Thanks. For the explanation – I was suspected it might be like this and I wanted to double check before building inconsistent programs ☺)
Would it be interesting for the community to have also something that would also be able to share/broadcast items from one task to the other tasks. Spark for example has this as the broadcast behavior, and it can be useful for different algorithms and applications. For example if one build a clustering algorithm running in parallel, then you can have a task updating the cluster centroids, that would advertise to the others afterwards (just one of the examples). I am thinking we can achieve this using a combination of the broadcast stream and the iterative operators. Basically if a task makes an update to the state, this value should be push back (upstream) such that afterwards it can be broadcasted to all the task using it. Let me know what you think.. Dr. Radu Tudoran Staff Research Engineer - Big Data Expert IT R&D Division [cid:image007.jpg@01CD52EB.AD060EE0] HUAWEI TECHNOLOGIES Duesseldorf GmbH German Research Center Munich Office Riesstrasse 25, 80992 München E-mail: radu.tudo...@huawei.com<mailto:radu.tudo...@huawei.com> Mobile: +49 15209084330 Telephone: +49 891588344173 HUAWEI TECHNOLOGIES Duesseldorf GmbH Hansaallee 205, 40549 Düsseldorf, Germany, www.huawei.com<http://www.huawei.com/> Registered Office: Düsseldorf, Register Court Düsseldorf, HRB 56063, Managing Director: Bo PENG, Qiuen Peng, Shengli Wang Sitz der Gesellschaft: Düsseldorf, Amtsgericht Düsseldorf, HRB 56063, Geschäftsführer: Bo PENG, Qiuen Peng, Shengli Wang This e-mail and its attachments contain confidential information from HUAWEI, which is intended only for the person or entity whose address is listed above. Any use of the information contained herein in any way (including, but not limited to, total or partial disclosure, reproduction, or dissemination) by persons other than the intended recipient(s) is prohibited. If you receive this e-mail in error, please notify the sender by phone or email immediately and delete it! From: Fabian Hueske [mailto:fhue...@gmail.com] Sent: Tuesday, August 28, 2018 9:53 AM To: Xingcan Cui <xingc...@gmail.com<mailto:xingc...@gmail.com>> Cc: Radu Tudoran <radu.tudo...@huawei.com<mailto:radu.tudo...@huawei.com>>; user <user@flink.apache.org<mailto:user@flink.apache.org>> Subject: Re: What's the advantage of using BroadcastState? Hi, Xingcan is right. There is no hidden state synchronization happening. You have to ensure that the broadcast state is the same at every parallel instance. Hence, it should only be modified by the processBroadcastElement() method that receives the same broadcasted elements on all task instance. The API tries to help users to not violate the contract, however it is not bullet proof. Side-passing information in a local variable (as suggested by you) cannot be prevented and would lead to inconsistencies. Best, Fabian Am Mo., 27. Aug. 2018 um 16:51 Uhr schrieb Xingcan Cui <xingc...@gmail.com<mailto:xingc...@gmail.com>>: Hi Radu, I cannot make a full understanding of your question but I guess the answer is NO. The broadcast state pattern just provides you with an automatic data broadcasting and a bunch of map states to cache the "low-throughput” patterns. Also, to keep consistency, it forbid the `processElement()` to modify the states. But this API does not really broadcast the states. You should keep the logic for `processBraodcastElement()` deterministic. Maybe the equation below could make the pattern clear. <identical input> + <deterministic logic> = <identical states> = <broadcast state> Best, Xingcan On Aug 27, 2018, at 10:23 PM, Radu Tudoran <radu.tudo...@huawei.com<mailto:radu.tudo...@huawei.com>> wrote: Hi Fabian, Thanks for the blog post about broadcast state. I have a question with respect to the update capabilities of the broadcast state: Assume you do whatever processing logic in the main processElement function .. and at a given context marker you 1) would change a local field marker, to 2) signal that next time the broadcast function is triggered a special pattern should be created and broadcasted. My question is: is such a behavior allowed? Would the new special Pattern that originates in an operator be shared across the other instances of the KeyedProcessFunction? public static class PatternEvaluator extends KeyedBroadcastProcessFunction<Long, Action, Pattern, Tuple2<Long, Pattern>> { public bolean test = false; @Override public void processElement( Action action, ReadOnlyContext ctx, Collector<Tuple2<Long, Pattern>> out) throws Exception { //…logic