On Mon, Apr 24, 2017 at 6:34 PM, Philip Oakley <philipoakley@xxxxxxx> wrote: > > Sorry if I'm bikeshedding here, but it does look like adding an alternate > 'bisect' strategy may be useful for this style of integration testing. > > To me, given the multiplicity of independent branches being brought > together, it should be possible to do a check on each of the branches > separately, before the tests along the line of integration . The tests would > not be a true 'bisection' as they are separate single point tests, but they > do establish good commits at the tips of those branches. > > Thus, for each of the merges in the --first-parent traversal, the option > could test (in the OS of choice) the *second parent* commit of the merge. > This sets the known good points. The breakages during the integration then > become easier to bisect, as it is only looking for the integration of a good > series into mainline that breaks. [1] I think what you suggest is very similar in practice as the --first-parent strategy that we have been suggesting for some years to add to bisect as a GSoC project: https://git.github.io/SoC-2017-Ideas/ > In addition, an aside that dscho made earlier about the merge-base of some > branches relative to Windows may have been missed. The normal bisect process > assumes that we start from a set of good merge bases. When a good commit is not an ancestor of the bad commit, we test the merge bases first and we abort if one of them is not good. > However, dscho noticed > that occasionally some series may choose an older point on maint (etc.) that > itself would _not_ be classed as good when tested on Windows (or on other > OS's). Those older mergebases can make the classic bisect across all the > commits in the DAG between here and there a tortuous process, especially if > the local OS implementation percieves the base points as bad! (which breaks > expectations). My understanding of the bisecting problem that Dscho is reporting is that the process is sometimes buggy or tortuous because bisect is not smart enough to optimize the testing of the merge bases when there are a lot of them. It is not because bisect expects or assumes that some merge bases are good or bad when in fact they are not.