+1 for this feature. I think this will be appreciated by users, as a way to use the HeapStateBackend with a safety-net against OOM errors. And having had major production exposure is great.
>From the implementation plan, it looks like this exists purely in a new module and does not require any changes in other parts of Flink's code. Can you confirm that? Other that that, I have no further questions and we could proceed to vote on this FLIP, from my side. Best, Stephan On Tue, Aug 13, 2019 at 10:00 PM Yu Li <car...@gmail.com> wrote: > Sorry for forgetting to give the link of the FLIP, here it is: > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-50%3A+Spill-able+Heap+Keyed+State+Backend > > Thanks! > > Best Regards, > Yu > > > On Tue, 13 Aug 2019 at 18:06, Yu Li <car...@gmail.com> wrote: > > > Hi All, > > > > We ever held a discussion about this feature before  but now opening > > another thread because after a second thought introducing a new backend > > instead of modifying the existing heap backend is a better option to > > prevent causing any regression or surprise to existing in-production > usage. > > And since introducing a new backend is relatively big change, we regard > it > > as a FLIP and need another discussion and voting process according to our > > newly drafted bylaw . > > > > Please allow me to quote the brief description from the old thread  > for > > the convenience of those who noticed this feature for the first time: > > > > > > *HeapKeyedStateBackend is one of the two KeyedStateBackends in Flink, > > since state lives as Java objects on the heap in HeapKeyedStateBackend > and > > the de/serialization only happens during state snapshot and restore, it > > outperforms RocksDBKeyeStateBackend when all data could reside in > memory.**However, > > along with the advantage, HeapKeyedStateBackend also has its > shortcomings, > > and the most painful one is the difficulty to estimate the maximum heap > > size (Xmx) to set, and we will suffer from GC impact once the heap memory > > is not enough to hold all state data. There’re several (inevitable) > causes > > for such scenario, including (but not limited to):* > > > > > > > > ** Memory overhead of Java object representation (tens of times of the > > serialized data size).* Data flood caused by burst traffic.* Data > > accumulation caused by source malfunction.**To resolve this problem, we > > proposed a solution to support spilling state data to disk before heap > > memory is exhausted. We will monitor the heap usage and choose the > coldest > > data to spill, and reload them when heap memory is regained after data > > removing or TTL expiration, automatically. Furthermore, *to prevent > > causing unexpected regression to existing usage of HeapKeyedStateBackend, > > we plan to introduce a new SpillableHeapKeyedStateBackend and change it > to > > default in future if proven to be stable. > > > > Please let us know your point of the feature and any comment is > > welcomed/appreciated. Thanks. > > > >  https://s.apache.org/pxeif > >  > > > https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=120731026 > > > > Best Regards, > > Yu > > >