if (..whatever context) { Test = true; } } @Override public void processBroadcastElement( Pattern pattern, Context ctx, Collector<Tuple2<Long, Pattern>> out) throws Exception { // store the new pattern by updating the broadcast state BroadcastState<Void, Pattern> bcState = ctx.getBroadcastState(new MapStateDescriptor<>("patterns", Types.VOID, Types.POJO(Pattern.class))); // storing in MapState with null as VOID default value bcState.put(null, pattern); If (test) { bcState.put(null, new Pattern(test) ); } } } Dr. Radu Tudoran Staff Research Engineer - Big Data Expert IT R&D Division <image001.png> HUAWEI TECHNOLOGIES Duesseldorf GmbH German Research Center Munich Office Riesstrasse 25, 80992 München E-mail: radu.tudo...@huawei.com<mailto:radu.tudo...@huawei.com> Mobile: +49 15209084330 Telephone: +49 891588344173 HUAWEI TECHNOLOGIES Duesseldorf GmbH Hansaallee 205, 40549 Düsseldorf, Germany, www.huawei.com<http://www.huawei.com/> Registered Office: Düsseldorf, Register Court Düsseldorf, HRB 56063, Managing Director: Bo PENG, Qiuen Peng, Shengli Wang Sitz der Gesellschaft: Düsseldorf, Amtsgericht Düsseldorf, HRB 56063, Geschäftsführer: Bo PENG, Qiuen Peng, Shengli Wang This e-mail and its attachments contain confidential information from HUAWEI, which is intended only for the person or entity whose address is listed above. Any use of the information contained herein in any way (including, but not limited to, total or partial disclosure, reproduction, or dissemination) by persons other than the intended recipient(s) is prohibited. If you receive this e-mail in error, please notify the sender by phone or email immediately and delete it! From: Fabian Hueske [mailto:fhue...@gmail.com] Sent: Monday, August 20, 2018 9:40 AM To: Paul Lam <paullin3...@gmail.com<mailto:paullin3...@gmail.com>> Cc: Rong Rong <walter...@gmail.com<mailto:walter...@gmail.com>>; Hequn Cheng <chenghe...@gmail.com<mailto:chenghe...@gmail.com>>; user <user@flink.apache.org<mailto:user@flink.apache.org>> Subject: Re: What's the advantage of using BroadcastState? Hi, I've recently published a blog post about Broadcast State [1]. Cheers, Fabian [1] https://data-artisans.com/blog/a-practical-guide-to-broadcast-state-in-apache-flink 2018-08-20 3:58 GMT+02:00 Paul Lam <paullin3...@gmail.com<mailto:paullin3...@gmail.com>>: Hi Rong, Hequn Your answers are very helpful! Thank you! Best Regards, Paul Lam 在 2018年8月19日，23:30，Rong Rong <walter...@gmail.com<mailto:walter...@gmail.com>> 写道： Hi Paul, To add to Hequn's answer. Broadcast state can typically be used as "a low-throughput stream containing a set of rules which we want to evaluate against all elements coming from another stream" [1] So to add to the difference list is: whether it is "broadcast" across all keys if processing a keyed stream. This is typically when it is not possible to derive same key field using KeySelector in CoStream. Another additional difference is performance: BroadcastStream is "stored locally and is used to process all incoming elements on the other stream" thus requires to carefully manage the size of the BroadcastStream. [1]: https://ci.apache.org/projects/flink/flink-docs-release-1.6/dev/stream/state/broadcast_state.html On Sun, Aug 19, 2018 at 1:40 AM Hequn Cheng <chenghe...@gmail.com<mailto:chenghe...@gmail.com>> wrote: Hi Paul, There are some differences: 1. The BroadcastStream can broadcast data for you, i.e, data will be broadcasted to all downstream tasks automatically. 2. To guarantee that the contents in the Broadcast State are the same across all parallel instances of our operator, read-write access is only given to the broadcast side 3. For BroadcastState, flink guarantees that upon restoring/rescaling there will be no duplicates and no missing data. In case of recovery with the same or smaller parallelism, each task reads its checkpointed state. Upon scaling up, each task reads its own state, and the remaining tasks (p_new-p_old) read checkpoints of previous tasks in a round-robin manner. While MapState doesn't have such abilities. Best, Hequn On Sun, Aug 19, 2018 at 11:18 AM, Paul Lam <paullin3...@gmail.com<mailto:paullin3...@gmail.com>> wrote: Hi, AFAIK, the difference between a BroadcastStream and a normal DataStream is that the BroadcastStream is with a BroadcastState, but it seems that the functionality of BroadcastState can also be achieved by MapState in a CoMapFunction or something since the control stream is still broadcasted without being turned into BroadcastStream. So, I’m wondering what’s the advantage of using BroadcastState? Thanks a lot! Best Regards, Paul Lam
