On Wed, Mar 09 2022, Johannes Schindelin wrote: > Hi Junio, > > On Mon, 7 Mar 2022, Junio C Hamano wrote: > >> Johannes Schindelin <Johannes.Schindelin@xxxxxx> writes: >> >> > I said that the current output is only useful to veterans. The output that >> > hides the detailed log, under a separate job that is marked as >> > non-failing. >> > >> > That's still as true as when I said it. :-) >> >> I think getting rid of the separate "print failures" CI step and >> making it more discoverable how to reveal the details of failing >> test step is a usability improvement. > > I'm so glad you're saying that! I was starting to doubt whether my caring > about getting rid of that `print failures` step was maybe misguided. I don't think anyone's been maintaining that getting rid of it wouldn't be ideal. I for one have just been commenting on issues in the proposed implementation. I think we might still want to retain some such steps in the future, i.e. if we have a failure have subsequent steps that on failure() bump varying levels of verbosity / raw logs etc., or even try re-running the test in different ways (e.g. narrow it down with --run). But the failure step you see when something fails should ideally have the failure plus the relevant error, just as we do with compile errors. >> I am not Ævar, but I think what was questioned was the improvement >> justifies multi dozens of seconds extra waiting time, which is a >> usability dis-improvement. I do not have a good answer to that >> question. > > It is definitely worrisome that we have to pay such a price. And if there > was a good answer how to improve that (without sacrificing the > discoverability of the command trace corresponding to the test failure), I > would be more than just eager to hear it. Isn't the answer to that what I suggested in [1]; I.e. the performance problem being that we include N number of lines of the output that *didn't fail*, and that's what slows down showing the relevant output that *did* fail. I.e. if say t3070-wildmatch.sh fails in a couple of tests we'd show a *lot* of lines between the relevant failing tests, let's just elide the non-failing ones and show the output for the failing ones specifically. *Sometimes* (but very rarely) it's relevant to still look at the full output, since the failure might be due to an earlier silent failure in a previous test (or the state it left behind), but I think that's rare enough that the right thing to do is just to stick that in a subsequent "verbose dump" step or whatever. >> But new "non-veteran" users may not share that. That is something a >> bit worrying about the new UI. > > Indeed. My goal, after all, is to improve the experience of contributors, > not to make it harder. > > Still, given that you currently have to click quite a few times until you > get to where you need to be, I have my doubts that what this patch series > does is actually making things slower, measured in terms of the total time > from seeing a failed build to being able to diagnose the cause by > inspecting the command trace. Yes, but wouldn't the "Test Summary Report" in [1] be the best of both worlds[1] (with some minor changes to adapt it to the GitHub "grouping" output, perhaps)? Then you'd always see the specific of the failing test at the end, if you had N number of failures you might have a lot of output above that, but even that we could always tweak with some smart heuristic. I.e. show verbose "not ok" output if failures <10, if 10..100 elide some for the raw log, if >100 just print "this is completely screwed" or whatever :) 1. https://lore.kernel.org/git/220302.86mti87cj2.gmgdl@xxxxxxxxxxxxxxxxxxx/
