Hi Ken & David, Thanks for following up. I've responded to your questions below.
If the number of unique keys isn’t huge, I could think of yet another > helicopter stunt that you could try :) Unfortunately the number of keys in our case is huge, they're unique per handful of events. If your data are already pre-partitioned and the partitioning matches (hash > partitioning on the JAVA representation of the key yielded by the > KeySelector), you can use `reinterpretAsKeyedStream` [1] to skip the > shuffle. That comes with the additional constraints that Ken mentioned, correct? It could break immediately in cases if a key comes through on a different partition, or if the number of partitions happen to change? I'm concerned about that for our use case as we don't have 100% control of the upstream data source. I feel you'd be blocked by the state access downstream (with RocksDB). Are > you sure it isn't the case? Yes, you are right that state access is also a limiting factor and some optimizations to limit that have helped quite a bit (both in our implementation and in using local SSDs for rocksdb). One other path we looked at is using memory-backed volumes for rocksdb, but ran into a limitation that we cannot configure Flink's process memory lower than the k8s container memory, leading to OOMs. More details at https://stackoverflow.com/questions/74118022/flink-pods-ooming-using-memory-backed-volume-with-k8s-operator . I don't have a dashboard currently to immediately point to data shuffling as the primary bottleneck, but I thought it could be a huge optimization if we can tell Flink to take advantage of the pre-partitioned datasource, given we're shuffling near 1 Gb/sec right now. I can see that the join is causing the backpressure on the sources though, and figured that network and state acces would be the two primary contributors there. Let me know if you have any good debugging tools to narrow in on this more. Thanks, Tommy On Mon, Mar 6, 2023 at 4:42 AM David Morávek <d...@apache.org> wrote: > Using an operator state for a stateful join isn't great because it's meant > to hold only a minimal state related to the operator (e.g., partition > tracking). > > If your data are already pre-partitioned and the partitioning matches > (hash partitioning on the JAVA representation of the key yielded by the > KeySelector), you can use `reinterpretAsKeyedStream` [1] to skip the > shuffle. > > > What we see is that the join step causes backpressure on the kafka > sources and lag slowly starts to accumulate. > > I feel you'd be blocked by the state access downstream (with RocksDB). Are > you sure it isn't the case? > > [1] > https://javadoc.io/static/org.apache.flink/flink-streaming-java/1.16.1/org/apache/flink/streaming/api/datastream/DataStreamUtils.html#reinterpretAsKeyedStream-org.apache.flink.streaming.api.datastream.DataStream-org.apache.flink.api.java.functions.KeySelector- > > Best, > D. > > On Sun, Mar 5, 2023 at 5:31 AM Ken Krugler <kkrugler_li...@transpac.com> > wrote: > >> Hi Tommy, >> >> To use stateful timers, you need to have a keyed stream, which gets >> tricky when you’re trying to avoid network traffic caused by the keyBy() >> >> If the number of unique keys isn’t huge, I could think of yet another >> helicopter stunt that you could try :) >> >> It’s possible to calculate a composite key, based on the “real” key and a >> synthetic value, that will wind up on in the same slot where you’re doing >> this calculation. >> >> So that would let you create a keyed stream which would have >> serialization/deserialization cost, but wouldn’t actually go through the >> network stack. >> >> Since the composite key generation is deterministic, you can do the same >> thing on both streams, and join on the composite key. >> >> You’d want to cache the mapping from the real key to the synthetic value, >> to avoid doing this calculation for every record. >> >> If that sounds promising, lmk and I can post some code. >> >> — Ken >> >> >> On Mar 4, 2023, at 12:37 PM, Tommy May <tvma...@gmail.com> wrote: >> >> Hello Ken, >> >> Thanks for the quick response! That is an interesting workaround. In our >> case though we are using a CoProcessFunction with stateful timers. Is there >> a similar workaround path available in that case? The one possible way I >> could find required partitioning data in kafka in a very specific way >> to match what Flink's keyBy is doing, and that it'd have additional >> constraints to the method you described that would be difficult to handle >> in a prod environment where we don't have full control over the producers & >> input topics. >> >> Regarding the addition of a more flexible way to take advantage of >> pre-partitioned sources like in FLIP-186, would you suggest I forward this >> chain over to the dev Flink mailing list? >> >> Thanks, >> Tommy >> >> >> >> On Sat, Mar 4, 2023 at 11:32 AM Ken Krugler <kkrugler_li...@transpac.com> >> wrote: >> >>> Hi Tommy, >>> >>> I believe there is a way to make this work currently, but with lots of >>> caveats and constraints. >>> >>> This assumes you want to avoid any network shuffle. >>> >>> 1. Both topics have names that return the same value for >>> ((topicName.hashCode() * 31) & 0x7FFFF) % parallelism. >>> 2. Both topics have the same number of partitions. >>> 3. The parallelism of your join function exactly matches the number of >>> partitions. >>> 4. You can’t change any of the above without losing state. >>> 5. You don’t need stateful timers. >>> >>> If the above is true, then you could use a CoFlatMapFunction and >>> operator state to implement a stateful join. >>> >>> If it’s something like a left outer join without any state TTL or need >>> to keep both sides in state, then it’s pretty easy. >>> >>> — Ken >>> >>> PS - it’s pretty easy to figure out a “-xxx” value to append to a topic >>> name to get the hashCode() result you need. >>> >>> On Mar 3, 2023, at 4:56 PM, Tommy May <tvma...@gmail.com> wrote: >>> >>> Hello, >>> >>> My team has a Flink streaming job that does a stateful join across two >>> high throughput kafka topics. This results in a large amount of data ser/de >>> and shuffling (about 1gb/s for context). We're running into a bottleneck on >>> this shuffling step. We've attempted to optimize our flink configuration, >>> join logic, scale out the kafka topics & flink job, and speed up state >>> access. What we see is that the join step causes backpressure on the kafka >>> sources and lag slowly starts to accumulate. >>> >>> One idea we had to optimize this is to pre-partition the data in kafka >>> on the same key that the join is happening on. This'll effectively reduce >>> data shuffling to 0 and remove the bottleneck that we're seeing. I've done >>> some research into the topic and from what I understand this is not >>> straightforward to take advantage of in Flink. It looks to be a fairly >>> commonly requested feature based on the many StackOverflow posts and slack >>> questions, and I noticed there is FLIP-186 which attempts to address this >>> topic as well. >>> >>> Are there any upcoming plans to add this feature to a future Flink >>> release? I believe it'd be super impactful for similar large scale jobs out >>> there. I'd be interested in helping as well, but admittedly I'm relatively >>> new to Flink. I poked around the code a bit, and it certainly did not seem >>> like a straightforward addition, so it may be best handled by someone with >>> more internal knowledge. >>> >>> Thanks, >>> Tommy >>> >>> >>> -------------------------- >>> Ken Krugler >>> http://www.scaleunlimited.com >>> Custom big data solutions >>> Flink, Pinot, Solr, Elasticsearch >>> >>> >>> >>> >> -------------------------- >> Ken Krugler >> http://www.scaleunlimited.com >> Custom big data solutions >> Flink, Pinot, Solr, Elasticsearch >> >> >> >>
