreamPartitionAssignor.java You can, if you wish write your own assignor and set it via StreamsConfig. However, be aware that this might be tricky to get right and also might have runtime implications with regard to rebalancing and state store recovery. We recently improve the current implementation to avoid costly task movements: https://issues.apache.org/jira/browse/KAFKA-4677 Thus, I would not recommend to implement an own `PartitionAssignor`. However, the root question is, why do you need this exact assignment you are looking for in the first place? Why is it "bad" if "types" of tasks are not distinguished? I would like to understand yo
Re: more uniform task assignment across kafka stream nodes
Thanks for the response Mathias! The reason we want this exact task assignment to happen is that a critical part of our pipeline involves grouping relevant records together (that’s what the aggregate function in the topology is for). And for hot keys this can lead to sometimes 100s of records to get grouped together. Even worse, these records are session bound, we use session windows. Hence we see lots of activity around the store backing the aggregate function and even though we use SSD drives we’re not seeing the kind of performance we want to see. It seems like the aggregate function leads to lots of updates to these hot keys which lead to lots of rocksdb activity. Now there are many ways to fix this problem: - just don’t aggregate, create an algorithm which is not reliant on grouping/aggregating records. Not what we can do with our tight schedule right now. - do grouping/aggregating but employ n instances and rely on uniform distribution of these tasks. This is the easiest solution and what we expected to work but didn’t work as you can tell from this thread. We threw 4 instances at it but only 2 got used. - tune rocksdb? I tried this actually but it didn’t really help us much, aside from the fact that tuning rocksdb is very tricky. - use in-memory store instead? Unfortunately we have to use session windows for this aggregate function and apparently there’s no in-memory session store impl? I tried to create one but soon realized it’s too much work :) I looked at default PartitionAssigner code too, but that ain’t trivial either. So I’m a bit hopeless :( Ara. On Mar 27, 2017, at 1:35 PM, Matthias J. Sax <matth...@confluent.io<mailto:matth...@confluent.io>> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 27, 2017 at 1:35:30 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> Ara, thanks for the detailed information. If I parse this correctly, both instances run the same number of tasks (12 each). That is all Streams promises. To come back to your initial question: Is there a way to tell kafka streams to uniformly assign partitions across instances? If I have n kafka streams instances running, I want each to handle EXACTLY 1/nth number of partitions. No dynamic task assignment logic. Just dumb 1/n assignment. That is exactly what you get: each of you two instances get 24/2 = 12 tasks assigned. That is dump 1/n assignment, isn't it? So my original response was correct. However, I now understand better what you are actually meaning by your question. Note that Streams does not distinguish "type" of tasks -- it only sees 24 tasks and assigns those in a balanced way. Thus, currently there is no easy way to get the assignment you want to have, except, you implement you own `PartitionAssignor`. This is the current implementation for 0.10.2 https://github.com/apache/kafka/blob/0.10.2/streams/src/main/java/org/apache/kafka/streams/processor/internals/StreamPartitionAssignor.java You can, if you wish write your own assignor and set it via StreamsConfig. However, be aware that this might be tricky to get right and also might have runtime implications with regard to rebalancing and state store recovery. We recently improve the current implementation to avoid costly task movements: https://issues.apache.org/jira/browse/KAFKA-4677 Thus, I would not recommend to implement an own `PartitionAssignor`. However, the root question is, why do you need this exact assignment you are looking for in the first place? Why is it "bad" if "types" of tasks are not distinguished? I would like to understand your requirement better -- it might be worth to improve Streams here. -Matthias On 3/27/17 12:57 PM, Ara Ebrahimi wrote: Hi, So, I simplified the topology by making sure we have only 1 source topic. Now I have 1 source topic, 8 partitions, 2 instances. And here’s how the topology looks like: instance 1: KafkaStreams processID: 48b58bc0-f600-4ec8-bc92-8cb3ea081aac StreamsThread appId: mar-23-modular StreamsThread clientId: mar-23-modular StreamsThread threadId: StreamThread-1 Active tasks: StreamsTask taskId: 0_3 ProcessorTopology: KSTREAM-SOURCE-00: topics: [activities-avro-or] children: [KSTREAM-FILTER-01] KSTREAM-FILTER-01: children: [KSTREAM-MAP-02] KSTREAM-MAP-02: children: [KS
Re: more uniform task assignment across kafka stream nodes
ALUES-06]
> KSTREAM-MAPVALUES-06:
> children: [KSTREAM-FLATMAPVALUES-07]
> KSTREAM-FLATMAPVALUES-000007:
> children: [KSTREAM-MAP-000008]
> KSTREAM-MAP-08:
> children: [KSTREAM-FILTER-11]
> KSTREAM-FILTER-11:
> children: [KSTREAM-SINK-10]
> KSTREAM-SINK-10:
> topic: activities-by-phone-store-or-repartition
> KSTREAM-BRANCHCHILD-05:
> Partitions [activities-avro-or-5]
> Standby tasks:
>
>
> activities-avro-or is input topic. ml-features-avro-or is output topic. In
> the middle we have an aggregate (activities-by-phone-store-or-repartition).
