flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stephan Ewen <se...@apache.org>
Subject Re: sharded state, 2-step operation
Date Tue, 23 Aug 2016 18:41:46 GMT
Hi!

This is a tricky one. State access and changes are not shared across
operators in Flink.
We chose that design because it makes it possible to work on "local" state
in each operator
  - state automatically shards with the computation
  - no locking / concurrency implications
  - asynchronous persistence

Sharing state across operations between two operations in the same stage
works with the CoMap / CoFlatMap functions
Sharing state across successive nodes does not work, because the functions
could be executed on different machines and one would need to do remote and
synchronized state updates that way.

Do you think you can use the CoMap / CoFlatMap functions for this?

Greetings,
Stephan


On Tue, Aug 23, 2016 at 8:03 PM, Michael Warnock <michael@ripple.com> wrote:

> I'm trying to do something that seems like it should be possible, but my
> implementation doesn't behave as expected, and I'm not sure how else to
> express it.
>
> Let's say the stream is composed of tuples like this: (Alice, Bob, 1) and
> I want to keyBy(1), flatMap with state associated with Alice, then keyBy(2)
> with state associated with Bob.  The trick is, when I later get a tuple
> like (Bob, Alice, 1), I want the first operator to see the state that was
> updated in the second op previously.  Is this possible?  I tried
>  implementing both operators as one, getting the state by descriptor in the
> flatMap body, and even instantiating the operator only once; the behavior
> is, as you might guess, that the state in stage 1 doesn't include changes
> made previously in stage 2.
>
> Is there any way to do this without throwing away the parallelism?
>
> Thanks in advance!
> ~Michael
>

Mime
View raw message