geode-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Anilkumar Gingade <aging...@pivotal.io>
Subject Re: [DISCUSS] Controlling event dispatch to AsyncEventListener (review by Aug 22)
Date Tue, 20 Aug 2019 23:03:45 GMT
Thanks for all the great feedback and comments.

*API Name change:*
*Suggestion: *startPaused*,*setManualStart*,
*startWithEventDispatcherPaused*?, createPaused()*

*Start/Stop behavior:*


*- Manual start has caused a lot of trouble over the years.- Explain
starting AEQ in a paused state is different from creating gateway senders
with manual start*

Yes, we can change/adopt the name which is meaningful with the
functionality. The name suggested "pauseEventDispatchToListener()" is to
make its usage/action clear; and address any ambiguity with its usage;
between adding/removing event from AEQ and dispatching events to AEQ
listener.

To emphasis, this is not same as GatewaySender manual start (start/stop);
the manual start/stop is with enqueuing and dequeuing events from
GatewaySender itself, and as Mike pointed out there are issues with this
(during recovery with parallel gateway) and the reason its been deprecated.

The new functionality is similar to the "pause" and "resume" operation on
the GatewaySender. Except that here with the new api, the AEQ is created
with pause state.

The new api doesn't control adding and removing event from AEQ. Its to
control dispatching event to the listener. When created in paused state,
the events are continued be added into the AEQ and removed from it
(expiry). The new API will allow applications to create/manage any required
state/resource for the events before processing those events in the
application code.

*Cache level setting:*
*- Will it be more feasible if we can set the flag at cache level.*

The cache level configuration affects all the AEQs, which may not be the
requirement. Having at AEQ level help the application to use this
capability only at the required AEQ, gives more controlling capability.

-Anil.







On Tue, Aug 20, 2019 at 2:01 PM Nabarun Nag <nnag@apache.org> wrote:

