Hi Dawid,
Very thanks for the discussion and sorry for the delayed response
since I was hesitated on some points.
But as a whole, with some more thought, first I agree with that adding
the trigger() / cancle() methods to some kind of timer object is not necessary
for us to achieve the exactly-once for the operators. We could follow the
direction of "modifying the implementation of the operators" to achieve the
same target.
But continue to think with this direction, it now looks to me it is also not
needed to add the callback to the timer services:
1. For InternalTimerService, the operators could just call `InternalTimerService
#forEachProcessingTimer()` on finish to handle the pending timers.
2. For the timers registered to the underlying ProcessingTimerService, at least
in
the currently listed scenarios, the operators itself knows what is the
remaining work
(e.g., the FileWriter knows if it has in-progress file to flush).
Operators could handle the remaining timers in finish() method.
Then the only interface we need to consider is that added to the
ProcessFunction. The
current interface also looks ok to me.
If you think the above option works, I could first have a PoC that demonstrate
it is sufficient
to only modify the operator implementation to handling the remaining workers
properly on
finish(). If there are new issues I'll post here and we could have some more
discussion.
Best,
Yun Gao
------------------Original Mail ------------------
Sender:Dawid Wysakowicz <dwysakow...@apache.org>
Send Date:Fri Dec 2 21:21:25 2022
Recipients:Dev <dev@flink.apache.org>
Subject:Re: [DISCUSS] FLIP-269: Properly Handling the Processing Timers on Job
Termination
Ad. 1
I'd start with ProcessingTimerService as that's the only public interface. It
is exposed in the Sink V2 interface. In this scenario it would be the Sink
interface that need to extend from a EOFTimersHandler. I believe it would be
hard to pass it from there to the ProcessingTimeService as it is passed from
the outside e.g. in the ProcessingTimeServiceAware. For that reason I'd go with
a registration method in that interface.
In ProcessFunction I'd go with a mixin approach, so a ProcessFunction can
extend from EOFTimersHandler. I'd do that because ProcessFunction does not have
an init/open method where we could register the handler.
On operator level I'd have a registration method in InternalTimerService. I
believe that's the only way to handle the above ProcessFunction aproach. E.g.
in KeyedProcessOperator you need to check if the UDF extend from the interface
not the operator itself.
Ad. 2
I'd go with
(Keyed)ProcessFunction:
interface EOFTimersHandler {
void handleProcessingTimer(long timestamp, Context);
}
interface Context {
public abstract <X> void output(OutputTag<X> outputTag, X value);
public abstract K getCurrentKey();
// we can extend it for waitFor later
}
ProcessingTimeService:
interface EOFTimersHandler {
void handleProcessingTimer(long timestamp, Context);
}
interface Context {
// we can extend it for waitFor later
}
InternalTimeService:
interface EOFTimersHandler {
void handleProcessingTimer(InternalTimer<K,N> timer Context);
}
interface Context {
// we can extend it for waitFor later
}
Personally I'd not try to unify those places too much. They have also different
visibilities (public/internal), have access to different set of metadata
(key/namespace).
Ad 3.
I don't like the having the trigger/cancel methods, because:
1. I don't like the back and forth between system and UDF
2. Yes, the biggest issue I have is with the possibility with registering new
timers. I am trying to be on the safe side here. I don't like the idea of
dropping them, because it is again making assumptions what users do with those
timers. What if they e.g. emit counter if it reached certain threshold? We'd
need an additional flag in the method that is the final timer. My sentiment is
that we're making it questionably easier to trigger a timer for the cost of
openning up for unforeseen problems with follow up registration.
Best,
Dawid
On 30/11/2022 12:13, Yun Gao wrote:
Hi Dawid, PiotrVery thanks for the discussion!As a whole I think we are already
consistent with the callback option, and I don't think I opposed that we could
modify the current internal implementation. But from my side it is still not
clear what the actual interfaces are proposing. Let me first try to summarize
that a bit:1) Which object does the handlers register on?It seems there are two
options, one is to timer services (InternalTimerService/ ProcessingTimerService
or some equivalent things after refactoring), the otherone is as a lifecycle of
the operator. I'm now tending to the latter one, how do you think on this
part?2) What is the interface of the handler?Option 1 is that interface
SomeHandlerName { void processingTimer(Timer timer);}class Timer { long
getTimestamp(); void trigger(); void cancel(); // Other actions if required.
