flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From bjlovegithub <...@git.apache.org>
Subject [GitHub] flink issue #2629: [FLINK-4391] Provide support for asynchronous operations ...
Date Fri, 04 Nov 2016 02:48:17 GMT
Github user bjlovegithub commented on the issue:

    https://github.com/apache/flink/pull/2629
  
    Hi @tillrohrmann , Thanks for your review ;D I will check through each of your comments
and update the PR later.
    Coming to the first part of review, the first one is about `UNORDERED` mode against `Watermark`.
This combination is meaningless, of course. Maybe an error can be printed out and the graph
generator stops compiling the graph if  `UNORDERED` mode and `Watermark` are enabled at the
same time?
    
    Both of these two modes are guaranteed by `AsyncWaitOperator`. While doing checkpoint
for the chained operator and making the snapshot for the `AsyncWaitOperator`, it will first
try to  get all elements in the  `AsyncCollectorBuffer` by calling `getStreamElementsInBuffer()`,
which will try to get the lock first to block `Emitter` thread and set a flag named `isCheckpointing`
to idle `Emitter` thread. So any finished `AsyncCollector` will not be transferred to the
next operator. Calling the `snapshotState()` method is from the head operator to the tail
operator, making sure that all states can be taken correctly since `Emitter` threads in parent
operators have stopped working.
    
    I used to consider about using checkpoint lock in `Emitter` thread, but after testing
with the case chaining multiple `AsyncWaitOperator` together, all `Emitter` thread can not
fully utilize the the parallelism since they have to get the same lock while collecting outputs.
One way to optimize this is to put a conditional statement at `performCheckpoint()`, if there
is an `AsyncWaitOpeartor` in the chained operator, then it should broadcast barriers later
after `checkpointState()`, otherwise, we can use original design.
    
    At last, I will add more test cases based on the `OneInputStreamTaskTestHarness`.


---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastructure@apache.org or file a JIRA ticket
with INFRA.
---

Mime
View raw message