<div class="gmail_quote">On Fri, Jun 19, 2009 at 6:44 PM, Andrew Bennetts <<a href="mailto:andrew.bennetts@canonical.com">andrew.bennetts@canonical.com</a>> wrote: <blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"> My wild guess would be timing sensitivity, but the complete determinism (100% reproducible in 10 attempts) suggests not. Hmm, perhaps it's a garbage collection issue? Turning on -v will have the effect of producing slightly different object allocation/deallocation counts, which will cause the periodic cyclic garbage collector to kick in at slightly different points... it's not the first time I've seen an issue like that... I don't have time to do this right now, but if I were investigating this I'd be experimenting with adding "import gc; gc.disable()" early in the 'bzr' script and see how that affects results. </blockquote><div> Thanks for the suggestion. I was pretty excited about this, because I vaguely recall seeing a warning about garbage collection at some point in the last couple days. So I made a copy bzr with import gc; gc.disable() right at the top, ahead of import os. I regret to report that I found the same curious behaviors with bzr-no_gc (my modified copy) as I originally reported for 'selftest http_test'. So I did a grep "import gc" *.py in /usr/lib/python2.6/dist-packages/bzrlib to see if anything else is trivially pulling in gc and trumping gc.disable. No matches. I suppose there might be indirect ways to get import gc, but I have not looked any harder. A less specific grep for "gc." only turns up references to groupcompress, which is obviously unrelated to this discussion. Anyway, I've only tested test_http without garbage collection. I am going to go through my other unusual results and see if garbage collection makes any difference. Thanks! -M </div></div>