On Nov 22, 2012, at 11:41 AM, Vincent Massol <[email protected]> wrote:
On Nov 22, 2012, at 11:28 AM, Vincent Massol <[email protected]> wrote:
On Feb 15, 2012, at 4:58 PM, Vincent Massol <[email protected]> wrote:
On Feb 15, 2012, at 4:10 PM, Vincent Massol wrote:
On Jan 25, 2012, at 5:02 PM, Vincent Massol wrote:
On Jan 15, 2012, at 11:27 AM, Vincent Massol wrote:
Hi devs,
I've started an experiment to have colocated functional tests (CFT), which means having the functional tests located where the functional domain sources are located instead of in XE.
For example for the linkchecker module we have the following directories:
xwiki-platform-linkchecker/ |_ xwiki-platform-linkchecker-refresher (JAR) |_ xwiki-platform-linkchecker-ui (XAR) |_ xwiki-platform-linkchecker-tests (functional tests)
The rationale for this was: * Have everything about a functional domain self-contained (source and all tests) * Making it easy to run only tests for a given functional domain * Move page objects to the functional domain too
Here are some findings about this experiment:
A - It takes about 30 seconds to generate the adhoc packaging and start XWiki. This would be done for each module having functional tests compared to only once if all tests were executed in XE B- The package mojo created to generate a full packaging is quite nice and I plan to reuse it in lots of other places in our build (distributions, database, places where we need XWiki configuration files) C- We will not be able to run platform builds in Maven multithreaded mode since it would mean that several XWiki instance could be started at the same time on the same port D- The colocated functional test module
Solutions/ideas:
* One idea to overcome A and C would to have the following setup: ** Keep functional test modules colocated but have them generate a test JAR ** Still allow running functional tests from the colocated module (this makes it easy to verify no regression was introduced when making changes to a given domain) ** Have functional tests in XE depend on the colocated functional test module JARs and configure Jenkins to run all functional tests from XE only
* Another solution to overcome C is to auto-discover the port to use in our XWiki startup script (and save it in a file so that the stop script can use it).
I think the first proposal is the best one and brings the best of the 2 worlds.
Actually there's another one:
* Run functional tests in platform * Configure jenkins so that we have one jenkins job per platform top module
This would allow several things: * Faster feedback to user: for example if a commit is done in a module that happens late in the build order the whole platform wouldn't need to be rebuilt before noticing the problem) - Note: Jenkins supports a "Local subdirectory for repo" feature that would make this possible I think)
After researching this a bit I don't think it's possible…
This "Local subdirectory for repo" is something else: "Specify a local directory (relative to the workspace root) where the Git repository will be checked out. If left empty, the workspace root itself will be used."
We could check out the whole Git repo but 1) it'll take space (a lot of space) but more importantly 2) GitHub will trigger a build on all jobs for that repo which isn't good.
Regarding space, I've tested it and I've also tested the "shallow clone" option of Jenkins.
So for platform: * with shallow clone: 281MB * without shallow clone (ie full clone): 346MB
We have 91 modules in platform so that's 91*281=25571MB == 25GB Even with the full clone that's 31GB
We need to add the target/ dir size, so let's say 1GB more in total, to 26GB or 32GB. That's doable! :)
So I guess it's ok and that's not our biggest issue.
There are 2 more issues: * Automatic creation of jenkins job. That's doable and almost done: http://jira.xwiki.org/jira/browse/XCOMMONS-87 * Triggers. This is the hardest since we'll need to create a manual ordering of jobs as otherwise they'll all start building at the same time when there's a github event coming in… It's not hard to do but it's maintenance intensive. Of course this ordering would be defined in the automated jenkins job creations but it still needs to be maintained...
In conclusion it's doable and we need to work on http://jira.xwiki.org/jira/browse/XCOMMONS-87 first.
Actually maybe a better thing to do (much simpler) would simply be to use parallel builds for platform and instead ensure that our functional tests there use different ports so that they can run in //...
That's provided the agents are powerful enough to support a large # of threads. Would be interesting to compute the ideal # of threads based on our dependency graph :)
Yet another solution is to build platform without the -Pintegration-tests profile with parallel builds and then create one job per functional test in platform. -Vincent
Thanks -Vincent
So I think we'll have to drop the idea of having a single job per module.
See also http://stackoverflow.com/questions/8922857/jenkins-how-to-build-multiple-top...
hmm actually this is what we already do for functional tests but we don't trigger than on Git Hub notifications, we trigger them on xwiki-enterprise (after it's been built).
I guess we could do the same, i.e.:
* Only trigger the top level module in a given repo using a GitHub push * Only trigger the other modules on "Build whenever a SNAPSHOT dependency is built"
The problem is that it would still use a lot of space. For example my xwiki-platform/ dir takes 1.2GB of disk space (with target/ dirs) So with roughly 230+ modules that's 276 GB just for xwiki-platform…
This computation is wrong, see above.
Thanks -Vincent
Combined with commons, rendering and enterprise that's probably around 500GB easily…
In conclusion it doesn't seem like a good idea…
Thanks -Vincent
Thanks -Vincent
* Use the full power of our agents by distributing the jobs * Since functional tests run at the same time as unit tests if the module build works we can be sure it's good
Thus even though packaging and starting XE takes a lot of time (I've timed it to roughly 1.5 minutes on my machine) the balancing of jobs on multiple agents might overcome this - it would still be great to be able to improve the XE start time.
Note that creating jenkins job for so many modules might seem daunting but this can be easily automated with http://ci.xwiki.org/api/ or even better http://evgeny-goldin.com/wiki/Maven-jenkins-plugin
I'm still hesitating between the previous idea and this one but thought I would share my current thoughts with you in case you have some idea/preference… :)
Thanks -Vincent
WDYT?
Thanks -Vincent