ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Raul Kripalani <ra...@apache.org>
Subject Re: Ignite modularisation: client/server separation
Date Mon, 03 Aug 2015 15:01:19 GMT
Hi Sergey,

That's great if you think that the OSGi enablement can be achieved without
separating client from server. My concern with OSGifying all of current's
ignite-core is that we would be incurring the cost of OSGi-enabling the
server side when, quite honestly, there's probably little interest in that.
Folks will be interested in interacting as clients with Ignite servers from
OSGi, not in running Ignite servers within OSGi containers.

Out of over 200 Camel components hosted at the ASF, there are only around
12 that are *not* OSGi-friendly. When we create a component that leverages
a library/framework that's not OSGi-compatible, we try to engage with the
relevant communities to enhance the library in that direction. Or if the
communities are not in a position to collaborate and the OSGi enablement
can be achieved without structural changes (sometimes it's just a matter of
adding the right manifest headers – but I get the impression Ignite's
server-side will be more complex), we create a wrapper in the ServiceMix
Bundles community and publish it to Maven Central.

The ~ 12 components which are not OSGi-friendly are young, mostly
introduced in Camel >= 2.15. So it's ok to create a camel-ignite component
without OSGi support for now, but we should be having a roadmap to not
frustrate our OSGi users (as I said, we have a significant base) and for
the component to be accepted into Camel without objections.

With regards to the separation itself, I still think a JAR of nearly 7mb is
too big a burden for all clients. This is not an opinion about "right
engineering practices", but my practical opinion as an Ignite user. I get
excited with the applications of Ignite in the IoT – for pure caching, one
can use the Memcached client, but the compute capability, service registry,
messaging, etc. make it really appealing in this context. And given the
resource scarcity inherent in IoT devices, I don't see why a mobile client
or an embedded device should spare 7mb for the library when it won't even
use 20% of its capability given that it'll never act as a server.

Regards,

*Raúl Kripalani*
Apache Camel PMC Member & Committer | Enterprise Architect, Open Source
Integration specialist
http://about.me/raulkripalani | http://www.linkedin.com/in/raulkripalani
http://blog.raulkr.net | twitter: @raulvk

On Mon, Aug 3, 2015 at 2:23 PM, Sergi Vladykin <sergi.vladykin@gmail.com>
wrote:

> Raul,
>
> As far as I understand your main concern is that Ignite is incompatible
> with OSGi (other points seem to be more about your personal feelings about
> "right engineering practices" than about real problems). I believe that we
> can achieve this compatibility without such a big refactoring of
> everything. But do we really need that if it is not a requirement for
> Camel?
>
> BTW client/server API of Ignite is intentionally unified to make it easier
> to use (simply switch one flag in config instead of rewriting your code to
> completely different API). In this case unification works better than
> separation of concerns.
>
> Sergi
>
> 2015-08-03 14:31 GMT+03:00 Raul Kripalani <raulk@apache.org>:
>
> > Hello guys,
> >
> > Spoke to Dmitriy over the weekend about the modularisation of Ignite.
> >
> > At Apache Camel we strive to make all our components (including the
> future
> > camel-ignite) OSGi compatible because have a significant user base
> > deploying Camel apps on Apache Karaf.
> >
> > I see some groundwork before we can aspire to make OSGi components
> > communicate with Ignite:
> >
> >    * There seems to be no concept of a client. Client-side and
> server-side
> > coexist in ignite-core. There's no code separation, so a client wanting
> to
> > communicate with an Ignite topology will end up importing the server-side
> > implementation too.
> >
> >    * ... and the server-side implementation uses classloading constructs
> > (e.g. Zero Deployment) which may prove hard to engineer for compatibility
> > with OSGi. This re-enginering is a waste of time IMHO because there's no
> > interest in running Ignite servers on top of OSGi - only clients.
> >
> >    * The ignite-core JAR (6.7mb) is too heavy for a component that's
> > nothing but a client. Some average Java client API sizes for reference:
> > hazelcast (400kb), activemq-client (1.2 mb), kafka (300kb).
> >
> >    * Possible dependency leak. Currently not a problem as we lack 3rd
> party
> > deps in ignite-core. But if we introduce any, we'll impose our dep
> versions
> > on the client creating possible classpath hell. Or seen from another
> > perspective: I don't know the reason we have no 3rd party deps, and I'm
> > sure we've had to reinvent the wheel at some point... Perhaps we avoided
> > them because we knew we would be imposing them on the client? If that's
> the
> > case, separating server and client will bring lots of flexibility.
> >
> >    * This architecture/design entails a lack of separation of concerns,
> > IMO. While it's true that a client can start an embedded Ignite instance
> > (and for that it'll need the server-side code), not all clients will do.
> In
> > my opinion, ignite-java-client (hypothetical name) should discover
> > ignite-core (just the server-side impl) in the classpath if starting an
> > embedded Ignite is requested.
> >
> > What's your opinion? Obviously such a refactoring is a large undertaking
> > and not for immediate action. If the community shares this vision, I
> would
> > start thinking about this for 2.0.
> >
> > P.S.: With regards to Camel, a warning on the camel-ignite doc page for
> not
> > being OSGi-compliant will suffice.
> >
> > Regards,
> >
> > *Raúl Kripalani*
> > Apache Camel PMC Member & Committer | Enterprise Architect, Open Source
> > Integration specialist
> > http://about.me/raulkripalani | http://www.linkedin.com/in/raulkripalani
> > http://blog.raulkr.net | twitter: @raulvk
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message