}But it seems there is controversy on whether to add actions to the timer
class. If without that, with my understanding the interfaces of the Option 2
areinterface SomeHandlerName { void processTimer(Timer timer); }interface
KeyedSomeHandlerName<KEY, NAMESPACE> { void processKeyedTimer(KeyedTimer<KEY,
NAMESPACE> timer, Context ctx); }class Timer { long getTimestamp();}class
KeyedTimer<KEY, NAMESPACE> extends Timer { KEY getKey(); NAMESPACE
getNamespace();}void Context {void executeAtScheduledTime(Consumer<timer>
handler);}As Piotr has pointed out, if we could eliminate the logic of
namespace, we could thenremove the namespace related type parameter and method
from the interfaces.Do I understand right?Besides, I'm still fully got the
reason that why we should not add the actions to the timer class, in
consideration that it seems in most cases users could implement their logical
with simply calling timer.trigger() (I think the repeat registration is indeed
a problem, but I think we could ignore the timers registered during
termination). Could you further enlighten me a bit on this part?Best,Yun
Gao------------------------------------------------------------------From:Piotr
Nowojski <pnowoj...@apache.org>Send Time:2022 Nov. 30 (Wed.) 17:10To:dev
<dev@flink.apache.org>Subject:Re: [DISCUSS] FLIP-269: Properly Handling the
Processing Timers on Job TerminationHi,I have a couple of remarks.First a
general one. For me the important part in the design of this API ishow to
expose this to Flink users in public interfaces. NamelyProcessFunction and
StreamOperator. InternalTimerService is an internalclass, so we can change it
and break it as needed in the future.For registering a handler like proposed by
Dawid:interface SomeHandlerName { void onTimer(/* whatever type it is */ timer,
Context ctx ) { }}makes sense to me. For the InternalTimerService I think it
doesn't mattertoo much what we do. We could provide a similar interface as for
theProcessFunction/StreamOperator, it doesn't have to be the same one. On
thecontrary, I think it shouldn't be the same, as part of this effort
weshouldn't be exposing the concept of `Namespaces` to the public facing API.Re
the "waitFor". Theoretically I see arguments why users might want to usethis,
but I'm also not convinced whether that's necessary in practice. Iwould be +1
either way. First version can be without this functionality andwe can add it
later (given that we designed a good place to add it in thefuture, like the
`Context` proposed by Dawid). But I'm also fine adding itnow if others are
insisting.Best,Piotrekśr., 30 lis 2022 o 09:18 Dawid Wysakowicz
<dwysakow...@apache.org>napisał(a):
WindowOperator is not implemented by users. I can see that
forInternalTimerService we'll needinterface PendingTimerProcessor<KEY,
NAMESPACE> {void onTimer(InternalTimer<KEY, NAMESPACE> timer)
{doHandleTimer(timer);}I don't see a problem with that.As you said
ProcessingTimeService is a user facing interface andcompletely unrelated to the
InternalTimerService. I don't see a reasonwhy we'd need to unify those.As for
the waitFor behaviour. Personally, I have not been convinced itis necessary.
Maybe it's just my lack of vision, but I can't think of ascenario where I'd use
it. Still if we need it, I'd go for something like:void onTimer(/* whatever
type it is */ timer, Context ctx ) {}interface Context {void
executeAtScheduledTime(Consumer<timer> handler);}That way you have independent
simple interfaces that need to work onlyin a single well defined scenario and
you don't need to match aninterface to multiple different cases.Best,DawidOn
30/11/2022 07:27, Yun Gao wrote:
Hi Dawid,Thanks for the comments!As a whole I'm also open to the API and I also
prefer to use simplebut flexible interfaces, but it still looks there are some
problem tojust let users to implement the termination actions.Let's take the
WindowOperator as an example. As seen in [1],in the timer processing logic it
needs to acquire the key / namespaceinformation bound to the timer (which is
only supported by the
InternalTimerService).
