flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stephan Ewen <se...@apache.org>
Subject Re: Watermarks as "process completion" flags
Date Mon, 30 Nov 2015 13:13:56 GMT
Hi!

If you implement the "Checkpointed" interface, you get the function calls
to "snapshotState()" at the point when the checkpoint barrier arrives at an
operator. So, the call to "snapshotState()" in the sink is when the barrier
reaches the sink. The call to "checkpointComplete()" in the sources comes
after all barriers have reached all sinks.

Have a look here for an illustration about barriers flowing with the
stream:
https://ci.apache.org/projects/flink/flink-docs-release-0.10/internals/stream_checkpointing.html

Stephan


On Mon, Nov 30, 2015 at 11:51 AM, Anton Polyakov <polyakov.anton@gmail.com>
wrote:

> Hi Stephan
>
> thanks that looks super. But source needs then to emit checkpoint. At the
> source, while reading source events I can find out that - this is the
> source event I want to take actions after. So if at ssource I can then emit
> checkpoint and catch it at the end of the DAG that would solve my problem
> (well, I also need to somehow distinguish my checkpoint from Flink's
> auto-generated ones).
>
> Sorry for being too chatty, this is the topic where I need expert opinion,
> can't find out the answer by just googling.
>
>
> On Mon, Nov 30, 2015 at 11:07 AM, Stephan Ewen <sewen@apache.org> wrote:
>
>> Hi Anton!
>>
>> That you can do!
>>
>> You can look at the interfaces "Checkpointed" and "checkpointNotifier".
>> There you will get a call at every checkpoint (and can look at what records
>> are before that checkpoint). You also get a call once the checkpoint is
>> complete, which corresponds to the point when everything has flown through
>> the DAG.
>>
>> I think it is nice to implement it like that, because it works
>> non-blocking: The stream continues while the the records-you-wait-for flow
>> through the DAG, and you get an asynchronous notification once they have
>> flown all the way through.
>>
>> Greetings,
>> Stephan
>>
>>
>> On Mon, Nov 30, 2015 at 11:03 AM, Anton Polyakov <
>> polyakov.anton@gmail.com> wrote:
>>
>>> I think I can turn my problem into a simpler one.
>>>
>>> Effectively what I need - I need way to checkpoint certain events in
>>> input stream and once this checkpoint reaches end of DAG take some action.
>>> So I need a signal at the sink which can tell "all events in source before
>>> checkpointed event are now processed".
>>>
>>> As far as I understand flagged record don't quite work since DAG doesn't
>>> propagate source events one-to-one. Some transformations might create 3
>>> child events out of 1 source. If I want to make sure I fully processed
>>> source event, I need to wait till all childs are processed.
>>>
>>>
>>>
>>> On Sun, Nov 29, 2015 at 4:12 PM, Anton Polyakov <
>>> polyakov.anton@gmail.com> wrote:
>>>
>>>> Hi Fabian
>>>>
>>>> Defining a special flag for record seems like a checkpoint barrier. I
>>>> think I will end up re-implementing checkpointing myself. I found the
>>>> discussion in flink-dev:
>>>> mail-archives.apache.org/mod_mbox/flink-dev/201511.mbox/…
>>>> <http://mail-archives.apache.org/mod_mbox/flink-dev/201511.mbox/%3CCA+faj9xDFAUG_zi==E2H8s-8R4cn8ZBDON_hf+1Rud5pJqvZ4A@mail.gmail.com%3E>
which
>>>> seems to solve my task. Essentially they want to have a mechanism which
>>>> will mark record produced by job as “last” and then wait until it’s
fully
>>>> propagated through DAG. Similarly to what I need. Essentially my job which
>>>> produces trades can also thought as being finished once it produced all
>>>> trades, then I just need to wait till latest trade produced by this job is
>>>> processed.
>>>>
>>>> So although windows can probably also be applied, I think propagating
>>>> barrier through DAG and checkpointing at final job is what I need.
>>>>
>>>> Can I possibly utilize internal Flink’s checkpoint barriers (i.e. like
>>>> triggering a custom checkoint or finishing streaming job)?
>>>>
>>>> On 24 Nov 2015, at 21:53, Fabian Hueske <fhueske@gmail.com> wrote:
>>>>
>>>> Hi Anton,
>>>>
>>>> If I got your requirements right, you are looking for a solution that
>>>> continuously produces updated partial aggregates in a streaming fashion.
>>>> When a  special event (no more trades) is received, you would like to store
>>>> the last update as a final result. Is that correct?
>>>>
>>>> You can compute continuous updates using a reduce() or fold() function.
>>>> These will produce a new update for each incoming event.
>>>> For example:
>>>>
>>>> val s: DataStream[(Int, Long)] = ...
>>>> s.keyBy(_._1)
>>>>   .reduce( (x,y) => (x._1, y._2 + y._2) )
>>>>
