On Thu, Oct 15, 2020 at 01:08:22PM -0700, Junio C Hamano wrote: > > The sub-thread between me and Jonathan starting here: > > > > https://lore.kernel.org/git/20201005082448.GB2862927@xxxxxxxxxxxxxxxxxxxxxxx/ > > > > but specifically: > > > > https://lore.kernel.org/git/20201005084946.GE2862927@xxxxxxxxxxxxxxxxxxxxxxx/ > > Ah, I see. I actually do use "git ls-files t/ | grep ..." to look > for tests that are relevant to the issue I have at hand quite often, > so unlike what Jonathan said in the thread, having a good name does > matter to me. I'm not sure where he suggests worse names. I'd think if anything we'd have better names, because they'd be even more meaningful (if people start using them for test selectors). FWIW, I also grep like that when looking for scripts. > As far as I can tell, the numbers in the test names serve only two > purposes. > > One is as a unique key to avoid colliding in the test result > aggrevation database, and the other is as a unique key to use in > GIT_SKIP_TESTS (which in turn is used by the Meta/Make wrapper I > use, found on the todo branch, like 'Meta/Make --test=0050,1400 > test'). Yeah, that sounds right. We also frequently talk about "t5601" in human-readable messages. I often use that to either tab-complete or run "./t5601-*". So they do serve as short unique identifiers there. You could tell me "t-foo-bar-baz" and that would similarly be unique, but it's not as short. :) They do serve as unique keys in places like test-results/, but really the whole test name is the key. Most of them have unique text descriptions. If I do: git ls-files t/ | perl -lne '/t[0-9]{4}-(.*).sh$/ and print "$1"' | sort | uniq -d that only yields two duplicates (and we'd probably be better off giving them unique names anyway). > I would be heavily inconvenienced if we decide to remove numbers > becuase it would rob the latter use case from me, but I'd survive if > we just are going to lift the requirement that numbers must be > unique. My plan is that you'd be able to say: Meta/make --test=filesystem,update-ref for t0050 and t1400 respectively. Or more interesting things like "git-svn-*" to skip all of the t91xx tests (or "git-svn/*" if we split them by directories). > I may end up running irrelevant 0050 and 1400 when the > tests I really want to run are the other 0050 and 1400 with > "--test=0050,1400", but when I am trying to run only 2 among 900+ > scripts, running 2 extra ones I didn't have to run only because > their prefix collide is still much better and tolerable. I think you could already do this in general. The only thing standing in the way is test-lint-duplicates. I added that because I thought it would be better to catch the duplicates early rather than get confused by them later when you can't uniquely identify a test. But in practice I do feel like I spend more time de-duping tests when I rebase or merge (because my branch raced with somebody else allocating the test number) than I ever did dealing with unexpectedly skipped tests. I definitely don't want to make anybody's life harder. But if the numbers aren't unique and the text connected to them is, then it seems like we should just use the text primarily. -Peff
