Hi Paul,
Sorry for overlooking the "offset translation" MM2 feature. I'm glad
Ryanne was able to confirm this would work.
I'm just one voice, but FWIW, I think that the internal topic naming
scheme is a public API. We document the structure of the naming
scheme in several places. We also recommend making use of the fact
that the applicationId is a prefix of the topic name in conjunction with
Kafka Broker ACLs to grant access to the internal topics to the
applications that own them.
Actually, for this latter reason, I'm concerned that giving more control
over the names of internal topics might make topic security and
access control more difficult. Or maybe this concern is off-base, and
folks who take advanced control over the topic name would also take
on the responsibility to make sure their naming scheme works in
conjunction with their broker configs.
For whatever reason, I hadn't considered prefixing the application's
id with "pre-prod.". Offhand, I think this would achieve the desired
outcome. There may be some devil in the details, of course.
Glad to hear, by the way, that you've already considered the problem
of concurrent modifications to the changelogs (etc.). It sounds like
your plan should work, although it might become a management burden
if you start wanting to run a lot of these stream-app tests. In that case,
you could consider mirroring the relevant topics *again* into a
test-specific
prefix (like "pre-prod.test-1.", up to some point. Then, you could stop
the mirror, run the test, verify the results, and then just delete the
whole test dataset.
Does it seem like you have a good path forward? From what I'm
hearing, the "user-space" approach is at least worth exploring before
considering a new API. Of course, if it doesn't pan out for whatever reason,
I'd (personally) support adding whatever features are necessary to support
your use case.
Thanks,
-John
On Mon, Mar 25, 2019 at 9:40 PM Paul Whalen <pgwha...@gmail.com> wrote:
> John and Ryanne,
>
> Thanks for the responses! I think Ryanne's way of describing the question
> is actually a much better summary than my long winded description: "a
> Streams app can switch between topics with and without a cluster alias
> prefix when you migrate between prod and pre-prod, while preserving state."
>
> To address a few of John's points...
>
> But, the prod app will still be running, and its changelog will still be
> > mirrored into pre-prod when you start the pre-prod app.
> >
> The idea is actually to turn off the mirroring from prod to pre-prod during
> this period, so the environments can operate completely independently and
> their state can comfortably diverge during the testing period. After the
> testing period we'd be happy to throw away everything in pre-prod and start
> mirroring again from prod with a blank slate.
>
> Also, the pre-prod app won't be in the same consumer group as the prod app,
> > so it won't know from what offset to start processing input.
> >
> This is where I'm hoping the magic of MM2 will come in - at the time we
> shut off mirroring from prod to pre-prod in order to spin of the pre-prod
> environment, we will do an "offset translation" with RemoteClusterUtils
> like Ryanne mentioned, so new Streams apps in pre-prod will see consumer
> offsets that make sense for reading from pre-prod topics.
>
> I like both of your ideas around the "user space" solution: subscribing to
> multiple topics, or choosing a topic based on config. However, in order to
> populate their internal state properly, when the pre-prod apps come up they
> will need to look for repartition and changelog topics with the right
> prefix. This seems problematic to me since the user doesn't have direct
> control over those topic names, though it did just occur to me now that the
> user *sort of* does. Since the naming scheme is currently just
> applicationId + "-" + storeName + "-changelog", we could translate the
> consumer group offsets to a consumer group with a new name that has the
> same prefix as the mirrored topics do. That seems a bit clumsly/lucky to
> me (is the internal topic naming convention really a "public API"?), but I
> think it would work.
>
> I'd be curious to hear if folks think that solution would work and be an
> acceptable pattern, since my original proposal of more user control of
> internal topic naming did seem a bit heavy handed.
>
> Thanks very much for your help!
> Paul
>
> On Mon, Mar 25, 2019 at 3:14 PM Ryanne Dolan <ryannedo...@gmail.com>
> wrote:
>
> > Hey Paul, thanks for the kind words re MM2.