>>>> would continuously compute a sum for every key (_._1) and produce an
>>>> update for each incoming record.
>>>>
>>>> You could add a flag to the record and implement a ReduceFunction that
>>>> marks a record as final when the no-more-trades event is received.
>>>> With a filter and a data sink you could emit such final records to a
>>>> persistent data store.
>>>>
>>>> Btw.: You can also define custom trigger policies for windows. A custom
>>>> trigger is called for each element that is added to a window and when
>>>> certain timers expire. For example with a custom trigger, you can evaluate
>>>> a window for every second element that is added. You can also define
>>>> whether the elements in the window should be retained or removed after the
>>>> evaluation.
>>>>
>>>> Best, Fabian
>>>>
>>>>
>>>>
>>>> 2015-11-24 21:32 GMT+01:00 Anton Polyakov <polyakov.anton@gmail.com>:
>>>>
>>>>> Hi Max
>>>>>
>>>>> thanks for reply. From what I understand window works in a way that it
>>>>> buffers records while window is open, then apply transformation once
window
>>>>> close is triggered and pass transformed result.
>>>>> In my case then window will be open for few hours, then the whole
>>>>> amount of trades will be processed once window close is triggered. Actually
>>>>> I want to process events as they are produced without buffering them.
It is
>>>>> more like a stream with some special mark versus windowing seems more
like
>>>>> a batch (if I understand it correctly).
>>>>>
>>>>> In other words - buffering and waiting for window to close, then
>>>>> processing will be equal to simply doing one-off processing when all
events
>>>>> are produced. I am looking for a solution when I am processing events
as
>>>>> they are produced and when source signals "done" my processing is also
>>>>> nearly done.
>>>>>
>>>>>
>>>>> On Tue, Nov 24, 2015 at 2:41 PM, Maximilian Michels <mxm@apache.org>
>>>>> wrote:
>>>>>
>>>>>> Hi Anton,
>>>>>>
>>>>>> You should be able to model your problem using the Flink Streaming
>>>>>> API. The actions you want to perform on the streamed records
>>>>>> correspond to transformations on Windows. You can indeed use
>>>>>> Watermarks to signal the window that a threshold for an action has
>>>>>> been reached. Otherwise an eviction policy should also do it.
>>>>>>
>>>>>> Without more details about what you want to do I can only refer you
to
>>>>>> the streaming API documentation:
>>>>>> Please see
>>>>>> https://ci.apache.org/projects/flink/flink-docs-release-0.10/apis/streaming_guide.html
>>>>>>
>>>>>> Thanks,
>>>>>> Max
>>>>>>
>>>>>> On Sun, Nov 22, 2015 at 8:53 PM, Anton Polyakov
>>>>>> <polyakov.anton@gmail.com> wrote:
>>>>>> > Hi
>>>>>> >
>>>>>> > I am very new to Flink and in fact never used it. My task (which
I
>>>>>> currently solve using home grown Redis-based solution) is quite simple
- I
>>>>>> have a system which produces some events (trades, it is a financial
system)
>>>>>> and computational chain which computes some measure accumulatively
over
>>>>>> these events. Those events form a long but finite stream, they are
produced
>>>>>> as a result of end of day flow. Computational logic forms a processing
DAG
>>>>>> which computes some measure over these events (VaR). Each trade is
>>>>>> processed through DAG and at different stages might produce different
set
>>>>>> of subsequent events (like return vectors), eventually they all arrive
into
>>>>>> some aggregator which computes accumulated measure (reducer).
>>>>>> >
>>>>>> > Ideally I would like to process trades as they appear (i.e.
stream
>>>>>> them) and once producer reaches end of portfolio (there will be no
more
>>>>>> trades), I need to write final resulting measure and mark it as “end
of day
>>>>>> record”. Of course I also could use a classical batch - i.e. wait
until all
>>>>>> trades are produced and then batch process them, but this will be
too
>>>>>> inefficient.
>>>>>> >
>>>>>> > If I use Flink, I will need a sort of watermark saying - “done,
no
>>>>>> more trades” and once this watermark reaches end of DAG, final
measure can
>>>>>> be saved. More generally would be cool to have an indication at the
end of
>>>>>> DAG telling to which input stream position current measure corresponds.
>>>>>> >
>>>>>> > I feel my problem is very typical yet I can’t find any solution.
>>>>>> All examples operate either on infinite streams where nobody cares
about
>>>>>> completion or classical batch examples which rely on fact all input
data is
>>>>>> ready.
>>>>>> >
>>>>>> > Can you please hint me.
>>>>>> >
>>>>>> > Thank you vm
>>>>>> > Anton
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>
>>
>

Mime
View raw message