I fixed two reversions yesterday (neither related to v7 repos) during a day of triage in preparation for the release of git-annex 7.

One of the reversions broke adding remotes in the webapp, and was filed all the way back in January with lots of confirmations. I feel bad I didn't get around to even looking at that bug report until now.

My backlog is kind of large, it hovers around 400 messages most of the time now, there needs to be a better way to make sure I notice such bad bugs. Would someone like to help with git-annex bug triage, picking out bugs that multiple users have confirmed, or that have good intructions to reproduce them, and helping me prioritize them? No coding required, massive contribution to git-annex. Please get in touch.


Anyway, after that full day's work, I took a look at the autobuilders, and it was bad; the test suite was failing everywhere testing v7. For quite a while I've been seeing intermittent test suite failures involving the new repo version, that mostly only happened on the autobuilders. But now they were more reproducible; a recent change made them happen much more frequently. That was good; it made it easier to track down the problem.

Which was that git-annex was getting mtime information with 1 second granularity. So when the test suite modified a file several times in the same second, git-annex could fail to notice some of the modifications. I think when I origianlly developed the inode cache module in 2013, for direct mode, there was no easy way to access high-precision mtimes from haskell, but there is now, and git-annex will use them.

That left one other failure in the test suite, an intermittent crash of sqlite with ErrorIO on Linux. May be related to the known sqlite crashes in WSL. I've been trying various things today to try to fix it, but have to run the test suite in a loop for several hours to reproduce it reliably.