flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aljoscha Krettek <aljos...@apache.org>
Subject Re: events eviction
Date Mon, 15 Feb 2016 10:57:59 GMT
Hi,
you are right, the logic is in EvictingNonKeyedWindowOperator.emitWindow() for non-parallel
(non-keyed) windows and in EvictingWindow.processTriggerResult() in the case of keyed windows.

You are also right about the contract of the Evictor, it returns the number of elements to
be evicted from the beginning. This also means that eviction does not consider any timestamps
in the elements and is therefore quite arbitrary. The places in the code I mentioned above
simply get the value from the Evictor and evict that many elements from the internal buffer/state.

Right now it is not possible to replace the window operator that is used by flink. What you
can do is copy the window operator code and use it manually using DataStream.transform().

About the evictor state. I’m afraid this is not possible right now. It was a conscious decision
to make the Evictor stateless to make it easier for the system to handle. I would also strongly
advise against using Evictors if at all possible. They make it impossible to incrementally
aggregate window results (for example with a reduce function). This can have a huge performance/memory
footprint impact. In your case, what are you using them for?

I hope this helps somehow, but let us know if you need further explanations.

Cheers,
Aljoscha

> On 15 Feb 2016, at 11:09, Radu Tudoran <radu.tudoran@huawei.com> wrote:
> 
> Hello,
>  
> I am looking over the mechanisms of evicting events in Flink. I saw that either using
a default evictor or building a custom one the logic is that the evictor will provide the
number of events to be discarded.
> Could you please provide me with some additional pointers regarding the mechanism in
Flink where this actually happens:
> -          The class that implements this functionality of discarding the events? (my
initial expectations that this happens in the window class turn out to be wrong). I checked
and found the “EvictingNonKeyedWindowOperator” – is this the right place to look indeed?
> -          If yes, would it be possible to create a customizable class like this one
and somehow pass it to the framework? I would be curious if there is an option  other than
modifying the core classes and recompiling the framework?
>  
> On a slightly parallel topic - is there some way of creating a state in the evictor that
will be check pointed and restore in case of failure.  I would be interested if something
like an operator state is possible in the evictor.
>  
> Regards,
>  
> Dr. Radu Tudoran
> Research Engineer - Big Data Expert
> IT R&D Division
>  
> <image001.png>
> HUAWEI TECHNOLOGIES Duesseldorf GmbH
> European Research Center
> Riesstrasse 25, 80992 München
>  
> E-mail: radu.tudoran@huawei.com
> Mobile: +49 15209084330
> Telephone: +49 891588344173
>  
> HUAWEI TECHNOLOGIES Duesseldorf GmbH
> Hansaallee 205, 40549 Düsseldorf, Germany, www.huawei.com
> Registered Office: Düsseldorf, Register Court Düsseldorf, HRB 56063,
> Managing Director: Bo PENG, Wanzhou MENG, Lifang CHEN
> Sitz der Gesellschaft: Düsseldorf, Amtsgericht Düsseldorf, HRB 56063,
> Geschäftsführer: Bo PENG, Wanzhou MENG, Lifang CHEN
> This e-mail and its attachments contain confidential information from HUAWEI, which is
intended only for the person or entity whose address is listed above. Any use of the information
contained herein in any way (including, but not limited to, total or partial disclosure, reproduction,
or dissemination) by persons other than the intended recipient(s) is prohibited. If you receive
this e-mail in error, please notify the sender by phone or email immediately and delete it!


Mime
View raw message