flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aljoscha Krettek <aljos...@apache.org>
Subject Re: Keys distribution insights
Date Tue, 06 Jun 2017 13:48:53 GMT
Hi,

There is no way of doing it with any Flink UI but you could try and do it manually: in your
job, instead of doing the actual computation just count how many elements you have per key
(in your GroupReduce). Then put a MapPartition right after the GroupReduce (which should preserve
the same partitioning) and inside that see what keys you have and how many elements you had
per key. With this you know which partition, i.e. which parallel instance had which keys and
how many they were.

Best,
Aljoscha

> On 5. Jun 2017, at 12:01, Flavio Pompermaier <pompermaier@okkam.it> wrote:
> 
> Hi everybody,
> in my job I have a groupReduce operator with parallelism 4 and one of the sub-tasks takes
a huge amount of time (wrt the others).
> My guess is that the objects assigned to that slot have much more data to reduce (an
thus are somehow computationally heavy within the groupReduce operator). 
> What I'm trying to understand which keys are assigned to that slot: is there any way
(from the JobManager UI or from the logs) to investigate the keys distribution (that from
the plan visualization is the result of an hash partition)?
> 
> Best,
> Flavio


Mime
View raw message