So this probably doesn't belong in this thread, but here goes:
When you think of the metric system as source and reporters and sinks,
one has to consider what he source emits:
Either:
a) events for added/removed metrics
b) periodically emit the values of all metrics, with the plethora of
additional scope information the reporters might require
Approach a) obviously doesn't work in a distributed setting, but is
closest to the current approach.
Reporters access metrics / meta-info as needed, and we only use as much
resources as we actually need.
Approach b) does work in a distributed setting, and conceptually works
reasonable well with scheduled reporters (i.e., reporters that
periodically write data to the external system, but for reporters who
are polled on demand from some external system (prometheus, jmx) this
can cause resources to be wasted if the polling interval is larger than
the update interval or metrics aren't being polled at all.
Additionally, you'd have to include _a lot_ of metadata for each metric
to retain current functionality.
Naturally this approach also consumes additional network resources.
Another concern I have is that we're mixing concerns here; one being
about allowing users to process metrics in a datastream fashion (which
has quite a few caveats, like not being able to access metrics from the
dispatcher / RM since we surely aren't running user-code in them), the
other being about handling formats.
And I have to point out that the format problem only really applies to
kafka.
For other reporters we don't have this problem since the backends define
the format, and there are just easier ways to handle this without a
major rework (wrap kafka connector, have a factory for serialization
schemes, _done_)
I'd now suggest to move the process-as-a-datastream idea into a
different thread because it a) isn't _really_ connected to the reporter
it self and b) we already have enough points of contention.
As for having the connector in flink vs flink-packages, I'm constantly
amazed at how much value people attribute to the source of a reporter
being in Flink.
After all, there's _nothing_ stopping us from including additional
3rd-party reporters in the distribution.
There's also _nothing_ stopping us from linking to 3rd-party reporters
in the documentation.
In other words, all user-facing parts are agnostic to whether the
reporter is maintained by Flink or not, hence I'm not accepting the
argument anymore that something must be in Flink for it to be used. It
just doesn't make sense.
Having every widely used component within Flink is just not maintainable
in the long run, as well all know, hence I'm very much in favor of
having it maintained externally via flink-packages. That's the very
purpose of that site.
And, on a final note, there's also _nothing_ stopping us from adding
something to Flink after X <time_unit> if it becomes an integral in the
way Flink is being used.
On 20/11/2019 15:46, Gyula Fóra wrote:
@Becket , Yun:
Regarding the core/ecosystem project:
I don't completely agree with your arguments regarding why this should be
an external ecosystem project instead of part of the Flink repo.
A metric connector is relevant for the Flink users, not the metric store.
Metric storage systems don't care about where logs are coming but Flink job
authors need a way to get the metrics to whatever systems they have. The
same applies for other connectors. If we don't provide canonical ways of
communicating with external systems, be it sources, sinks or metrics that
makes everyones life a bit harder.
Historically most of the connectors went straight to Flink and over time
the maintenance of these has become quite a challenge with Flink core
itself growing rapidly. I agree that we have to make these decision and not
include every new external connector to the Flink core. I think this
decision should be based on the value it brings to users, and how often it
will be used. These are not easy questions and the Flink ecosystem website
is a great way for gauging the popularity/value of a specific connector.
Another way of deciding this would be to talk to the Flink community (like
we do with this thread) and see if this is a common pattern and if we can
come up with a good generic solution regarding Kafka versioning and formats
that will work for most. If we see big interest here and have a consensus
on the formats I don't see any reason why we shouldn't include it.
Regarding the message format:
The idea with the JSON format was that it could be an easy-to-use source
for downstream metric systems to integrate with it. I don't have much
experience with different metric storage systems so maybe Yun you are right
that you will always end up having another processor for this. But even in
that case JSON is a pretty safe format as it is easy to process no matter
what you use.