>
> On instance 1 I see 3 tasks for activities-avro-or and on instance 2 I see 5.
> Bad.
>
> On instance 1 see 4 tasks for ml-features-avro-or. And 4 on instance 2. Good.
>
> On instance 1 see 5 tasks for activities-by-phone-store-or-repartition. And 3
> on instance 2. Bad.
>
> As I said in terms of offsets for all these partitions I see uniform
> distribution, so we’re not dealing with a bad key scenario.
>
> Ara.
>
> On Mar 25, 2017, at 6:43 PM, Matthias J. Sax
> <matth...@confluent.io<mailto:matth...@confluent.io>> wrote:
>
>
>
>
>
>
> This message is for the designated recipient only and may contain privileged,
> proprietary, or otherwise confidential information. If you have received it
> in error, please notify the sender immediately and delete the original. Any
> other use of the e-mail by you is prohibited. Thank you in advance for your
> cooperation.
>
>
>
> From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>>
> Subject: Re: more uniform task assignment across kafka stream nodes
> Date: March 25, 2017 at 6:43:12 PM PDT
> To: users@kafka.apache.org<mailto:users@kafka.apache.org>
> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>>
>
>
> Please share the rest of your topology code (without any UDFs / business
> logic). Otherwise, I cannot give further advice.
>
> -Matthias
>
>
> On 3/25/17 6:08 PM, Ara Ebrahimi wrote:
> Via:
>
> builder.stream("topic1");
> builder.stream("topic2");
> builder.stream("topic3”);
>
> These are different kinds of topics consuming different avro objects.
>
> Ara.
>
> On Mar 25, 2017, at 6:04 PM, Matthias J. Sax
> <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>>
> wrote:
>
>
>
>
>
>
> This message is for the designated recipient only and may contain privileged,
> proprietary, or otherwise confidential information. If you have received it
> in error, please notify the sender immediately and delete the original. Any
> other use of the e-mail by you is prohibited. Thank you in advance for your
> cooperation.
>
>
>
> From: "Matthias J. Sax"
> <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>>
> Subject: Re: more uniform task assignment across kafka stream nodes
> Date: March 25, 2017 at 6:04:30 PM PDT
> To:
> users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>
> Reply-To:
> <users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>>
>
>
> Ara,
>
> How do you consume your topics? Via
>
> builder.stream("topic1", "topic2", "topic3);
>
> or via
>
> builder.stream("topic1");
> builder.stream("topic2");
> builder.stream("topic3");
>
> Both and handled differently with regard to creating tasks (partition to
> task assignment also depends on you downstream code though).
>
> If this does not help, can you maybe share the structure of processing?
> To dig deeper, we would need to know the topology DAG.
