Re: Error converting from 1.4.4.1 to 1.5.0?

On Wed, 14 Feb 2007, Bill Lear wrote:

This in itself could have been due to a historical buglet that shouldn't matter (SHA1 on pack-files got miscomputed). However, that's probably NOT the problem, since:

implies that the pack really is corrupt.

Even a single-bit error will corrupt a pack in bad ways, which is one reason why we're so careful with it and add its own SHA1 to the end.

The best way to proceed:

 - MAKE A BACKUP ("tar" up everything). If for no other reason than

   (a) then you don't have to worry about making things worse even by 
       mistake
   (b) it might be interesting for others (if you can make those 
       pack-files available) to try to figure out what exactly the 
       corruption was. We've done it before, when it turned out to be a 
       single-bit error.

 - if you have other git archives or just back-ups of everything, just use 
   them, and throw the corrupt one away entirely (but see above on why 
   it's nice to have an archive of the corruption for posterity anyway)

 - if you don't, you can try "git unpack-objects -r". See the man-page on 
   why you need to first _move_ the pack-file away:

	mv <bad-pack-file> .git/bad-pack.pack
	mv <bad-pack-index> .git/bad-index.index

	git unpack-objects -r < .git/bad-pack.pack

   this will unpack as many objects into loose format as it can. Hopefully 
   you haven't lost much.

 - after that, the ones that you *did* lose, you can hopefully find in 
   older git repos: even if you didn't have the *full* new repo anywhere 
   else, other git repositories may have the particular objects that got 
   corrupted. "git fsck" will tell you what is missing, and you can just 
   point your .git/info/alternates file at other repositories to "steal" 
   objects from automatically.

 - if you aren't missing any objects after that, you can now repack the 
   repository, and then remove the alternates file:

	git repack -a -d
	rm .git/info/alternates

   because the repack will steal all the objects you need, and thus you 
   don't need alternates any more.

Finally: it would be very interesting to hear if you do something strange 
or unusual that could have made your chances of getting corruption higher.

Have you ever seen random SIGSEV's or strange oopses, which could be a sign of memory corruption on your machine? Do you do a lot of things over NFS? (which really can corrupt things, especially in circumstances with dodgy ethernet chips: the UDP checksums are very weak, and some ethernet cards do not do a good job of checking the ethernet CRC's!).

Yes, don't do anything more (certainly do *not* repack or anything) until you have tarred up and saved the current state, and then only _after_ you have a good safe archive to restart from, try to fix it up.

And if you can make the git history available to outsiders, I'd love to see the corrupt tar-file (it doesn't have to be *public*, if you just can trust me and perhaps a few other people with the data).

So far, as far as I can recall, we've certainly had people screw up their own trees by mistake, but apart from that kind of "user error" things, the only real corruption I recall was the single-bit error in a pack-file. We were able to recover that, but in general, for safety, the best way to protect your data is to replicate it across multiple independent machines (something that git is _good_ at, happily).

		Linus