You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@flink.apache.org by Jan Nehring <ja...@dfki.de> on 2017/02/06 12:39:37 UTC
stream clustering in flink
Hi,
we want to cluster a stream of Tweets using Flink. Every incoming tweet
is compared to the last 100 tweets. After this comparison, a cluster ID
is assigned to the tweet. We try to find out the best approach how to
solve this:
1. Using a stream window of the last tweets seems to be difficult
because we would need to cross join this window with every incoming
tweet. According to my research the Flink API does not support cross
joins on stream windows.
2. We could also store the last 100 tweets in one operator with
parallelism=1. This would work but it introduces a bottleneck.
3. We could share the last 100 tweets as a "shared state" among the
operator that assigns the cluster. But every tweet changes the state so
there would be a lot of synchronization effort between the operators.
Are you aware of other possible solutions? Currently solution #2 seems
the most promising to me but I do not like the bottleneck.
Best regards Jan
Re: stream clustering in flink
Posted by Aljoscha Krettek <al...@apache.org>.
Hi,
I think distributed stream clustering is still a somewhat open field. I'm
not aware of popular open source systems that have implementations for that
(except maybe Apache SAMOA). Maybe you will have some luck if you try to
search for "distributed stream clustering" papers.
Cheers,
Aljoscha
On Mon, 6 Feb 2017 at 13:39 Jan Nehring <ja...@dfki.de> wrote:
> Hi,
>
> we want to cluster a stream of Tweets using Flink. Every incoming tweet
> is compared to the last 100 tweets. After this comparison, a cluster ID
> is assigned to the tweet. We try to find out the best approach how to
> solve this:
>
> 1. Using a stream window of the last tweets seems to be difficult
> because we would need to cross join this window with every incoming
> tweet. According to my research the Flink API does not support cross
> joins on stream windows.
> 2. We could also store the last 100 tweets in one operator with
> parallelism=1. This would work but it introduces a bottleneck.
> 3. We could share the last 100 tweets as a "shared state" among the
> operator that assigns the cluster. But every tweet changes the state so
> there would be a lot of synchronization effort between the operators.
>
> Are you aware of other possible solutions? Currently solution #2 seems
> the most promising to me but I do not like the bottleneck.
>
> Best regards Jan
>