>
>
> -Matthias
>
>
> On 3/25/17 5:56 PM, Ara Ebrahimi wrote:
> Mathias,
>
> This apparently happens because we have more than 1 source topic. We have 3
> source topics in the same application. So it seems like the task assignment
> algorithm creates topologies not for one specific topic at a time but the
> total partitions across all source topics consumed in an application
> instance. Because we have some code dependencies between these 3 source
> topics we can’t separate them into 3 applications at this time. Hence the
> reason I want to get the task assignment algorithm basically do a uniform and
> simple task assignment PER s
Re: more uniform task assignment across kafka stream nodes
:
StreamsThread appId: mar-23-modular
StreamsThread clientId: mar-23-modular
StreamsThread threadId: StreamThread-8
Active tasks:
StreamsTask taskId: 0_5
ProcessorTopology:
KSTREAM-SOURCE-00:
topics: [activities-avro-or]
children: [KSTREAM-FILTER-01]
KSTREAM-FILTER-01:
children: [KSTREAM-MAP-02]
KSTREAM-MAP-02:
children: [KSTREAM-BRANCH-03]
KSTREAM-BRANCH-03:
children: [KSTREAM-BRANCHCHILD-04, KSTREAM-BRANCHCHILD-05]
KSTREAM-BRANCHCHILD-04:
children: [KSTREAM-MAPVALUES-06]
KSTREAM-MAPVALUES-06:
children: [KSTREAM-FLATMAPVALUES-07]
KSTREAM-FLATMAPVALUES-07:
children: [KSTREAM-MAP-08]
KSTREAM-MAP-08:
children: [KSTREAM-FILTER-11]
KSTREAM-FILTER-11:
children: [KSTREAM-SINK-10]
KSTREAM-SINK-10:
topic: activities-by-phone-store-or-repartition
KSTREAM-BRANCHCHILD-05:
Partitions [activities-avro-or-5]
Standby tasks:
activities-avro-or is input topic. ml-features-avro-or is output topic. In the
middle we have an aggregate (activities-by-phone-store-or-repartition).
On instance 1 I see 3 tasks for activities-avro-or and on instance 2 I see 5.
Bad.
On instance 1 see 4 tasks for ml-features-avro-or. And 4 on instance 2. Good.
On instance 1 see 5 tasks for activities-by-phone-store-or-repartition. And 3
on instance 2. Bad.
As I said in terms of offsets for all these partitions I see uniform
distribution, so we’re not dealing with a bad key scenario.
Ara.
On Mar 25, 2017, at 6:43 PM, Matthias J. Sax
<matth...@confluent.io<mailto:matth...@confluent.io>> wrote:
This message is for the designated recipient only and may contain privileged,
proprietary, or otherwise confidential information. If you have received it in
error, please notify the sender immediately and delete the original. Any other
use of the e-mail by you is prohibited. Thank you in advance for your
cooperation.
From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>>
Subject: Re: more uniform task assignment across kafka stream nodes
Date: March 25, 2017 at 6:43:12 PM PDT
To: users@kafka.apache.org<mailto:users@kafka.apache.org>
Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>>
Please share the rest of your topology code (without any UDFs / business
logic). Otherwise, I cannot give further advice.
-Matthias
On 3/25/17 6:08 PM, Ara Ebrahimi wrote:
Via:
builder.stream("topic1");
builder.stream("topic2");
builder.stream("topic3”);
These are different kinds of topics consuming different avro objects.
Ara.
On Mar 25, 2017, at 6:04 PM, Matthias J. Sax
<matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>>
wrote:
This message is for the designated recipient only and may contain privileged,
proprietary, or otherwise confidential information. If you have received it in
error, please notify the sender immediately and delete the original. Any other
use of the e-mail by you is prohibited. Thank you in advance for your
cooperation.
From: "Matthias J. Sax"
<matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>>
Subject: Re: more uniform task assignment across kafka stream nodes
Date: March 25, 2017 at 6:04:30 PM PDT
To:
users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>
Reply-To:
<users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>>
Ara,
How do you consume your topics? Via
builder.stream("topic1", "topic2", "topic3);
or via
builder.stream("topic1");
builder.stream("topic2");
builder.stream("topic3");
Both and handled differently with regard to creating tasks (partition to
task assignment also depends on you downstream code though).