Thus if we want users to implement the same logic on termination, we
either let users
to trigger the timer handler directly or we also allows users to access
these piece of
information. If we go with the later direction, we might need to provide
interfaces like
interface PendingTimerProcessor<KEY, NAMESPACE> {void onTimer(Timer<KEY,
NAMESPACE> timer) {doHandleTimer(timer);}}class Timer<KEY, NAMESPACE> {long
getTimestamp();KEY getKey();NAMESPACE getNamespace();}Then we'll have the issue
that since we need the interface to handle
both of cases of
InternalTimerSerivce and raw ProcessTimeService, the later do not have
key and
namespace information attached, and its also be a bit inconsistency for
users to have to set
the KEY and NAMESPACE types.Besides, it looks to me that if we want to
implement behaviors like
waiting for, it might
be not simply reuse the time handler time, then it requires every
operator authors to
re-implement such waiting logics.
Moreover it still have the downside that if you call back to the
`onTimer` method after
`trigger` you have access to the Context which lets you register new
timers.
I think we could simply drop the timers registered during we start
processing the pending timers
on termination. Logically there should be no new data after termination.
I think I am not convinced to these arguments. First of all I'm afraid
there is no clear distinction
in that area what is runtime and what is not. I always found
`AbstracStreamOperator(*)` actually part
of runtime or Flink's internals and thus I don't find
`InternalTimerService` a utility, but a vital part
of the system. Let's be honest it is impossible to implement an
operator without extending from
`AbstractStreamOperator*`.What would be the problem with having a
proper implementation in
`InternalTimerService`? Can't we do it like this?:
I think the original paragraph is only explanation to that the interface
is harder to support if we
allows the users to implement the arbitrary logic. But since now we are
at the page with the callback
option, users could always be allowed to implement arbitrary logic no
matter we support timer.trigger()
or not, thus I think now there is no divergence on this point. I also
believe in we'll finally have some logic
similar to the proposed one that drain all the times and process it.Best,Yun
Gao[1]
https://github.com/apache/flink/blob/a7fdab8b23cddf568fa32ee7eb804d7c3eb23a35/flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/operators/windowing/WindowOperator.java#L488
<https://github.com/apache/flink/blob/a7fdab8b23cddf568fa32ee7eb804d7c3eb23a35/flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/operators/windowing/WindowOperator.java#L488
><https://github.com/apache/flink/blob/a7fdab8b23cddf568fa32ee7eb804d7c3eb23a35/flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/operators/windowing/WindowOperator.java#L488
<https://github.com/apache/flink/blob/a7fdab8b23cddf568fa32ee7eb804d7c3eb23a35/flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/operators/windowing/WindowOperator.java#L488
>
------------------------------------------------------------------From:Dawid
Wysakowicz <dwysakow...@apache.org>Send Time:2022 Nov. 28 (Mon.) 23:33To:dev
<dev@flink.apache.org>Subject:Re: [DISCUSS] FLIP-269: Properly Handling the
Processing Timers
on Job Termination
Do we really need to have separate methods for
triggering/waiting/cancelling. To me it sounds rather counterintuitive.
Whycan't users just execute whatever they want in the handler itself insteadof
additional back and forth with the system? Moreover it still have thedownside
that if you call back to the `onTimer` method after `trigger` youhave access to
the Context which lets you register new timers.
I find following approach much simpler:void onTimer(...)
{doHandleTimer(timestamp);}void processPendingTimer(...) {//
triggerdoHandleTimer(timestamp);// for cancel, simply do nothing...}Sorry I
might not make it very clear here. I think the difficulty with
supported setting the currentKey is a special issue for the callbackoptions (no
matter what the interface is) since it allows users to executelogic other than
the one registered with the timers. The complexity comesfrom that currently we
have two level of TimerServices: TheProcessingTimerService (there is no key)
and InternalTimerService (withkey). Currently only ProcessingTimerService is
exposed to the runtime andInternalTimerService is much more a utility to
implement the operator. Thenwith the current code, the runtime could only
access toProcessingTimerService on termination.
I think I am not convinced to these arguments. First of all I'm afraid
there is no clear distinction in that area what is runtime and what is not.I
always found `AbstracStreamOperator(*)` actually part of runtime orFlink's
internals and thus I don't find `InternalTimerService` a utility,but a vital
part of the system. Let's be honest it is impossible toimplement an operator
without extending from `AbstractStreamOperator*`.