> >
> > I'm not a Streams expert first off, but I think I understand your
> question:
> > if a Streams app can switch between topics with and without a cluster
> alias
> > prefix when you migrate between prod and pre-prod, while preserving
> state.
> > Streams supports regexes and lists of topics as input, so you can use
> e.g.
> > builder.stream(List.of("topic1", "prod.topic1")), which is a good place
> to
> > start. In this case, the combined subscription is still a single stream,
> > conceptually, but comprises partitions from both topics, i.e. partitions
> > from topic1 plus partitions from prod.topic1. At a high level, this is no
> > different than adding more partitions to a single topic. I think any
> > intermediate or downstream topics/tables would remain unchanged, since
> they
> > are still the result of this single stream.
> >
> > The trick is to correctly translate offsets for the input topics when
> > migrating the app between prod and pre-prod, which RemoteClusterUtils can
> > help with. You could do this with external tooling, e.g. a script
> > leveraging RemoteClusterUtils and kafka-streams-application-reset.sh. I
> > haven't tried this with a Streams app myself, but I suspect it would
> work.
> >
> > Ryanne
> >
> >
> > On Sun, Mar 24, 2019 at 12:31 PM Paul Whalen <pgwha...@gmail.com> wrote:
> >
> > > Hi all,
> > >
> > > With MirrorMaker 2.0 (
> > >
> > >
> >
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-382%3A+MirrorMaker+2.0
> > > )
> > > accepted and coming along very nicely in development, it has got me
> > > wondering if a certain use case is supported, and if not, can changes
> be
> > > made to Streams or MM2 to support it. I'll explain the use case, but
> the
> > > TL;DR here is "do we need more control over topic naming in MM2 or
> > > Streams?"
> > >
> > > My team foresees using MM2 as a way to mirror data from our prod
> > > environment to a pre-prod environment. The data is supplied by
> external
> > > vendors, introduced into our system through a Kafka Streams ETL
> pipeline,
> > > and consumed by our end-applications. Generally we would only like to
> > run
> > > the ETL pipeline in prod since there is an operational cost to running
> it
> > > in both prod and pre-prod (the data sometimes needs manual attention).
> > > This seems to fit MM2 well: pre-prod end-applications consume from the
> > > pre-prod Kafka cluster, which is entirely "remote" topics being
> mirrored
> > > from the prod cluster. We only have to keep one instance of the ETL
> > > pipeline running, but end-applications can be separate, connecting to
> > their
> > > respective prod and pre-prod Kafka clusters.
> > >
> > > However, when we want to test changes to the ETL pipeline itself, we
> > would
> > > like to turn off the mirroring from prod to pre-prod, and run the ETL
> > > pipeline also in pre-prod, picking up the most recent state of the prod
> > > pipeline from when mirroring was turned off (FWIW, downtime is not an
> > issue
> > > for our use case).
> > >
> > > My question/concern is basically, can Streams apps work when they're
> > > running against topics prepended with a cluster alias, like
> > > "pre-prod.App-statestore-changelog" as is the plan with MM2. From what
> I
> > > can tell the answer is no, and my proposal would be to give the Streams
> > > user more specific control over how Streams names its internal topics
> > > (repartition and changelogs) by defining an
> "InternalTopicNamingStrategy"
> > > or similar. Perhaps there is a solution on the MM2 side as well, but
> it
> > > seems much less desirable to budge on that convention.
> > >
> > > I phrased the question in terms of my team's problem, but it's worth
> > noting
> > > that this use case is passably similar to a potential DR use case,
> where
> > > there is a DR cluster that is normally just being mirrored to by MM2,
> but
> > > in a DR scenario would become the active cluster that Streams
> > applications
> > > are connected to.
> > >
> > > Thanks for considering this issue, and great job to those working on
> MM2
> > so
> > > far!
> > >
> > > Paul
> > >
> >
>