If this does not help, can you maybe share the structure of processing?
To dig deeper, we would need to know the topology DAG.
-Matthias
On 3/25/17 5:56 PM, Ara Ebrahimi wrote:
Mathias,
This apparently happens because we have more than 1 source topic. We have 3
source topics in the same application. So it seems like the task assignment
algorithm creates topologies not for one specific topic at a time but the total
partitions across all source topics consumed in an application instance.
Because we have some code dependencies between these 3 source topics we can’t
separate them into 3 applications at this time. Hence the reason I want to get
the task assignment algorithm basically do a uniform and simple task assignment
PER source topic.
Ara.
On Mar 25, 2017, at 5:21 PM, Matthias J. Sax
<matth...@confluent.io<
Re: more uniform task assignment across kafka stream nodes
Please share the rest of your topology code (without any UDFs / business
logic). Otherwise, I cannot give further advice.
-Matthias
On 3/25/17 6:08 PM, Ara Ebrahimi wrote:
> Via:
>
> builder.stream("topic1");
> builder.stream("topic2");
> builder.stream("topic3”);
>
> These are different kinds of topics consuming different avro objects.
>
> Ara.
>
> On Mar 25, 2017, at 6:04 PM, Matthias J. Sax
> <matth...@confluent.io<mailto:matth...@confluent.io>> wrote:
>
>
>
>
>
>
> This message is for the designated recipient only and may contain privileged,
> proprietary, or otherwise confidential information. If you have received it
> in error, please notify the sender immediately and delete the original. Any
> other use of the e-mail by you is prohibited. Thank you in advance for your
> cooperation.
>
> ____________
>
> From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>>
> Subject: Re: more uniform task assignment across kafka stream nodes
> Date: March 25, 2017 at 6:04:30 PM PDT
> To: users@kafka.apache.org<mailto:users@kafka.apache.org>
> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>>
>
>
> Ara,
>
> How do you consume your topics? Via
>
> builder.stream("topic1", "topic2", "topic3);
>
> or via
>
> builder.stream("topic1");
> builder.stream("topic2");
> builder.stream("topic3");
>
> Both and handled differently with regard to creating tasks (partition to
> task assignment also depends on you downstream code though).
>
> If this does not help, can you maybe share the structure of processing?
> To dig deeper, we would need to know the topology DAG.
>
>
> -Matthias
>
>
> On 3/25/17 5:56 PM, Ara Ebrahimi wrote:
> Mathias,
>
> This apparently happens because we have more than 1 source topic. We have 3
> source topics in the same application. So it seems like the task assignment
> algorithm creates topologies not for one specific topic at a time but the
> total partitions across all source topics consumed in an application
> instance. Because we have some code dependencies between these 3 source
> topics we can’t separate them into 3 applications at this time. Hence the
> reason I want to get the task assignment algorithm basically do a uniform and
> simple task assignment PER source topic.
>
> Ara.
>
> On Mar 25, 2017, at 5:21 PM, Matthias J. Sax
> <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>>
> wrote:
>
>
>
>
>
>
> This message is for the designated recipient only and may contain privileged,
> proprietary, or otherwise confidential information. If you have received it
> in error, please notify the sender immediately and delete the original. Any
> other use of the e-mail by you is prohibited. Thank you in advance for your
> cooperation.
>
>
>
> From: "Matthias J. Sax"
> <matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>>
> Subject: Re: more uniform task assignment across kafka stream nodes
> Date: March 25, 2017 at 5:21:47 PM PDT
> To:
> users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>
> Reply-To:
> <users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>>
>
>
> Hi,
>
> I am wondering why this happens in the first place. Streams,
> load-balanced over all running instances, and each instance should be
> the same number of tasks (and thus partitions) assigned.
>
> What is the overall assignment? Do you have StandyBy tasks configured?
> What version do you use?
>
>
> -Matthias
>
>
> On 3/24/17 8:09 PM, Ara Ebrahimi wrote:
> Hi,
>
> Is there a way to tell kafka streams to uniformly assign partitions across
> instances? If I have n kafka streams instances running, I want each to handle
> EXACTLY 1/nth number of partitions. No dynamic task assignment logic. Just
> dumb 1/n assignment.
