Hi Maciej, thanks for joining. I answer your comments below. > > the idea is quite interesting - although maintaining some coordination to > be able to handle checkpoints would probably pretty tricky. Did you figure > out how to handle proper distribution of tasks between TMs? As far as I > understand you have to guarantee that all sources reading from cache are on > the same TM as sinks writing data from Kafka? Or you think about some > distributed caches? > No, we haven't yet figured that out. Yes, I've heard that it is indeed a problem to force Flink to distribute the tasks as one wants it. I only hoped that we will be able to solve it when we get there :-) One of the ideas was to actually use in-memory grid co-located with Flink (like based on Apache Ignite), but then the problem of network shuffle just moved from Kafka to that grid. Which might be smaller problem, but still.
> As for your original question - we are also looking for solutions/ideas > for this problem in Nussknacker. We have similar problem, however we had > different constraints (on premise, not have to care too much about > bandwidth) and we went with "one job per scenario". It works ok, but the > biggest problem for me is that it does not scale with the number of jobs: > Flink job is quite heavy entity - all the threads, classloaders etc. Having > more than a few dozens of jobs is also not so easy to handle on JobManager > part - especially when it's restarted etc. I guess your idea would also > suffer from this problem? > Unfortunately yes, good point. Maybe it can be mitigated if I had the jobs distributed among several Flink clusters. Solution globally heavier, but lighter on each cluster. Then the data must go to the distributed in-memory grid with hopefully local reads only. I see a lot of difficulties in the discussed approach. But my willingness to use SQL/Table API and CEP is so strong, I want to do the PoC regardless. I hope we will be able to provide benchmarks which prove that the performance of such approach is significantly better justifying the work by us, maybe also the community, on overcoming these difficulties. > > thanks, > > maciek > > > > On 27/03/2020 10:18, Krzysztof Zarzycki wrote: > > I want to do a bit different hacky PoC: > * I will write a sink, that caches the results in "JVM global" memory. > Then I will write a source, that reads this cache. > * I will launch one job, that reads from Kafka source, shuffles the data > to the desired partitioning and then sinks to that cache. > * Then I will lunch multiple jobs (Datastream based or Flink SQL based) , > that uses the source from cache to read the data out and then reinterprets > it as keyed stream [1]. > * Using JVM global memory is necessary, because AFAIK the jobs use > different classloaders. The class of cached object also needs to be > available in the parent classloader i.e. in the cluster's classpath. > This is just to prove the idea, the performance and usefulness of it. All > the problems of checkpointing this data I will leave for later. > > I'm very very interested in your, community, comments about this idea and > later productization of it. > Thanks! > > Answering your comments: > >> Unless you need reprocessing for newly added rules, I'd probably just >> cancel with savepoint and restart the application with the new rules. Of >> course, it depends on the rules themselves and how much state they require >> if a restart is viable. That's up to a POC. >> > No, I don't need reprocessing (yet). The rule starts processing the data > from the moment it is defined. > The cancellation with savepoint was considered, but because the number of > new rules defined/changed daily might be large enough, that will generate > too much of downtime. There is a lot of state kept in those rules making > the restart heavy. What's worse, that would be cross-tenant downtime, > unless the job was somehow per team/tenant. Therefore we reject this option. > BTW, the current design of our system is similar to the one from the blog > series by Alexander Fedulov about dynamic rules pattern [2] he's just > publishing. > > >> They will consume the same high intensive source(s) therefore I want to >>> optimize for that by consuming the messages in Flink only once. >>> >> That's why I proposed to run one big query instead of 500 small ones. >> Have a POC where you add two of your rules manually to a Table and see how >> the optimized logical plan looks like. I'd bet that the source is only >> tapped once. >> > > I can do that PoC, no problem. But AFAIK it will only work with the > "restart with savepoint" pattern discussed above. > > > [1] > https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/experimental.html#reinterpreting-a-pre-partitioned-data-stream-as-keyed-stream > [2] https://flink.apache.org/news/2020/03/24/demo-fraud-detection-2.html > > > >> On Wed, Mar 25, 2020 at 6:15 PM Krzysztof Zarzycki <k.zarzy...