Thanks for the work, it’s definitely needed!
I’d like to suggest to take it one step further.
To me, I’d like to see Kafka Connect the same way we have Docker and Docker
repositories.
Here’s how I would envision the flow:
- Kafka Connect workers are just workers. They come with no jars whatsoever
- The REST API allow you to add a config to the connect cluster
- The workers, seeing the config, pull the jars from the available (maven?)
repositories (public or private)
- Classpath isolation comes into play so that the pulled jar doesn’t interact
with other connectors
- Additionally, I believe the config should have a “tag” or “version” (like
docker really), so that
o you can run different versions of the same connector on your connect cluster
o configurations are strongly linked to a connector (right now if I update my
connector jars, I may break my configuration)
I know this is a bit out of scope, but if major changes are coming to connect
then these are my two cents.
Finally, maybe extend that construct to Transformers. The ability to
externalise transformers as jars would democratize their usage IMO
On 3/5/17, 4:24 am, "Ewen Cheslack-Postava" <e...@confluent.io> wrote:
Thanks for the KIP.
A few responses inline, followed by additional comments.
On Mon, May 1, 2017 at 9:50 PM, Konstantine Karantasis <
konstant...@confluent.io> wrote:
> Gwen, Randall thank you for your very insightful observations. I'm glad
you
> find this first draft to be an adequate platform for discussion.
>
> I'll attempt replying to your comments in order.
>
> Gwen, I also debated exactly the same two options: a) interpreting absence
> of module path as a user's intention to turn off isolation and b)
> explicitly using an additional boolean property. A few reasons why I went
> with b) in this first draft are:
> 1) As Randall mentions, to leave the option of using a default value open.
> If not immediately in the first version of isolation, maybe in the future.
> 2) I didn't like the implicit character of the choice of interpreting an
> empty string as a clear intention to turn isolation off by the user. Half
> the time could be just that users forget to set a location, although
they'd
> like to use class loading isolation.
> 3) There's a slim possibility that in rare occasions a user might want to
> avoid even the slightest increase in memory consumption due to class
> loading duplication. I admit this should be very rare, but given the other
> concerns and that we would really like to keep the isolation
implementation
> simple, the option to turn off this feature by using only one additional
> config property might not seem too excessive. At least at the start of
this
> discussion.
> 4) Debugging during development might be simpler in some cases.
> 5) Finally, as you mention, this could allow for smoother upgrades.
>
I'm not sure any of these keep you from removing the extra config. Is there
any reason you couldn't have clean support for relying on the CLASSPATH
while still supporting the classloaders? Then getting people onto the new
classloaders does require documentation for how to install connectors, but
that's pretty minimal. And we don't break existing installations where
people are just adding to the CLASSPATH. It seems like this:
1. Allows you to set a default. Isolation is always enabled, but we won't
include any paths/directories we already use. Setting a default just
requires specifying a new location where we'd hold these directories.
2. It doesn't require the implicit choice -- you actually never turn off
isolation, but still support the regular CLASSPATH with an empty list of
isolated loaders
3. The user can still use CLASSPATH if they want to minimize classloader
overhead
4. Debugging can still use CLASSPATH
5. Upgrades just work.
>
> Randall, regarding your comments:
> 1) To keep its focus narrow, this KIP, as well as the first implementation
> of isolation in Connect, assume filesystem based discovery. With careful
> implementation, transitioning to discovery schemes that support broader
> URIs I believe should be easy in the future.
>
Maybe just mention a couple of quick examples in the KIP. When described
inline it might be more obvious that it will extend cleanly.
> 2) The example you give makes a good point. However I'm inclined to say
> that such cases should be addressed more as exceptions rather than as
being
> the common case. Therefore, I wouldn't see all dependencies imported by
the
> framework as required to be filtered out, because in that case we lose the
> advantage of isolation between the framework and the connectors (and we
are
> left only with isolation between connectors).
3) I tried to abstract implementation details in this the KIP, but you are
> right. Even though filtering here is mainly used semantically rather than
> literally, it gives an implementation hint that we could avoid.
>
I think we're missing another option -- don't do filtering and require that
those dependencies are correctly filtered out of the modules. If we want to
be nicer about this, we could also detect maybe 2 or 3 classes while
scanning for Connectors/Converters/Transformations that indicate the
classloader has jars that it shouldn't and warn about it. I can't think of
that many that would be an issue -- basically connect-api, connect-runtime
if they really mess it up, and maybe slf4j.
> 4) In the same spirit as in 3) I believe we should reserve enough
> flexibility to the implementation to discover and load classes, when they
> appear in multiple locations under the general module location.
>
>
And a couple of addition comments:
- module.path should be module.paths if it accepts multiple paths
- I think the description of the module paths is a bit confusing because I
think for simpler configuration we actually want to specify the *parent*
directory of module paths. I definitely prefer this since it's simpler
although I am not certain how it will mix with future extensions to other
URL handlers (where a zip file or http URL wouldn't do the same discovery
within subdirectories)
-Ewen
> Thanks again! Let me know what you think.
> Konstantine
>
>
> On Mon, May 1, 2017 at 10:12 AM, Randall Hauch <rha...@gmail.com> wrote:
>
> > Very nice work, Konstantine. Conflicting dependencies of connectors is
> > indeed a big issue that makes it hard to manage installed connectors.