>
> Here’s our scenario. Lets say we have an “source" topic with 8 partitions. We
> also have 2 kafka streams instances. Each instances get assigned to handle 4
> “source" topic partitions. BUT then we do a few maps and an aggregate. So
> data gets shuffled around. The map function uniformly distributes these
> across all partitions (I can verify that by looking at the partition
> offsets). After the m
Re: more uniform task assignment across kafka stream nodes
Via:
builder.stream("topic1");
builder.stream("topic2");
builder.stream("topic3”);
These are different kinds of topics consuming different avro objects.
Ara.
On Mar 25, 2017, at 6:04 PM, Matthias J. Sax
<matth...@confluent.io<mailto:matth...@confluent.io>> wrote:
This message is for the designated recipient only and may contain privileged,
proprietary, or otherwise confidential information. If you have received it in
error, please notify the sender immediately and delete the original. Any other
use of the e-mail by you is prohibited. Thank you in advance for your
cooperation.
From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>>
Subject: Re: more uniform task assignment across kafka stream nodes
Date: March 25, 2017 at 6:04:30 PM PDT
To: users@kafka.apache.org<mailto:users@kafka.apache.org>
Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>>
Ara,
How do you consume your topics? Via
builder.stream("topic1", "topic2", "topic3);
or via
builder.stream("topic1");
builder.stream("topic2");
builder.stream("topic3");
Both and handled differently with regard to creating tasks (partition to
task assignment also depends on you downstream code though).
If this does not help, can you maybe share the structure of processing?
To dig deeper, we would need to know the topology DAG.
-Matthias
On 3/25/17 5:56 PM, Ara Ebrahimi wrote:
Mathias,
This apparently happens because we have more than 1 source topic. We have 3
source topics in the same application. So it seems like the task assignment
algorithm creates topologies not for one specific topic at a time but the total
partitions across all source topics consumed in an application instance.
Because we have some code dependencies between these 3 source topics we can’t
separate them into 3 applications at this time. Hence the reason I want to get
the task assignment algorithm basically do a uniform and simple task assignment
PER source topic.
Ara.
On Mar 25, 2017, at 5:21 PM, Matthias J. Sax
<matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>>
wrote:
This message is for the designated recipient only and may contain privileged,
proprietary, or otherwise confidential information. If you have received it in
error, please notify the sender immediately and delete the original. Any other
use of the e-mail by you is prohibited. Thank you in advance for your
cooperation.
____________
From: "Matthias J. Sax"
<matth...@confluent.io<mailto:matth...@confluent.io><mailto:matth...@confluent.io>>
Subject: Re: more uniform task assignment across kafka stream nodes
Date: March 25, 2017 at 5:21:47 PM PDT
To:
users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>
Reply-To:
<users@kafka.apache.org<mailto:users@kafka.apache.org><mailto:users@kafka.apache.org>>
Hi,
I am wondering why this happens in the first place. Streams,
load-balanced over all running instances, and each instance should be
the same number of tasks (and thus partitions) assigned.
What is the overall assignment? Do you have StandyBy tasks configured?
What version do you use?
-Matthias
On 3/24/17 8:09 PM, Ara Ebrahimi wrote:
Hi,
Is there a way to tell kafka streams to uniformly assign partitions across
instances? If I have n kafka streams instances running, I want each to handle
EXACTLY 1/nth number of partitions. No dynamic task assignment logic. Just dumb
1/n assignment.
Here’s our scenario. Lets say we have an “source" topic with 8 partitions. We
also have 2 kafka streams instances. Each instances get assigned to handle 4
“source" topic partitions. BUT then we do a few maps and an aggregate. So data
gets shuffled around. The map function uniformly distributes these across all
partitions (I can verify that by looking at the partition offsets). After the
map what I notice by looking at the topology is that one kafka streams instance
get assigned to handle say 2 aggregate repartition topics and the other one
gets assigned 6. Even worse, on bigger clusters (say 4 instances) we see say 2
nodes gets assigned downstream aggregate repartition topics and 2 other nodes
assigned NOTHING to handle.
