I've thought about not unpacking car files but rather flattening
everything in them and writing a classloader that can deal with
offsets within a jar file. This would completely solve the problem
except for peoples desire to fiddle with jsp files on the fly.
For this, maybe we should allow a car file has the config.ser (in
xml :-) as now but the rest of it is a map to the actual locations of
the files? I think it might be possible to include this in either a
second "development-friendly" config store (in a special location) or
the standard config store.
thanks
david jencks
On Apr 7, 2006, at 10:29 AM, Dain Sundstrom wrote:
Please stop blaming the m2 repo structure for this problem. The m2
repo structure only increased the path of our longest path by 36
characters. The true problem is that David and I moved the
unpacked configurations into the repository. We did this because
of the chunkiness of the numbered directories in the config-store
directory. The m2 repository structure makes querying the
repository for version numbers possible and it is this querying
that makes optional version numbers possible.
I think we have two issues that both must be addressed:
1) The ears we generate in our build have very long internal paths,
154 characters. This is just bad form, and vastly reduces the user
path head room.
2) We need to move the unpacked ears our of the repository and into
a separate flat directory structure.
I can look at the second one later today after fixing the redeploy
command. Can someone take a look at getting our build to jar up
the classes and compiled jsps in our build. I'll fix the generated
classes in our build.
-dain
On Apr 7, 2006, at 6:50 AM, Matt Hogstrom wrote:
Thinking about this some more I believe we need to make a good
decision here as having to revisit this issue in the future will
cause users to have to change how the server works. I've been
talking to a new user that has a larger server farm and is very
interested in the Geronimo server as their new foundation.
However, they run a few thousand servers and are VERY sensitive to
changes in the behaviour of the server in terms of how it impacts
them. Changes to the repsoistory will affect their operational
experience dramatically and they do run Windows (go Bill Gates).
They are watching this thread with keen interest. Their biggest
concern is changing how their build and distribution system works
and changes in this area is highly disruptive for them.
My view of the problem is that there are really three distinct
areas of a path. They are the user area, the server area and the
application area. Let me splain...
| 0000000000000000000000000000 |
11111111111111111111111111111111111111111 | 2222222222222222222 ...
C:\my\directory\before\geronimo\geronimo-1.1\repository
\com.apache.geronimo\console-1.1\appArtifacts
The area in the 0's are controlled by the user and we need to
leave more headroom than a few characters so they can manage
multiple deployments of Geronimo; this could include multiple
versions or multiple deployments. The users probably enjoy
flexibility in naming as much as we do. We don't have control
over this but we influence how much headroom is available.
The 1's is really the area we have control over as this is the
server proper. This includes the area from the top of the tree to
the end of where the files we create end. So, for instance, this
includes var, repository, etc. Since were currently experiencing
this problem in the respository I think we should focus on this area.
Finally, the 2's are the area that include the application and
Maven dependent information. The Maven naming convention is
verbose. The current implementation needs to be changed, the
question is how and can the change survive several releases so
that our users are not forced to change their deployments on each
subsequent release. *One immediate thought I had was to place
applications back into the config-store (or equivalent name).
Rather than simply use a number as we did previously perhaps the
configId of the deployment would be appropriate. Its human
readable and would be shorter than the current maven structure.*
I highlighted the previous as I think this is the best option
based on what I know today.
Perhaps there some way to provide a Maven abstraction that would
map Maven dir names to an internal format for us. I expect if we
are running into this its only a matter of time befoew other Maven
users experience the same issues. For us its the nesting of Maven
articacts / configurations that is causing us the problem. Jason,
thoughts?
Whatever we decide we need to ensure that it is stable enough to
work for a period of time.
Matt
Dain Sundstrom wrote:
Man I hate Windows....
Anyway, if you have a real OS and list the files in an assembly,
you will see that the problem is caused by the combination of
two changes: we now keep configurations in the repository and we
unpack them. If you look closer you will see that the big
offenders are unpacked ears and wars.
I believe the following are the longest paths in the server:
(270)
geronimo-1.1-SNAPSHOT/repository/geronimo/daytrader-derby-jetty/
1.1- SNAPSHOT/daytrader-derby-jetty-1.1-SNAPSHOT.car/daytrader-
web-1.1- SNAPSHOT.war/META-INF/geronimo-generated/org/apache/
geronimo/axis/ client/GenericServiceEndpointWrapper$
$EnhancerByCGLIB$$36344d29.class
(264)
geronimo-1.1-SNAPSHOT/repository/geronimo/webconsole-jetty/1.1-
SNAPSHOT/webconsole-jetty-1.1-SNAPSHOT.car/geronimo-console-
standard-1.1-SNAPSHOT.war/WEB-INF/classes/org/apache/geronimo/
console/ databasemanager/wizard/DatabasePoolPortlet
$ResourceAdapterParams.class
One thing to note here is that the longest paths are all classes
generated by Geronimo, nested classes in wars or compiled JSP
pages. Someone should look into makeing maven jar the latter
two and Geronimo should be creating jars when generating classes
(actually we should stop generating classes a head of time but
that is another story).
Breaking down the longest path, we have:
GeronimoName (22)
geronimo-1.1-SNAPSHOT
RepositoryPath (55)
repository/geronimo/daytrader-derby-jetty/1.1-SNAPSHOT
FileName (39)
daytrader-derby-jetty-1.1-SNAPSHOT.car
NestedPath (154)
daytrader-web-1.1-SNAPSHOT.war/META-INF/geronimo-generated/org/
apache/geronimo/axis/client/GenericServiceEndpointWrapper$
$EnhancerByCGLIB$$36344d29.class
The first thing to note is if we simply replace "SNAPSHOT" with
"0", we drop 28 characters which makes the longest path 242; not
enough head room. Of course, when we switch our groupId to the
maven standard org.apache.geronimo we eat up 20 more
characters. If we are going to unpack war files there is very
little we can do about the NestedPath, so we have very few
choices left. If we simply combine combine ${GeronimoName}/$
{FileName}/${NestedPath} we are up to 115 characters leaving
only 41 characters for anything else, but when you add back the
28 from "SNAPSHOT", you get to a more comfortable level.
I think if we combine this problem with Sachin's request for a
separate directory for applications, we could do something like
this:
${GeronimoName}/apps/${FileName}/${NestedPath}
There are several problems with this. I think users will confuse
the hot-deploy directory "deploy" with the "apps" directory
[1]. Then again, if you look at the problem configurations they
are all apps the users may want to remove (sample apps and the
console), so may be we should just put these in the hot-deploy
directory. Another problem is that it will be much more
difficult to query a repository without a directory structure.
The server will basically have to read the configuration from
these apps on startup to determine what they are, so again we
may just want to use the hot-deploy directory. I'm not a fan of
the hot-deploy directory, but I'm not sure there is a better
solution.
Again I renew my hate of Windows...
/me shakes his fist at Bill Gates
-dain
[1] As a side issue, I prefer the name "apps" because it will be
most familiar to tomcat users.