What would be the problem with having a proper implementation in
`InternalTimerService`? Can't we do it like this?:
AbstractStreamOperator#finish()
{internalTimerService.finish();}InternalTimerService#finish() {while ((timer =
processingTimeTimersQueue.peek()) != null)
{keyContext.setCurrentKey(timer.getKey());processingTimeTimersQueue.poll();onEndOfInputHandler.processPendingTimer(timer);}}If
we only executes some predefined actions, we do not need to worry
about the implementation of InternalTimerService and just execute theregistered
timers. But if we allow users to execute arbitrary logic, weneed to be also
aware of the InternalTimerServices and parse the key fromthe timers stored in
it. I think we should always have method to overcomethis issue, but to support
the callback options would be more complex.
I am not sure, having "predefined actions" would be good enough that we
do not need to set a key. As a user I'd anyhow expect the proper key to beset
in processPendingTimer.
Best,DawidOn 24/11/2022 08:51, Yun Gao wrote:Hi Piotr / Divye, Very thanks for
the discussion! First IMO it seems we
have reached the consensus on the high-level API: Most operators shouldusually
have only one reasonable action to the pending timers ontermination, thus we
could let the operators to implement its own actionswith the low-level
interface provided. The only exception is theProcessFunction, with which users
might register customized timers, thususers might also defines the actions on
termination (If I havemisunderstandings here, please correct me). For the
low-level API, I couldget the benefits with the callback options: since in most
cases an operatorhas only one action to all the timers, its a waste for us to
store the sameflag for all the timers, also with a lot of code / state format
changes.But since it is enough for most users to simply trigger / cacnel
thetimers, it would be redundant for users to implement the logic twice.
Thusperhaps we might combine the benefits of the two options: We might have
aseparate interface public interface TimerHandlersOnTermination {
voidprocessPendingTimer(Timer timer, long currentTime); } public class Timer
{long getRegisteredTimestamp(); void trigger(); void waitFor(); voidcancel(); }
Then if an operator have implemented theTimerHandlersOnTermination interface,
on termination we could callprocessPendingTimer(xx) for every pending timers.
Users might simplytrigger / waitFor / cancel it, or execute some other logics
if needed. Thenfor the ProcessFunction we might have a similar interface
toprocessPendingTimer, except we might need to provide Context / Collector
tothe ProcessFunction. Do you think this would be a good direction? Also@Piotr
I don't see a problem here. Interface doesn't have to reflect that,only the
runtime must set the correct key context before executing thehandler dealing
with the processing time timers at the end of input/time.Sorry I might not make
it very clear here. I think the difficulty withsupported setting the currentKey
is a special issue for the callbackoptions (no matter what the interface is)
since it allows users to executelogic other than the one registered with the
timers. The complexity comesfrom that currently we have two level of
TimerServices: TheProcessingTimerService (there is no key) and
InternalTimerService (withkey). Currently only ProcessingTimerService is
exposed to the runtime andInternalTimerService is much more a utility to
implement the operator. Thenwith the current code, the runtime could only
access toProcessingTimerService on termination. If we only executes some
predefinedactions, we do not need to worry about the implementation
ofInternalTimerService and just execute the registered timers. But if weallow
users to execute arbitrary logic, we need to be also aware of
theInternalTimerServices and parse the key from the timers stored in it. Ithink
we should always have method to overcome this issue, but to supportthe callback
options would be more complex. Best, Yun
Gao------------------------------------------------------------------From:Divye
Kapoor <dkap...@pinterest.com.INVALID> <mailto:dkap...@pinterest.com.INVALID >
Send Time:2022 Nov. 24 (Thu.) 08:50To:dev <dev@flink.apache.org>
<mailto:dev@flink.apache.org > Cc:XenonDevelopment Team
<xenon-...@pinterest.com> <mailto:xenon-...@pinterest.com
Subject:Re: Re: [DISCUSS] FLIP-269: Properly Handling the Processing
Timers on Job Termination Sounds good. Looks like we're on the same
page.Thanks! Divye On Wed, Nov 23, 2022 at 2:41 AM Piotr Nowojski
<pnowoj...@apache.org> <mailto:pnowoj...@apache.org > wrote: Hi Divye Ithink we
are mostly on the same page. Just to clarify/rephrase: One thingto think about
- on EOF “trigger immediately” will mean that theasynchronous wait timeout
timers will also fire - which is undesirable Ididn't mean to fire all timers
immediately in all of the built-inoperators. Just that each built-in operator
can have a hard coded way(without a way for users to change it) to handle those
