2:09 a.m.
Hi Cheng and Tomek and anyone else on wildfly-dev,
One of the main things I'm hoping we can do for WildFly 20 is get a
first-cut (perhaps tech preview) implementation of bootable jar support. A
big part of the bootable jar story is the user using Galleon during the
build of their application + server jar to produce a server customized to
their requirements. The key thing that makes that usable is having Galleon
layers for the functionality they want.
EJB is a big part of our functionality and we don't have layers for it yet,
so I'd like to get started on adding some. I've been thinking some about
what that might look like.
A bit of context: a layer is a form of API, in that once we support a
layer, if users use it to get some functionality, in the future they should
be able to continue to use that layer and get that same functionality. So
before introducing a layer we should be sure it does what we want, and also
that it doesn't by mistake do 'extra' things that we couldn't remove in
the
future.
Our EJB subsystem provides a wide variety of features, and I can think of a
great number of use cases where users might want just some of them (say
local SLSBs only) and would like a slimmer server without the rest. It
would be great if in the future we could provide a number of layers to help
with that. But for the first cut at this I think we should keep it simple.
(Also, the more layers we have the more testing permutations we need to
deal with, which is doable but takes time.)
If a user is trying to slim their server I think of four key things they
want to do in descending order of importance:
1) Eliminate unnecessary open ports.
2) Eliminate unnecessary exposure
3) Eliminate jars, both to save filesystem footprint and to reduce any
theoretical attack surface.
4) Reduce configuration clutter and unnecessary services.
Given that I suggest we have the following layers:
1) ejb3-lite. This is like what we provide in standalone.xml, but without
the MDB instance pool or the remote connector resource. (A
webservices layer would depend on this.)
2) ejb3. Builds (i.e. depends) on ejb3-lite and adds the discovery
subsystem, the MDB pool and the remote connector resource. This looks what
we provide in standalone-full.xml, except no IIOP integration.
3) ejb3-iiop. Builds (i.e. depends) on *ejb3-lite* and adds the iiop
subsystem plus the integration resource in the EJB3 subsystem. This *could*
depend on *ejb3* instead, but that means EJB is unnecessarily exposed over
the HTTP interface and a lot of dependencies are brought in for messaging
integration.
A user could provision ejb3 + ejb3-iiop and get the equivalent to what's in
standalone-full.xml.
There'd be a couple ancillary layers as well:
4) ejb-local-cache. This provides the infinispan subsystem resources
related to local ejb caching (i.e. the ones in standalone.xml and
standalone-full.xml.) The ejb3-lite layer *optionally* depends on this.
It's an optional dependency because the user when provisioning can
*exclude* this layer and instead provision...
5) ejb-dist-cache. This provides the infinispan subsystem resources related
to distributed ejb caching (i.e. the ones in standalone0ha.xml and
standalone-full-ha.xml.)
This pattern of "exclude an optional local caching layer and add a
distributed variant" is how web session and jpa caching are handled in the
layers for those features.
https://github.com/bstansberry/wildfly/commits/ejb-layers2 illustrates what
the layer-specs could look like. (There's no effort there to adapt the
subsystem to require fewer deps; it's just illustrating the config
generation aspects.)
I think the main work here (besides testing) would be examining what if any
module dependencies in the org.jboss.as.ejb3 module could be made optional.
TBH I don't see big FS footprint improvements coming out of these
permutations. Big EJB dependencies include (in no particular order.)
a) IIOP. But the TM uses IIOP libs so they will be there regardless.
b) Remoting. It would be nice to be able to eliminate the need for
remoting, e.g. for ejb3-lite if the user app isn't a remote ejb client. But
that doesn't seem trivial and if a server is manageable remoting will be
there anyway to support the CLI.
c) JCA stuff. But if there's a datasource it's there anyway.
d) Infinispan. A benefit of a future SLSB-only layer would be this dep
could be eliminated. (But it might be used for web caching etc anyway.)
e) Remote artemis integration. This we do save if ejb3-lite is used.
The general testing strategy we've been doing as we've developed layers is
in testsuite/integration/smoke and basic we provision a server with a
particular set of layers and then add a surefire execution that runs a
subset of the tests that can work against the features in that server. That
hasn't been too painful.
WDYT?
Best regards,
- Brian