Otherwise I agree that a pluggable format would be much better and more
generic. We just need to find a way to keep it simple :D
@Bowen, Chensay
The whole idea of making a metrics reporter source sounds pretty great at
first :) If we could do this that would definitely make this more flexible
but even then you probably need some sort of a serialization schema
implementation for the metrics by default. Which is basically what the
kafka reporter would do + minimal client.
Chesnay, I don't completely understand what you mean by:
"Periodically emitting the values of all metrics goes against the
convention which we established"
Isn't this exactly what the Kafka metrics reporter would do anyways?
If you could elaborate on this a bit more that would be very helpful for me
because I don't have a good overview of the key design principals of the
current metric system.
Cheers,
Gyula
On Wed, Nov 20, 2019 at 10:22 AM Chesnay Schepler <ches...@apache.org>
wrote:
@Bowen I can see where you're coming from, but I don't think this would
work too well. Your "stream" would have to contain events for
added/removed metrics, but metrics are inherently not Serializable. I
think this would end up being a weird special case.
(Periodically emitting the values of all metrics goes against the
convention which we established from the very beginning that metrics
should only incur costs if necessary; as such a reporter that polls on
demand should only consume resources if it was called)
Additionally, there are plans to add additional methods to the reporter
in the future, at which point the source interface would no longer
suffice. At that point you'd need a separate interface again, and
wrappers for your sinks.
This would result in what is the trivial solution for this reporter
right now anyway: have the reporter use a kafka connector internally,
with all the features that if offers.
Overall I think we'd be unnecessarily coupling reporters to the source
interface, and i don't see a true benefit.
On 19/11/2019 19:47, Bowen Li wrote:
Hi,
What still unclear to me so far is - As I don't see any yet., what would
be
the fundamental differences between this Kafka reporter and Flink’s
existing Kafka producer?
I’ve been thinking of Flink metrics for a while, and the “metric
reporter”
feels a bit redundant to me. As you may already knew, Flink has been used
to process external metrics in various companies. If you think about it,
Flink’s own metric system is no different from external ones and actually
just another stream source, and metric reporters are just some data sinks
writing to external storage, with no guarantee or checkpointing.
So instead of adding Kafka or other MQ reporters and worrying about
message
format (which are already solved by Flink’s sinks), we can generalize and
expose Flink’s metrics system to be a simple built-in stream source, and
"metric reporters" are just some customized sink tailored for this
source.
Users may even be able to access and process it in stream environment
with
data stream api. That give users full flexibility on manipulating Flink
metrics with Flink, and it’s more of a “eat your own dogfood” philosophy.
This seems too good to be true, and I haven't had time to think of the
details. Let me know if I miss anything here.
On Mon, Nov 18, 2019 at 09:51 Yun Tang <myas...@live.com> wrote:
Hi all
Glad to see this topic in community.
We at Alibaba also implemented a kafka metrics reporter and extend it to
other message queues like Alibaba cloud log service [1] half a year ago.
The reason why we not launch a similar discussion is that we previously
thought we only provide a way to report metrics to kafka. Unlike current
supported metrics reporter, e.g. InfluxDB, Graphite, they all have an
easy-to-use data source in grafana to visualize metrics. Even with kafka
metrics reporter, we still need another way to consume data out and
work as
a data source for observability platform, and this would be diverse for
different companies.
I think this is the main concern to include this in a popular
open-source
main repo, and I pretty agree with Becket's suggestion to contribute
this
as a flink-package and we could offer an end-to-end solution including
how
to visualize these metrics data.
[1] https://www.alibabacloud.com/help/doc-detail/29003.htm
Best
Yun Tang
On 11/18/19, 8:19 AM, "Becket Qin" <becket....@gmail.com> wrote:
Hi Gyula,
Thanks for bringing this up. It is a useful addition to have a
Kafka
metrics reporter. I understand that we already have Prometheus and
DataDog
reporters in the Flink main repo. However, personally speaking, I
would
slightly prefer to have the Kafka metrics reporter as an ecosystem
project
instead of in the main repo due to the following reasons:
1. To keep core Flink more focused. So in general if a component is
more
relevant to external system rather than Flink, it might be good to
keep it
as an ecosystem project. And metrics reporter seems a good example
of
that.
