On 2/1/2011 3:15 AM, Vladimir Prus wrote:
John Maddock wrote:
Maybe I suggest that for some time, we outright ban freeform discussion about process, and instead, we restrict them to threads started by a Boost developers and saying this: "I am maintainer of X, and had N commits and M trac changes in the last year. I most hate P1, P2 and P3. I would propose that we use T1, T2, and T3 to fix that". Then, everybody could join to suggest better way of fixing P1, P2 and P3 -- without making up other supposed problems.
OK let me give my pet hates:
* The only tool comment I have is that SVN is awfully slow for big merges (Math lib docs for example), I probably need to find a better way of using the tool better though.
I can't shake the feeling that SVN performance is specific to our instance, at least other SVN servers I use feel faster. It would be worthwhile to experiment with different setups, including using svn+ssh instead of https, or using FSFS repository format on the server (if it uses BDB) now.
There are various problems with SVN. Like using HTTPS, which is known to be unstable. The contention between Trac and SVN is problematic because we have a heavily used Trac and it conflicts with regular SVN use frequently. We do use FSFS repo format, but it's not the latest sharded structure. Both the HTTPS and un-sharded aspects are something I intend to change. But I'm giving priority at the moment to the test reporting problems. Since they seem the most critical.
Alas, I am not sure anybody is in position to try this.
As for trying the plain SVN server configuration.. I'm also not sure if we can try it (and obviously I don't have time at the moment) as I don't know what firewall changes, or server management changes, might need to happen. And that's something I can't do.
* I think we could organize the testing more efficiently for faster turnaround and better integration testing, and much to my surprise I'm coming round to Robert Ramey's suggestion that we reorganize testing on a library-by-library basis, with each library tested against the current release/stable branch.
* Test machine pulls changes for lib X from version control (whatever tool that is). * Iff there are changes (either to lib X or to release), only then run the tests for that library against current release branch. * The testers machine builds it's own test results pages - ideally these should go into some form of version control as well so we can roll back and see what broke when. * When a tester first starts testing they would add a short meta-description to a script, and run the script to generate the test results index pages. ie there would be no need for a separate machine collecting and processing the results. * The test script should run much of the above *in parallel* if requested.
The aim would be to speed processing of testing by reducing the cycle time (most libraries most of the time don't need re-testing).
I suppose an alternative approach would be just make the incremental testing work. Boost.Build, obviously, can rebuild and rerun just the necessary tests, but the regression framework used to have issues, like reporting stale tests. I think it should give the same increase in testing time, and not really sure which approach is harder to implement.
Implementing the incremental testing is the easiest, assuming we are reimplementing the test reporting. And it's a major reason why I'm reimplementing the test reporting :-) The fix will be possible because the new reporting will not rely on process_jam_log to get information. But instead use the BBv2 XML output directly, which has tons more accurate information about the testing results.
I have one concern about this model - from time to time my stuff depends upon some bleeding edge feature from another library or Boost tool - sometimes too development of that new feature goes hand in hand with my usage - which is to say it's developed specifically to handle problem X, and the only way to really shake down the new feature is to put it to work. For example Boost.Build's "check-target-builds" rule was developed for and tested with Boost.Regex's ICU usage requirements. Development of the Boost.Build and Regex went hand in hand. Not sure how we deal with this in the new model?
That's why I prefer 'test whole trunk, incrementally' model to the 'test each library individually, against last release' model.
I tend to prefer both. That is, I don't think we can live without full trunk testing. But we also want the partial-integration testing that using single-library-agaist-release provides. I'm perfectly fine without having the dependencies of those release-tested libraries not be available, and having them fail the pre-integration, as it would show which parts that library depends on clearly. That may shock you ;-) But I'd rather see failures that show likely integration hot-spots, than try and be ultra smart about making a fully working partially integrated release. So to summarize I'd like to see testing: 1. incremental full trunk 2. single-library against full release (incremental if tester disk space allows it) 3. incremental fully integrated release Note, "trunk" and "release" are just shorthands for the corresponding concepts in our current procedures. So adjust for possible future procedures as needed ;-) -- -- Grafik - Don't Assume Anything -- Redshift Software, Inc. - http://redshift-software.com -- rrivera/acm.org (msn) - grafik/redshift-software.com -- 102708583/icq - grafikrobot/aim,yahoo,skype,efnet,gmail