On Thu, Mar 19, 2015 at 12:14:53AM -0400, Jeff King wrote: > On Thu, Mar 19, 2015 at 11:01:17AM +0900, Mike Hommey wrote: > > > > I don't think packing the unreachables is a good plan. They just end up > > > accumulating then, and they never expire, because we keep refreshing > > > their mtime at each pack (unless you pack them once and then leave them > > > to expire, but then you end up with a large number of packs). > > > > Note, sometimes I wish unreachables were packed. Recently, I ended up in > > a situation where running gc created something like 3GB of data as per > > du, because I suddenly had something like 600K unreachable objects, each > > of them, as a loose object, taking at least 4K on disk. This made my > > .git take 5GB instead of 2GB. That surely didn't feel like garbage > > collection. > > That's definitely a thing that happens, but it is a bit of a corner > case. It's unusual to have such a large number of unreferenced objects > all at once. > > I don't suppose you happen to remember the details, but would a lower > expiration time (e.g., 1 day or 1 hour) have made all of those objects > go away? Or were they really from some extremely recent event (of > course, "event" here might just have been "I did a full repack right > before rewriting history" which would freshen the mtimes on everything > in the pack). Unfortunately, I don't know the exact details. But yes, I guess a lower expiration time might have helped. > Certainly the "loosening" behavior for unreachable objects has corner > cases like this, and they suck when you hit one. Leaving the objects > packed would be better, but IMHO is not a viable alternative unless > somebody comes up with a plan for segregating the "old" objects in a way > that they actually expire eventually, and don't just keep getting > repacked and freshened over and over. It sure is a corner case, otoh, when it happens, every single git operation calls git gc --auto, which happily spends 5 minutes sucking CPU to end up doing nothing in practice. And add more salt on the injury if you are on battery 6700 loose objects seems easy to reach on a repo with 6M objects... Mike -- To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html