Hi Tai, OK, thanks for confirming. I understand that streaming shuffle is cheaper than batch-spill shuffle, but nevertheless may be unacceptable in large volume applications - it is still a network shuffle and that's the biggest part of the cost.
Now on the point of the trade off between load-balancing the initial work in Flink vs. one-less-shuffle. The problem with it is that this trade-off doesn't exist. Using physical partition ID will not only preserve existing co-partitioning but will also produce more balanced initial work for Flink. This is because partitioning at the transport layer (i.e. Kafka or other) is already balanced with respect to the actual volumes within each topic. Putting all topic-partitions in a flat list and taking index for assignment, you have a good chance that all partitions from say high-volume topic will end up in the same flink task and a set of very sparse topics, which may have the same number of partitions for whatever reasons will go to the remainder of flink tasks leading to a massive skew in balancing of the work. On the point of design philosophy of partitioning I can only give you my opinion because Kafka, and probably any distributed data infrastructure will impose partitioning but leave the actual partitioning scheme to the application. The particular partitioning scheme is indeed a concern of stream-processing layer, however, it is most efficient if it is provided by the transport layer. From the experience of large landscapes of data sources and stream-processing applications where a given partitioning scheme may be useful to more than just a single workflow, it is quite essential for the stream-processors to have an understanding which partitioning schemes exist and can be leveraged for their computation. That is how for example Samza co-exists with Kafka for creating real-time dynamic queries. Intermediate shuffles can be exposed via Kafka and several queries as well as any stream-processing framework can re-use this shuffle for their own computation. I guess, Flink is trying encapsulate this whole process within its own optimizer entirely, which is not a bad idea and can lead to simplification in how such a system would be architected, but it should still accept that partitioning schemes that exist at the transport layer are more efficient and better balanced to start with. On 6/21/16, Tai Gordon <tzuli...@gmail.com> wrote: > Hi Michal, > > I see, thanks for the description. I think you’ve definitely raised an > interesting point. > Yes, there may be unnecessary shuffle in this case (the partitionCustom on > consumed DataStream doesn’t override the assignPartitions in the Kafka > connector; custom partitioners are applied after the UDF transformation). > It’s noteworthy that this data exchange between tasks will be a “streaming > shuffle”, so I think the cost won’t be as high compared to other streaming > systems that require shuffle spills. I’m not familiar with how streaming > data shuffle in Flink works exactly though, I’m simply referring to this > doc [1], so anyone who’s knowledgable of this part please correct me if I’m > wrong. > > On the other hand, I’m curious about how the design philosophy of > partitioning for systems like Kafka (and AWS Kinesis like-wise), and how it > should work with Flink. > As far as I know, Kafka’s partitioning is meant for distributing messages > for scale and load-balancing, as well as for retaining even-order for > messages with the same partition key. It doesn’t provide information about > whether 2 topics are actually co-partitioned. If consuming streaming > frameworks like Flink is to take existing co-partitioning of messages in > Kafka into account, it’d have to assume this every time, and, like you > mentioned, use the partition id for assigning and sacrifice work balance > between subtasks. Perhaps we could make it an option in the connector, but > I’m in doubt if it is reasonable for users to sacrifice work balance and > initial throughput for the excessive streaming shuffle. Flink provides a > lot of ways for key extraction on data streams, so for complex topologies > there will most likely be shuffle downstream anyways. > > Regards, > Gordon > > > [1] > https://cwiki.apache.org/confluence/display/FLINK/Data+exchange+between+tasks > > > On June 21, 2016 at 4:02:01 AM, Michal Hariš (michal.har...@gmail.com) > wrote: > > Hi Tai, I was referring to co-partitioning, not co-location of leaders, > i.e. multiple topics that share the same partitioning scheme. By example, > say I have 2 topics which share the same keyspace and which are produced by > something other than Flink using identical partitioner. The data in these 2 > topics is already co-partitioned and shouldn't require any shuffle for > aggregations by key. From the partition assignment code and your > explanation (and the docs) I understand that the kafka connector will > assign the partitions to operator subtasks by iterating over the list, > thereby most likely breaking the existing co-partitioning, and so even if I > provide partitionCustom with the exact same code that was used to partition > the data in the external producer, any aggregation by the existing message > key will have to still incur the unnecessary shuffle. Correct ? Or does > partitionCustom on the data stream somehow override the behaviour of > assignPartitions on the source ? > > > On Mon, Jun 20, 2016 at 11:01 AM, Tai Gordon <tzuli...@gmail.com> wrote: > >> Hi Michal, >> >> Whether or not the external system's partitioning scheme is referenced > when >> assigning >> partitions to the consumer parallel subtasks depends on the > implementation >> of each connector / source. >> >> First, clarification on “co-partitioning": from your context I’m assuming >> you’re referring to co-location of Kafka partition leaders? If so, as you >> correctly identified, the current Kafka consumer connecter does not take >> the co-location of Kafka partitions into account when assigning > partitions. >> To exploit this, though, even if the deterministic assignment takes > leader >> location into account, subtasks will also need to be co-located with the >> leaders. As far as I know, there’s currently no way for subtask’s to > access >> their location info at runtime. >> On the other hand, if you’re referring to “co-partitioning the data" >> consumed from Kafka, you can use a custom partitioner on the consumed > data >> stream: >> >> > https://ci.apache.org/projects/flink/flink-docs-master/apis/streaming/#physical-partitioning >> . >> >> Hope this helps! >> >> Regards, >> Gordon >> >> On June 16, 2016 at 7:50:05 PM, Michal Hariš (michal.har...@gmail.com) >> wrote: >> >> Hi, I was recently looking into a kafka connector issue (FLINK-4023 / >> FLINK-4069), when it was pointed out that partition assignment will not > be >> deterministic if the partition discovery is imply moved to the open() >> method. In the assignPartitions of FlinkKafkaConsumerBase a modulo on the >> __index__ of the total list of topic-partitions subscribed to is used. It >> is clear that calling it from the open() method in each task can produce >> different lists and so some partitions can be consumed multiple times > while >> others not consumed at all. In a related discussion it was suggested to >> build this list in a deterministic way so that each partitions sees the >> same index for the same topic-partition. This would work for the issues >> above, but it highlighted for me another issue which relates to partition >> assignment itself - hence starting a different thread. >> >> It don't understand at this point the way how Flink does co-group on >> multiple topics but having worked in the Kafka zone for a number of > years, >> ignoring the physical partition id which is deterministic at the Kafka >> cluster level, and using a transient list (even if it is constructed >> deterministically) means that co-partitioning cannot be exploited for a >> straight co-group and Flink has to always do its own shuffle. I think > using >> getPartition() on each topic-partition instead of list index in the >> assignPartition is necessary, even if it may result in an unbalanced work >> distribution among Flink consumer instances. But it seems to me that in >> Flink, the partitioning schemes that exist outside its runtime are > ignored. >> Is it because any source outside Flink's realm is treated as to be > imported >> and no partitioning is assumed for simplicity/control or is it because > this >> is expected to produce even load-balancing of work? What else am I > missing? >> >> Michal >> >
