I feel reporting the current low watermark for each operator is better than
just reporting the source watermark when I see Flink 1.3 web frontend.
We want the smallest watermark in all splits. But Some runners, like
FlinkRunner, don't have a way to get the global smallest watermark, and the
metric's type(Counter, Guage, Distribution) can not express it.
Best,JingsongLee
------------------------------------------------------------------
From:Ben Chambers <bchamb...@google.com.INVALID>
Time:2017 Jun 2 (Fri) 21:46
To:dev <dev@beam.apache.org>; JingsongLee <lzljs3620...@aliyun.com>
Cc:Aviem Zur <aviem...@gmail.com>; Ben Chambers <bchamb...@google.com.invalid>
Subject:Re: [DISCUSS] Source Watermark Metrics
I think having runners report important, general properties such as the
source watermark is great. It is much easier than requiring every source to
expose it.
I'm not sure how we would require this or do so in a general way. Each
runner has seperate code for handling the watermark as well as different
ways information should be reported.
Where would the runner do this? Where would the runner.put these values?
Maybe this is just part of the documentation about what we would like
runners to do?
On Fri, Jun 2, 2017, 3:09 AM Aljoscha Krettek <aljos...@apache.org> wrote:
> Hi,
>
> Thanks for reviving this thread. I think having the watermark is very
> good. Some runners, for example Dataflow and Flink have their own internal
> metric for the watermark but having it cross-runner seems beneficial (if
> maybe a bit wasteful).
>
> Best,
> Aljoscha
>
> > On 2. Jun 2017, at 03:52, JingsongLee <lzljs3620...@aliyun.com> wrote:
> >
> > @Aviem Zur @Ben Chambers What do you think about the value of
> METRIC_MAX_SPLITS?
> >
> >
>
>------------------------------------------------------------------From:JingsongLee
> <lzljs3620...@aliyun.com>Time:2017 May 11 (Thu)
> 16:37To:dev@beam.apache.org <dev@beam.apache.org>Subject:[DISCUSS] Source
> Watermark Metrics
> > Hi everyone,
> >
> > The source watermark metrics show the consumer latency of Source.
> > It allows the user to know the health of the job, or it can be used to
> > monitor and alarm.
> > We should have the runner report the watermark metricsrather than
> > having the source report it using metrics. This addresses the fact that
> even
> > if the source has advanced to 8:00, the runner may still know about
> buffered
> > elements at 7:00, and so not advance the watermark all the way to 8:00.
> > The metrics Includes:
> > 1.Source watermark (`min` amongst all splits):
> > type = Gauge, namespace = io, name = source_watermark
> > 2.Source watermark per split:
> > type = Gauge, namespace = io.splits, name = <split_id>.source_watermark
> >
> > Min Source watermark amongst all splits seems difficult to implement
> since
> > some runners(like FlinkRunner) can't access to all the splits to
> aggregate
> > and there is no such AggregatorMetric.
> >
> > So We could report watermark per split and users could use a `min`
> > aggregation on this in their metrics backends. However, as was mentioned
> > in the IO metrics proposal by several people this could be problematic in
> > sources with many splits.
> >
> > So we do a check when report metrics to solve the problem of too many
> splits.
> > {code}
> > if (splitsNum <= METRIC_MAX_SPLITS) {
> > // set the sourceWatermarkOfSplit
> > }
> > {code}
> >
> > So I'd like to take a discussion to the implement of source watermark
> metrics
> > and specific how many splits is too many. (the value of
> METRIC_MAX_SPLITS)
> >
> > JIRA:
> > IO metrics (https://issues.apache.org/jira/browse/BEAM-1919)
> > Source watermark (https://issues.apache.org/jira/browse/BEAM-1941)
> >
>
>