@gmail.com> >> wrote: >> >>> Hello Arvid, >>> Thanks for joining to the thread! >>> First, did you take into consideration that I would like to dynamically >>> add queries on the same source? That means first define one query, later >>> the day add another one , then another one, and so on. A Week later kill >>> one of those, start yet another one, etc... There will be hundreds of these >>> queries running at once, but the set of queries change several times a day. >>> They will consume the same high intensive source(s) therefore I want to >>> optimize for that by consuming the messages in Flink only once. >>> >>> Regarding the temporary tables AFAIK they are only the metadata (let's >>> say Kafka topic detals) and store it in the scope of a SQL session. >>> Therefore multiple queries against that temp table will behave the same way >>> as querying normal table, that is will read the datasource multiple times. >>> >>> It looks like the feature I want or could use is defined by the way of >>> FLIP-36 about Interactive Programming, more precisely caching the stream >>> table [1]. >>> While I wouldn't like to limit the discussion to that non-existing yet >>> feature. Maybe there are other ways of achieving this danymic querying >>> capability. >>> >>> Kind Regards, >>> Krzysztof >>> >>> [1] >>> https://cwiki.apache.org/confluence/display/FLINK/FLIP-36%3A+Support+Interactive+Programming+in+Flink#FLIP-36:SupportInteractiveProgramminginFlink-Cacheastreamtable >>> >>> >>> >>> * You want to use primary Table API as that allows you to >>>> programmatically introduce structural variance (changing rules). >>>> >>> * You start by registering the source as temporary table. >>>> >>> * Then you add your rules as SQL through `TableEnvironment#sqlQuery`. >>>> * Lastly you unionAll the results. >>>> >>>> Then I'd perform some experiment if indeed the optimizer figured out >>>> that it needs to only read the source once. The resulting code would be >>>> minimal and easy to maintain. If the performance is not satisfying, you can >>>> always make it more complicated. >>>> >>>> Best, >>>> >>>> Arvid >>>> >>>> >>>> On Mon, Mar 23, 2020 at 7:02 PM Krzysztof Zarzycki < >>>> k.zarzy...@gmail.com> wrote: >>>> >>>>> Dear Flink community! >>>>> >>>>> In our company we have implemented a system that realize the dynamic >>>>> business rules pattern. We spoke about it during Flink Forward 2019 >>>>> https://www.youtube.com/watch?v=CyrQ5B0exqU. >>>>> The system is a great success and we would like to improve it. Let me >>>>> shortly mention what the system does: >>>>> * We have a Flink job with the engine that applies business rules on >>>>> multiple data streams. These rules find patterns in data, produce complex >>>>> events on these patterns. >>>>> * The engine is built on top of CoProcessFunction, the rules are >>>>> preimplemented using state and timers. >>>>> * The engine accepts control messages, that deliver configuration of >>>>> the rules, and start the instances of the rules. There might be many rule >>>>> instances with different configurations running in parallel. >>>>> * Data streams are routed to those rules, to all instances. >>>>> >>>>> The *advantages* of this design are: >>>>> * *The performance is superb. *The key to it is that we read data >>>>> from the Kafka topic once, deserialize once, shuffle it once (thankfully >>>>> we >>>>> have one partitioning key) and then apply over 100 rule instances needing >>>>> the same data. >>>>> * We are able to deploy multiple rule instances dynamically without >>>>> starting/stopping the job. >>>>> >>>>> Especially the performance is crucial, we have up to 500K events/s >>>>> processed by 100 of rules on less than 100 of cores. I can't imagine >>>>> having >>>>> 100 of Flink SQL queries each consuming these streams from Kafka on such a >>>>> cluster. >>>>> >>>>> The main *painpoints *of the design is: >>>>> * to deploy new business rule kind, we need to predevelop the rule >>>>> template with use of our SDK. *We can't use* *great Flink CEP*, *Flink >>>>> SQL libraries.* Which are getting stronger every day. Flink SQL with >>>>> MATCH_RECOGNIZE would fit perfectly for our cases. >>>>> * The isolation of the rules is weak. There are many rules running per >>>>> job. One fails, the whole job fails. >>>>> * There is one set of Kafka offsets, one watermark, one checkpoint for >>>>> all the rules. >>>>> * We have one just distribution key. Although that can be overcome. >>>>> >>>>> I would like to focus on solving the *first point*. We can live with >>>>> the rest. >>>>> >>>>> *Question to the community*: Do you have ideas how to make it >>>>> possible to develop with use of Flink SQL with MATCH_RECOGNIZE? >>>>> >>>>> My current ideas are: >>>>> 1. *A possibility to dynamically modify the job topology. * >>>>> Then I imagine dynamically attaching Flink SQL jobs to the same Kafka >>>>> sources. >>>>> >>>>> 2. *A possibility to save data streams internally to Flink, >>>>> predistributed*. Then Flink SQL queries should be able to read these >>>>> streams. >>>>> >>>>> The ideal imaginary solution would look that simple in use: >>>>> CREATE TABLE my_stream(...) with (<kafka properties>, >>>>> cached = 'true') >>>>> PARTITIONED BY my_partition_key >>>>> >>>>> (the cached table can also be a result of CREATE TABLE and INSERT INTO >>>>> my_stream_cached SELECT ... FROM my_stream). >>>>> >>>>> then I can run multiple parallel Flink SQL queries reading from that >>>>> cached table in Flink. >>>>> These >>>>> >>>>> Technical implementation: Ideally, I imagine saving events in Flink >>>>> state before they are consumed. Then implement a Flink source, that can >>>>> read the Flink state of the state-filling job. It's a different job, I >>>>> know! Of course it needs to run on the same Flink cluster. >>>>> A lot of options are possible: building on top of Flink, modifying >>>>> Flink (even keeping own fork for the time being), using an external >>>>> component. >>>>> >>>>> In my opinion the key to the maximized performance are: >>>>> * avoid pulling data through network from Kafka >>>>> * avoid deserialization of messages for each of queries/ processors. >>>>> >>>>> Comments, ideas - Any feedback is welcome! >>>>> Thank you! >>>>> Krzysztof >>>>> >>>>> P.S. I'm writing to both dev and users groups because I suspect I >>>>> would need to modify Flink to achieve what I wrote above. >>>>> >>>>