2. This helps encourage more contributions to Flink ecosystem
instead
of
giving the impression that anything in Flink ecosystem must be in
Flink
main repo.
3. To facilitate our ecosystem project authors, we have launched a
website[1] to help the community keep track of and advertise the
ecosystem
projects. It looks a good place to put the Kafka metrics reporter.
Regarding the message format, while I think use JSON by default is
fine as
it does not introduce much external dependency, I wonder if we
should
make
the message format pluggable. Many companies probably already have
their
own serde format for all the Kafka messages. For example, maybe
they
would
like to just use an Avro record for their metrics instead of
introducing a
new JSON format. Also in many cases, there could be a lot of metric
messages sent by the Flink jobs. JSON format is less efficient and
might
have too much overhead in that case.
Thanks,
Jiangjie (Becket) Qin
[1] https://flink-packages.org/
On Mon, Nov 18, 2019 at 3:30 AM Konstantin Knauf <
konstan...@ververica.com>
wrote:
> Hi Gyula,
>
> thank you for proposing this. +1 for adding a
KafkaMetricsReporter.
In
> terms of the dependency we could go a similar route as for the
"universal"
> Flink Kafka Connector which to my knowledge always tracks the
latest
Kafka
> version as of the Flink release and relies on compatibility of
the
> underlying KafkaClient. JSON sounds good to me.
>
> Cheers,
>
> Konstantin
>
>
>
>
>
> On Sun, Nov 17, 2019 at 1:46 PM Gyula Fóra <gyf...@apache.org>
wrote:
>
> > Hi all!
> >
> > Several users have asked in the past about a Kafka based
metrics
reporter
> > which can serve as a natural connector between arbitrary metric
storage
> > systems and a straightforward way to process Flink metrics
downstream.
> >
> > I think this would be an extremely useful addition but I would
like to
> hear
> > what others in the dev community think about it before
submitting a
> proper
> > proposal.
> >
> > There are at least 3 questions to discuss here:
> >
> >
> > *1. Do we want the Kafka metrics reporter in the Flink repo?*
As it is
> > much more generic than other metrics reporters already
included, I
would
> > say yes. Also as almost everyone uses Flink with Kafka it
would be
a
> > natural reporter choice for a lot of users.
> > *2. How should we handle the Kafka dependency of the
connector?*
> > I think it would be an overkill to add different Kafka
versions here,
> > so I would use Kafka 2.+ which has the best compatibility and
is
future
> > proof
> > *3. What message format should we use?*
> > I would go with JSON for readability and compatibility
> >
> > There is a relevant JIRA open for this already.
> > https://issues.apache.org/jira/browse/FLINK-14531
> >
> > We at Cloudera also promote this as a scalable way of pushing
metrics to
> > other systems so we are very happy to contribute an
implementation
or
> > cooperate with others on building it.
> >
> > Please let me know what you think!
> >
> > Cheers,
> > Gyula
> >
>
>
> --
>
> Konstantin Knauf | Solutions Architect
>
> +49 160 91394525
>
>
> Follow us @VervericaData Ververica <https://www.ververica.com/>
>
>
> --
>
> Join Flink Forward <https://flink-forward.org/> - The Apache
Flink
> Conference
>
> Stream Processing | Event Driven | Real Time
>
> --
>
> Ververica GmbH | Invalidenstrasse 115, 10115 Berlin, Germany
>
> --
> Ververica GmbH
> Registered at Amtsgericht Charlottenburg: HRB 158244 B
> Managing Directors: Timothy Alexander Steinert, Yip Park Tung
Jason,
Ji
> (Tony) Cheng
>