> Hi Anil,
>
> Will it be possible to explain to the community how the starting AEQ in a
> paused state is different from creating gateway senders with manual start
> set to true. It may be of concern as  'manual start'  for gateways is a
> deprecated.
>
> Just thinking out loud, will it be more feasible if we can set the flag at
> cache level. Any framework that is starting up Apache Geode (E.g: Spring) ,
> creates the cache -> cache.pauseProcessing(); -> create regions -> create
> AEQs -> cache.unpauseProcessing()
>
> We can gate the processing of all event listener at dispatchBatch().
>
> The advantage I feel is that
>  - we avoid introducing a new API to the AEQ creation factory.
>  - if we created 100 AEQs in paused state then we avoid having to have 100
> AEQ.unpause calls.
>
>
> Regards
> Naba
>
>
> On Tue, Aug 20, 2019 at 9:07 AM Michael Stolz <mstolz@pivotal.io> wrote:
>
> > Manual start has caused a lot of trouble over the years. We should
> > definitely circle back on those issues before traveling very far down
> this
> > road.
> >
> > --
> > Mike Stolz
> > Principal Engineer, Pivotal Cloud Cache
> > Mobile: +1-631-835-4771
> >
> >
> >
> > On Tue, Aug 20, 2019 at 11:56 AM Juan José Ramos <jramos@pivotal.io>
> > wrote:
> >
> > > Hello Anil,
> > >
> > > +1 for the proposed solution.
> > > I'd change the method name from *pauseEventDispatchToListener* to
> > something
> > > more meaningful and understandable for our users, maybe *startPaused*?,
> > > *setManualStart* (as we currently have for the
> *GatewaySenderFactory*)?,
> > > *startWithEventDispatcherPaused*?.
> > > Best regards.
> > >
> > >
> > >
> > > On Sat, Aug 17, 2019 at 12:55 AM Anilkumar Gingade <
> agingade@pivotal.io>
> > > wrote:
> > >
> > > > I have updated the wiki based on Dan's comment.
> > > > Changes with api:
> > > >
> > > > *On "AsyncEventQueueFactory" interface - *
> > > >
> > > > *AsyncEventQueueFactory pauseEventDispatchToListener();  *// This
> > causes
> > > > AEQ to be created with paused state.
> > > >
> > > >
> > > >
> > > >
> > > > On Fri, Aug 16, 2019 at 4:36 PM Anilkumar Gingade <
> agingade@pivotal.io
> > >
> > > > wrote:
> > > >
> > > > > Dan,
> > > > >
> > > > > If you look into the API; the AEQ will be created with the pause
> > state.
> > > > > The user (application) has to call resume to dispatch the events.
> > > > >
> > > > > It will be slightly different from GatewaySender behavior; where
> > > > > GatewaySender will be created with run mode and then application
> has
> > to
> > > > > call pause on it. Here in this case AEQ will be created with paused
> > > > state.
> > > > >
> > > > > -Anil.
> > > > >
> > > > >
> > > > > On Fri, Aug 16, 2019 at 4:31 PM Dan Smith <dsmith@pivotal.io>
> wrote:
> > > > >
> > > > >> Hi Anil,
> > > > >>
> > > > >> While I like the idea of matching the API of GatewaySender, I'm
> not
> > > > sure I
> > > > >> see how this solves the problem. Is it required of the user to
> call
> > > > pause
> > > > >> on the AsyncEventQueue as soon as it is created? How would someone
> > do
> > > > that
> > > > >> when creating AEQs with xml or cluster configuration? Maybe it
> would
> > > be
> > > > >> better to not dispatch any events until we are done creating
all
> > > > regions?
> > > > >>
> > > > >> -Dan
> > > > >>
> > > > >> On Fri, Aug 16, 2019 at 2:31 PM Anilkumar Gingade <
> > > agingade@pivotal.io>
> > > > >> wrote:
> > > > >>
> > > > >> > Proposal to support controlling capability with event dispatch
> to
> > > > >> > AsyncEventQueue Listener.
> > > > >> >
> > > > >> > Wiki proposal page:
> > > > >> >
> > > > >> >
> > > > >>
> > > >
> > >
> >
> https://cwiki.apache.org/confluence/display/GEODE/%5BDraft%5D+Controlling+event+dispatch+to+AsyncEventListener
> > > > >> >
> > > > >> > Here is the details from the wiki page:
> > > > >> > *Problem*
> > > > >> >
> > > > >> > *The Geode system requires AEQs to be configured before
regions
> > are
> > > > >> > created. If an AEQ listener is operating on a secondary
region,
> > this
> > > > >> could
> > > > >> > cause listener to operate on a region which is not yet created
> or
> > > > fully
> > > > >> > initialized (for region with co-located regions) which could
> > result
> > > in
> > > > >> > missing events or dead-lock scenario between region (co-located
> > > > region)
> > > > >> > creation threads. This scenario is likely to happen during
> > > persistence
> > > > >> > recovery; when AEQs are created in the start, the recovered
AEQ
> > > events
> > > > >> are
> > > > >> > dispatched immediately, thus invoking the AEQ listeners.*
> > > > >> > Anti-Goals
> > > > >> >
> > > > >> > None
> > > > >> > *Solution*
> > > > >> >
> > > > >> > *The proposed solution is to provide a way to control
> dispatching
> > > AEQ
> > > > >> > events to the AEQ Listeners, this could be done by adding
> "pause"
> > > and
> > > > >> > "resume" capability to the AEQ, which will allow application
to
> > > decide
> > > > >> when
> > > > >> > to dispatch events to the listeners. *
> > > > >> >
> > > > >> >
> > > > >> > *The proposal is similar to existing "pause" and "resume"
> behavior
> > > on
> > > > >> the
> > > > >> > GatewaySender, on which the AEQ is based on (AEQ implementation
> > is a
> > > > >> > wrapper around GatewaySender). *
> > > > >> > Changes and Additions to Public Interfaces
> > > > >> >
> > > > >> > *The proposed APIs are:*
> > > > >> >
> > > > >> > *On "AsyncEventQueueFactory" interface - *
> > > > >> >
> > > > >> > *AsyncEventQueue pauseEventDispatchToListener();*
> > > > >> >
> > > > >> > *On "AsyncEventQueue" interface -*
> > > > >> >
> > > > >> > *boolean resumeEventDispatchToListener(); **returns true
or
> false
> > if
> > > > the
> > > > >> > event dispatch is resumed successfully.*
> > > > >> >
> > > > >> >
> > > > >> > *The constraints on the pauseEventDispatchToListener() will
> remain
> > > > >> similar
> > > > >> > to as in "GatewaySender.pause()" :*
> > > > >> >
> > > > >> > "It should be kept in mind that the events will still be
getting
> > > > queued
> > > > >> > into the queue. The scope of this operation is the VM on
which
> it
> > is
> > > > >> > invoked. In case the AEQ is parallel, the AEQ will be paused
on
> > > > >> individual
> > > > >> > node where this API is called and the AEQ on other VM's
can
> still
> > > > >> dispatch
> > > > >> > events. In case the AEQ is not parallel, and the running
AEQ on
> > > which
> > > > >> this
> > > > >> > API is invoked is not primary then primary AEQ will still
> continue
> > > > >> > dispatching events."
> > > > >> > Performance Impact
> > > > >> >
> > > > >> >
> > > > >> > *This will have similar performance and resource implication
as
> > with
> > > > the
> > > > >> > "GatewaySender.pause()" functionality. If the AEQ is not
resumed
> > or
> > > > >> kept in
> > > > >> > "pause" state for long, it may start consuming the configured
> > memory
> > > > and
> > > > >> > overflow it into disk and may cause disk full scenario.*
> > > > >> > Backwards Compatibility and Upgrade Path
> > > > >> >
> > > > >> > *Impact with rolling upgrade: *
> > > > >> >
> > > > >> > *As the api is applicable at individual VM level, there
is no
> > > message
> > > > >> > serialization changes involved. And only applicable to the
> events
> > > > >> getting
> > > > >> > dispatched to the listeners on that VM. And the AEQ which
are
> > > > replicated
> > > > >> > (for redundancy) continues to work as before.*
> > > > >> >
> > > > >> > *Backward compatibility requirements: *
> > > > >> >
> > > > >> > *None. The AEQs are configured and managed at the server
side.
> > There
> > > > is
> > > > >> no
> > > > >> > messaging involved between client/server.*
> > > > >> >
> > > > >> > *Disk formatting changes:*
> > > > >> >
> > > > >> > *None.*
> > > > >> >
> > > > >> > *Deprecation and Application Changes:*
> > > > >> >
> > > > >> >
> > > > >> > *None. If needed, the existing application can be modified
to
> > > control
> > > > >> event
> > > > >> > dispatch with AEQ listener.*
> > > > >> > Prior Art
> > > > >> >
> > > > >> > *Without this, the AEQ listeners operating on other regions
> could
> > > > >> > experience missing events or dead lock, if there are co-located
> > > > >> regions.*
> > > > >> >
> > > > >> > *This approach is simple and can take advantage of the existing
> > > > >> > functionality that is already supported in GatewaySender
on
> which
> > > AEQ
> > > > is
> > > > >> > based on.*
> > > > >> >
> > > > >>
> > > > >
> > > >
> > >
> > >
> > > --
> > > Juan José Ramos Cassella
> > > Senior Software Engineer
> > > Email: jramos@pivotal.io
> > >
> >
>

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