flink-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aljoscha Krettek <aljos...@apache.org>
Subject Re: [DISCUSS] Hold copies in HeapStateBackend
Date Mon, 21 Nov 2016 17:41:36 GMT
To bring over my comment from the Github PR that started this discussion:

@wuchong <https://github.com/wuchong>, yes this is a problem with the
HeapStateBackend. The RocksDB backend does not suffer from this problem. I
think in the long run we should migrate the HeapStateBackend to always keep
data in serialised form, then we also won't have this problem anymore.

So I'm very much in favour of keeping data serialised. Copying data would
only ever be a stopgap solution.

On Mon, 21 Nov 2016 at 15:56 Fabian Hueske <fhueske@gmail.com> wrote:

> Another approach that would solve the problem for our use case (object
> re-usage for incremental window ReduceFunctions) would be to copy the first
> object that is put into the state.
> This would be a change on the ReduceState, not on the overall state
> backend, which should be feasible, no?
>
>
>
> 2016-11-21 15:43 GMT+01:00 Stephan Ewen <sewen@apache.org>:
>
> > -1 for copying objects.
> >
> > Storing a serialized data where possible is good, but copying all objects
> > by default is not a good idea, in my opinion.
> > A lot of scenarios use data types that are hellishly expensive to copy.
> > Even the current copy on chain handover is a problem.
> >
> > Let's not introduce even more copies.
> >
> > On Mon, Nov 21, 2016 at 3:16 PM, Maciek Próchniak <mpr@touk.pl> wrote:
> >
> > > Hi,
> > >
> > > it will come with performance overhead when updating the state, but I
> > > think it'll be possible to perform asynchronous snapshots using
> > > HeapStateBackend (probably some changes to underlying data structures
> > would
> > > be needed) - which would bring more predictable performance.
> > >
> > > thanks,
> > > maciek
> > >
> > >
> > > On 21/11/2016 13:48, Aljoscha Krettek wrote:
> > >
> > >> Hi,
> > >> I would be in favour of this since it brings things in line with the
> > >> RocksDB backend. This will, however, come with quite the performance
> > >> overhead, depending on how fast the TypeSerializer can copy.
> > >>
> > >> Cheers,
> > >> Aljoscha
> > >>
> > >> On Mon, 21 Nov 2016 at 11:30 Fabian Hueske <fhueske@gmail.com> wrote:
> > >>
> > >> Hi everybody,
> > >>>
> > >>> when implementing a ReduceFunction for incremental aggregation of
> SQL /
> > >>> Table API window aggregates we noticed that the HeapStateBackend does
> > not
> > >>> store copies but holds references to the original objects. In case
> of a
> > >>> SlidingWindow, the same object is referenced from different window
> > panes.
> > >>> Therefore, it is not possible to modify these objects (in order to
> > avoid
> > >>> object instantiations, see discussion [1]).
> > >>>
> > >>> Other state backends serialize their data such that the behavior is
> not
> > >>> consistent across backends.
> > >>> If we want to have light-weight tests, we have to create new objects
> in
> > >>> the
> > >>> ReduceFunction causing unnecessary overhead.
> > >>>
> > >>> I would propose to copy objects when storing them in a
> > HeapStateBackend.
> > >>> This would ensure that objects returned from state to the user behave
> > >>> identical for different state backends.
> > >>>
> > >>> We created a related JIRA [2] that asks to copy records that go into
> an
> > >>> incremental ReduceFunction. The scope is more narrow and would solve
> > our
> > >>> problem, but would leave the inconsistent behavior of state backends
> in
> > >>> place.
> > >>>
> > >>> What do others think?
> > >>>
> > >>> Cheers, Fabian
> > >>>
> > >>> [1] https://github.com/apache/flink/pull/2792#discussion_r88653721
> > >>> [2] https://issues.apache.org/jira/browse/FLINK-5105
> > >>>
> > >>>
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message