Ara.
This message is for the designated recipient only and may contain privileged,
proprietary, or otherwise confidential information. If you have received it in
error, please notify the sender immediately and delete the original. Any other
use of the e-mail by you is prohibited. Thank you in advance for your
cooperation.
_
Re: more uniform task assignment across kafka stream nodes
Ara,
How do you consume your topics? Via
> builder.stream("topic1", "topic2", "topic3);
or via
> builder.stream("topic1");
> builder.stream("topic2");
> builder.stream("topic3");
Both and handled differently with regard to creating tasks (partition to
task assignment also depends on you downstream code though).
If this does not help, can you maybe share the structure of processing?
To dig deeper, we would need to know the topology DAG.
-Matthias
On 3/25/17 5:56 PM, Ara Ebrahimi wrote:
> Mathias,
>
> This apparently happens because we have more than 1 source topic. We have 3
> source topics in the same application. So it seems like the task assignment
> algorithm creates topologies not for one specific topic at a time but the
> total partitions across all source topics consumed in an application
> instance. Because we have some code dependencies between these 3 source
> topics we can’t separate them into 3 applications at this time. Hence the
> reason I want to get the task assignment algorithm basically do a uniform and
> simple task assignment PER source topic.
>
> Ara.
>
> On Mar 25, 2017, at 5:21 PM, Matthias J. Sax
> <matth...@confluent.io<mailto:matth...@confluent.io>> wrote:
>
>
>
>
>
>
> This message is for the designated recipient only and may contain privileged,
> proprietary, or otherwise confidential information. If you have received it
> in error, please notify the sender immediately and delete the original. Any
> other use of the e-mail by you is prohibited. Thank you in advance for your
> cooperation.
>
> ____________
>
> From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>>
> Subject: Re: more uniform task assignment across kafka stream nodes
> Date: March 25, 2017 at 5:21:47 PM PDT
> To: users@kafka.apache.org<mailto:users@kafka.apache.org>
> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>>
>
>
> Hi,
>
> I am wondering why this happens in the first place. Streams,
> load-balanced over all running instances, and each instance should be
> the same number of tasks (and thus partitions) assigned.
>
> What is the overall assignment? Do you have StandyBy tasks configured?
> What version do you use?
>
>
> -Matthias
>
>
> On 3/24/17 8:09 PM, Ara Ebrahimi wrote:
> Hi,
>
> Is there a way to tell kafka streams to uniformly assign partitions across
> instances? If I have n kafka streams instances running, I want each to handle
> EXACTLY 1/nth number of partitions. No dynamic task assignment logic. Just
> dumb 1/n assignment.
>
> Here’s our scenario. Lets say we have an “source" topic with 8 partitions. We
> also have 2 kafka streams instances. Each instances get assigned to handle 4
> “source" topic partitions. BUT then we do a few maps and an aggregate. So
> data gets shuffled around. The map function uniformly distributes these
> across all partitions (I can verify that by looking at the partition
> offsets). After the map what I notice by looking at the topology is that one
> kafka streams instance get assigned to handle say 2 aggregate repartition
> topics and the other one gets assigned 6. Even worse, on bigger clusters (say
> 4 instances) we see say 2 nodes gets assigned downstream aggregate
> repartition topics and 2 other nodes assigned NOTHING to handle.
>
> Ara.
>
>
>
>
>
> This message is for the designated recipient only and may contain privileged,
> proprietary, or otherwise confidential information. If you have received it
> in error, please notify the sender immediately and delete the original. Any
> other use of the e-mail by you is prohibited. Thank you in advance for your
> cooperation.
>
>
>
>
>
>
>
>
>
>
>
>
> This message is for the designated recipient only and may contain privileged,
> proprietary, or otherwise confidential information. If you have received it
> in error, please notify the sender immediately and delete the original. Any
> other use of the e-mail by you is prohibited. Thank you in advance for your
> cooperation.
