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] Removing delete*Timer from the WindowOperator.Context
Date Tue, 11 Oct 2016 12:02:51 GMT
+Konstantin Knauf <konstantin.knauf@tngtech.com> looping you in directly
because you used the "delete timer" feature in the past and even did some
changes to the timer system. Are you still relying on the fact that deleted
timers are actually deleted.

The main reason for wanting to get rid of delete timer is IMHO that
deleting a timer is difficult, depending on the data structure that you use
for timers. Especially if you want a data structure that can grow out of
core. By the way, the current data structure for timers is a Java Queue (a
heap) so deletes from this are O(n), i.e. possibly slow.

On Wed, 28 Sep 2016 at 15:21 Maximilian Michels <mxm@apache.org> wrote:

> What are the use cases where you actually need to delete a timer? How
>
> about we only let users delete timers which they created themselves?
>
>
>
> I guessing most of these use cases will be obsolete with the new
>
> Trigger DSL because the trigger logic can be expressed more easily. So
>
> +1 for removing the delete methods from the context.
>
>
>
> On Tue, Sep 27, 2016 at 3:43 PM, Kostas Kloudas
>
> <k.kloudas@data-artisans.com> wrote:
>
> > Hi all,
>
> >
>
> > As the title of this email suggests, I am proposing to remove the
> methods
>
> > deleteProcessingTimeTimer(long time) and deleteEventTimeTimer(long time)
>
> > from the WindowOperator.Context. With this change, registered timers that
>
> > have nothing to do (e.g. because their state has already been cleaned up)
>
> > will be simply ignored by the windowOperator, when their time comes.
>
> >
>
> > The reason for the change is that by allowing custom user code, e.g. a
> custom Trigger,
>
> > to delete timers we may have unpredictable behavior.
>
> >
>
> > As an example, one can imagine the case where we have allowed_lateness =
> 0 and the cleanup
>
> > timer for a window collides with the end_of_window one. In this case, by
> deleting the end_of_window
>
> > timer from the trigger (possibly a custom one), we end up also deleting
> the cleanup one,
>
> > which in turn can lead to the window state never being garbage collected.
>
> >
>
> > To see what can be the consequences apart from memory leaks, this can
> easily lead
>
> > to wrong session windows, as a session that should have been garbage
> collected, will
>
> > still be around and ready to accept new data.
>
> >
>
> > With this change, timers that should correctly be deleted will now
> remain in the queue of
>
> > pending timers, but they will do nothing, while cleanup timers will
> cleanup the state of their
>
> > corresponding window.
>
> >
>
> > Other possible solutions like keeping a separate list for cleanup timers
> would complicate
>
> > the codebase and also introduce memory overheads which can be avoided
> using the
>
> > solution above (i.e. just ignoring timers the have nothing to do
> anymore).
>
> >
>
> > What do you think?
>
> >
>
> > Kostas
>
> >
>
>

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