The tuples are not buffered until the snapshot is globally complete (a snapshot is globally complete iff all operators have successfully taken a snapshot). They are only buffered until the corresponding checkpoint barrier on the second input is received. Once this is the case, the checkpoint barrier will directly be send to the downstream operators. Next a snapshot is taken. Depending on the state backend this can happen asynchronously or synchronously. After this is done, the operator continues processing elements (for the first input, the buffered elements are consumed first).
With multiple inputs I referred to a coFlatMap operator or a join operator which have both two inputs. Cheers, Till On Tue, Nov 1, 2016 at 3:29 PM, Renjie Liu <liurenjie2...@gmail.com> wrote: > Hi, Till: > By operator with multiple inputs, do you mean inputs from multiple > subtasks? > > On Tue, Nov 1, 2016 at 8:56 PM Till Rohrmann <trohrm...@apache.org> wrote: > >> Hi Li, >> >> the statement refers to operators with multiple inputs (two in this >> case). With the current implementation you will indeed block one of the >> inputs after receiving a checkpoint barrier n until you've received the >> corresponding checkpoint barrier n on the other input as well. This is what >> we call checkpoint barrier alignment. If the processing time on both input >> paths is similar and thus there is no back pressure on any of the inputs, >> the alignment should not take too long. In case where one of the inputs is >> considerably slower than the other, you should an additional delay. >> >> For single input operators, you don't have to align the checkpoint >> barriers. >> >> The checkpoint barrier alignment is not strictly necessary, but it allows >> us to not having to store all in flight records from the second input which >> arrive between the checkpoint barrier on the first input and the >> corresponding barrier on the second input. We might change this >> implementation in the future, though. >> >> Cheers, >> Till >> >> On Tue, Nov 1, 2016 at 8:05 AM, Li Wang <wangli1...@gmail.com> wrote: >> >> Hi all, >> >> I have a question regarding to the state checkpoint mechanism in Flink. I >> find the statement "Once the last stream has received barrier n, the >> operator emits all pending outgoing records, and then emits >> snapshot n barriers itself” on the document https://ci.apache.org/ >> projects/flink/flink-docs-master/internals/stream_ >> checkpointing.html#exactly-once-vs-at-least-once. >> >> Does this mean that to achieve exactly-once semantic, instead of sending >> tuples downstream immediately the operator buffers its outgoing tuples in a >> pending queue until the current snapshot is committed? If yes, will this >> introduce significant processing delay? >> >> Thanks, >> Li >> >> >> -- > Liu, Renjie > Software Engineer, MVAD >