> >
> > I do like Gwen's idea about removing the 'module.isolation.enabled'
> > property. However, I would have anticipated always using classpath
> > isolation for *only* those components registered under the module path
> and
> > not really for anything else already on the normal classpath. So, people
> > could continue to place custom connector JARs onto the classpath, though
> > this would become deprecated in favor of installing custom connector
> JARs /
> > modules via the module path. This keeps configuration simple, gives
> people
> > time to migrate, but let's people that need classpath isolation get it
to
> > install a variety of connectors each with their dependencies that
> > potentially conflict with other components.
> >
> > The challenge is whether there should be a default for 'module.path'.
> > Ideally there would be so that users know where they can install their
> > connectors. However, I suspect that this might be difficult to do unless
> it
> > can make use of system properties such as "${kafka.home}" so that
> relative
> > directories can be specified.
> >
> > A few other questions/comments:
> >
> > 1) Does the KIP have to specify how are components / modules installed,
> > discovered, or recognized by Kafka Connect? Or perhaps the KIP needs to
> > just specify the semantics of the file system module path (e.g., the
> > directories below those specified in the module path are to be unique
and
> > identify an installed component).
> >
> > 2) Will the module classloader filtering also have to exclude Kafka
> Connect
> > dependencies? The only one that I can think of is the SLF4J API, which
> > can't be loaded from the module's classloader if the connector is to
send
> > its log messages to the same logging system.
> >
> > 3) Rather than specify filtering, would be it a bit more flexible to
> simply
> > say that the implementation will need to ensure that Java, Kafka
Connect,
> > and other third party APIs (e.g., SLF4J API) will not be loaded from the
> > module classloaders? It'd be better to avoid specifying how it will be
> > done, just in case the implementation needs to evolve or use a different
> > technique (e.g., load the Java and public Kafka Connect APIs via one
> > classloader that is reused and that always appears before the module
> > classloader, while Kafka Connect implementation JARs appear after the
> > component's classloader.
> >
> > 4) Perhaps to address #2 and #3 above, perhaps the KIP could explicitly
> > specify the classloader order for a deployed connector. For example,
> > 'java', 'kafka-connect-apis', 'connector-module', 'smt-module-1', ...,
> > 'kafka-connect-impls', where 'connector-module' is the classloader for
> the
> > (first) module where the connector is found, 'smt-module-1' is the
> > classloader for the (first) module where the first SMT class is found
(if
> > specified and found in a separate module), 'smt-module-2' is the
> > classloader .... Might also need to say that the KIP does not specify
how
> > the implementation will pick the module if a specified class if found in
> > more than one module.
> >
> > Thoughts?
> >
> > Randall
> >
> > On Mon, May 1, 2017 at 6:43 AM, Gwen Shapira <g...@confluent.io> wrote:
> >
> > > Hi Konstantine,
> > >
> > > Thank you so much for driving this! The connector classpath mess is
> > driving
> > > me nuts (or worse, driving me to use Docker).
> > >
> > > I like the proposal for micro-benchmarks to test the context switching
> > > overhead.
> > >
> > > I have a difficult time figuring out the module.isolation.enabled.
> > > Especially with a default to false. I can't think of a reason that
> anyone
> > > will not want classpath isolation. "No! I want my connectors to mess
up
> > > each other's dependencies" said no one ever.
> > >
> > > So it looks like this is mostly for upgrade purpose? Because the
> initial
> > > upgrade will not have the module.path set and therefore classpath
> > isolation
> > > will simply not work by default?
> > >
> > > In that case, why don't we simply use the existence of non-empty
> > > module.path as an indicator of whether isolation should work or not?
> seem
> > > simpler and intuitive to me.
> > >
> > > Thanks!
> > >
> > > Gwen
> > >
> > >
> > >
> > >
> > >
> > > On Sat, Apr 29, 2017 at 9:16 AM, Konstantine Karantasis <
> > > konstant...@confluent.io> wrote:
> > >
> > > > * Because of KIP number collision, please disregard my previous KIP
> > > > announcement and use this thread for discussion instead *
> > > >
> > > >
> > > > Hi everyone,
> > > >
> > > > we aim to address dependency conflicts in Kafka Connect soon by
> > applying
> > > > class loading isolation.
> > > >
> > > > Feel free to take a look at KIP-146 here:
> > > >
> > > > *https://cwiki.apache.org/confluence/display/KAFKA/KIP-
> > > > 146+-+Classloading+Isolation+in+Connect
> > > > <https://cwiki.apache.org/confluence/display/KAFKA/KIP-
> > > > 146+-+Classloading+Isolation+in+Connect>*
> > > >
> > > > which describes minimal required changes to public interfaces and
the
> > > > general implementation approach.
> > > >
> > > > This is a much wanted feature for Kafka Connect. Your feedback is
> > highly
> > > > appreciated.
> > > >
> > > > -Konstantine
> > > >
> > >
> > >
> > >
> > > --
> > > *Gwen Shapira*
> > > Product Manager | Confluent
> > > 650.450.2760 | @gwenshap
> > > Follow us: Twitter <https://twitter.com/ConfluentInc> | blog
> > > <http://www.confluent.io/blog>
> > >
> >
>