timers. Windowedoperators would trigger the lingering timers (flush
outputs),AsyncWaitOperator could just ignore them. The same way users could
registerEOF timer handlers in the ProcessFunction as Dawid Wysakowicz proposed,
we(as flink developers) could use the same mechanism to implement anybehaviour
we want for the built-in operators. There should be no need toadd any separate
mechanism. Best, Piotrek śr., 23 lis 2022 o 08:21 DivyeKapoor
<dkap...@pinterest.com.invalid> <mailto:dkap...@pinterest.com.invalid >
napisał(a): Thanks Yun/Piotrek, Somebrief comments inline below. On Tue, Nov
22, 2022 at 1:37 AM Piotr Nowojski<pnowoj...@apache.org>
<mailto:pnowoj...@apache.org > wrote: Hi, All inall I would agree with Dawid's
proposal. +1 We can add the flexibility ofhow to deal with the timers in the
low level API via adding a handler - ifsomeone needs to customize it, he will
always have a workaround. Note aftergiving it more thought, I agree that
registering some handlers is betterthan overloading the register timer method
and modifying the timer's state.+1. At the same time, we can force the most
sensible semantic that we thinkfor the couple of built-in operators, which
should be prettystraightforward (either ignore the timers, or fire them at
once). I agreethere might be some edge cases, that theoretically user might
want to waitfor the timer to fire naturally, but: 1. I'm not sure how common
inpractice this will be. If not at all, then why should we be complicatingthe
API/system? That’s fair. However, the specifics are very importanthere. One
thing to think about - on EOF “trigger immediately” will meanthat the
asynchronous wait timeout timers will also fire - which isundesirable (because
they are racing with the last async call). However,the issue is cleanly
resolved by waiting for the timer to be canceled whenthe last event is
processed. (“Wait for” case). Ignoring the timer has theleast justification.
Registering the handler as per Dawid’s proposal andhaving that handler
unregister the timers on EOF makes best sense. Thissolution also unifies the
trigger immediately case as that handler canreregister the timers for early
termination. The proposal: 1. Operatorreceives EOF 2. EOF timer handler
triggers 3. EOF handler adjusts theregistered timers for early trigger or
ignore. If wait-for behavior isdesired, timers are not changed. This is
controlled in client code. 4.Operator waits for all timers to drain/trigger.
(“Always”). There is nospecial handling for ignore/early trigger. 5. Operator
allows job toproceed with shutdown. The only api change needed is an EOF
handler. Theother agreement we need is that “Wait for” is the desired behavior
inprocessing time and that processing time is fundamentally different fromevent
time in this respect. (I have changed my thinking since the lastmail). 2. We
can always expand the API in the future, and let the useroverride the default
built-in behaviour of the operators via some setter onthe stream transformation
(`SingleOutputStreamOperator`), or via somecustom API DSL style in each of the
operators separately. This is notrequired. See above. Re forcing the same
semantics for processing timetimers as for event time ones - this is tempting,
but indeed I see apossibility that users need to adhere to some external
constraints whenusing processing time. +1. As above, we should consider the 2
casesfundamentally different in this area. Re: Yun - b) Another issue is
thatwhat if users use timers with different termination actions in the
sameoperator / UDF? For example, users use some kind of timeout (like
throwsexception if some thing not happen after some other thing), and also
somekind of window aggregation logic. In this case, without additional
tags,users might not be able to distinguish which timer should be canceled
andwhich time should be triggered ? as above. The EOF handler makes thechoice.
4. How could these scenarios adjust their APIs ? From the currentlisted
scenarios, I'm more tend to that as @Dawid pointed out, there mightbe only one
expected behavior for each scenario, thus it does not seems toneed to allow
users to adjust the behavior. Thus @Divye may I have a doubleconfirmation
currently do we have explicit scenarios that is expected tochange the different
behaviors for the same scenario? Wait-for behavior isprobably the only expected
behavior and any alterations should be from theEOF handler managing the
registered timers. Besides @Divye from the listedscenarios, I have another
concern for global configuration is that for onejob, different operators seems
to still might have different expectedbehaviors. For example, A job using both
Window operator andAsyncWaitOperator might have different requirements for
timers ontermination? Thank you for raising this case. This changed my
thinking.Based on your point, we should try and align on the “Wait-for” with
EOFhandler proposal. I’m withdrawing the “single-runtime-config” proposal.Best,
Divye