>
>
>
signature.asc
Description: OpenPGP digital signature
Re: more uniform task assignment across kafka stream nodes
Mathias, This apparently happens because we have more than 1 source topic. We have 3 source topics in the same application. So it seems like the task assignment algorithm creates topologies not for one specific topic at a time but the total partitions across all source topics consumed in an application instance. Because we have some code dependencies between these 3 source topics we can’t separate them into 3 applications at this time. Hence the reason I want to get the task assignment algorithm basically do a uniform and simple task assignment PER source topic. Ara. On Mar 25, 2017, at 5:21 PM, Matthias J. Sax <matth...@confluent.io<mailto:matth...@confluent.io>> wrote: This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. From: "Matthias J. Sax" <matth...@confluent.io<mailto:matth...@confluent.io>> Subject: Re: more uniform task assignment across kafka stream nodes Date: March 25, 2017 at 5:21:47 PM PDT To: users@kafka.apache.org<mailto:users@kafka.apache.org> Reply-To: <users@kafka.apache.org<mailto:users@kafka.apache.org>> Hi, I am wondering why this happens in the first place. Streams, load-balanced over all running instances, and each instance should be the same number of tasks (and thus partitions) assigned. What is the overall assignment? Do you have StandyBy tasks configured? What version do you use? -Matthias On 3/24/17 8:09 PM, Ara Ebrahimi wrote: Hi, Is there a way to tell kafka streams to uniformly assign partitions across instances? If I have n kafka streams instances running, I want each to handle EXACTLY 1/nth number of partitions. No dynamic task assignment logic. Just dumb 1/n assignment. Here’s our scenario. Lets say we have an “source" topic with 8 partitions. We also have 2 kafka streams instances. Each instances get assigned to handle 4 “source" topic partitions. BUT then we do a few maps and an aggregate. So data gets shuffled around. The map function uniformly distributes these across all partitions (I can verify that by looking at the partition offsets). After the map what I notice by looking at the topology is that one kafka streams instance get assigned to handle say 2 aggregate repartition topics and the other one gets assigned 6. Even worse, on bigger clusters (say 4 instances) we see say 2 nodes gets assigned downstream aggregate repartition topics and 2 other nodes assigned NOTHING to handle. Ara. This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation. This message is for the designated recipient only and may contain privileged, proprietary, or otherwise confidential information. If you have received it in error, please notify the sender immediately and delete the original. Any other use of the e-mail by you is prohibited. Thank you in advance for your cooperation.
Re: more uniform task assignment across kafka stream nodes
Hi, I am wondering why this happens in the first place. Streams, load-balanced over all running instances, and each instance should be the same number of tasks (and thus partitions) assigned. What is the overall assignment? Do you have StandyBy tasks configured? What version do you use? -Matthias On 3/24/17 8:09 PM, Ara Ebrahimi wrote: > Hi, > > Is there a way to tell kafka streams to uniformly assign partitions across > instances? If I have n kafka streams instances running, I want each to handle > EXACTLY 1/nth number of partitions. No dynamic task assignment logic. Just > dumb 1/n assignment. > > Here’s our scenario. Lets say we have an “source" topic with 8 partitions. We > also have 2 kafka streams instances. Each instances get assigned to handle 4 > “source" topic partitions. BUT then we do a few maps and an aggregate. So > data gets shuffled around. The map function uniformly distributes these > across all partitions (I can verify that by looking at the partition > offsets). After the map what I notice by looking at the topology is that one > kafka streams instance get assigned to handle say 2 aggregate repartition > topics and the other one gets assigned 6. Even worse, on bigger clusters (say > 4 instances) we see say 2 nodes gets assigned downstream aggregate > repartition topics and 2 other nodes assigned NOTHING to handle. > > Ara. > > > > > > This message is for the designated recipient only and may contain privileged, > proprietary, or otherwise confidential information. If you have received it > in error, please notify the sender immediately and delete the original. Any > other use of the e-mail by you is prohibited. Thank you in advance for your > cooperation. > > > signature.asc Description: OpenPGP digital signature
