'Revision X not present': when branching from stacked branch with ghosts

spiv notes: "Although I do notice that that branch [1.13] has the AbsentContentFactory bug!"

The traceback shows that it's in the "if missing_keys" branch of fetch.py's _fetch_everything_for_search. i.e. fetch has decided that that key is either a missing compression parent or missing parent inventory, but the source does not have that key.

Using

Bazaar (bzr) 1.14rc2
  from bzr checkout /tmp/bzr-1.14
    revision: 4254

Successfully branches for me.

$ /tmp/bzr-1.14/bzr branch http://code.launchpad.net/~bzr/bzr/bzr.1.13 bzr-1.13.2
Branched 4114 revision(s).

I can reproduce this when branching standalone, but not in my shared repository. That revision is a ghost in my shared repository though.

I also cannot reproduce by branching *from* my shared repository — but that takes a different code path. This bug happens when the source is stacked, which makes InterPackRepo use the generic RepoFetcher rather than its own logic. When I make a trivially stacked branch from shared repository and branch that, then I can reproduce locally. The missing_key it finds is due to the new parent inventory check in bzr.dev, I think. Perhaps that code isn't taking ghosts into account adequately?

So, to reproduce, take your local bzr.dev and do this:

   $ bzr branch --stacked bzr.dev trivial-stacked
   $ bzr branch trivial-stacked new-standalone

If you instrument the fetch code you'll see:

missing keys: set([('inventories', '<email address hidden>'), ('inventories', '<email address hidden>')])

Those two ghosts are referenced from these two revisions: <email address hidden> and <email address hidden>

So basically ghosts are tripping up the new sanity checks in StreamSink.insert_stream. If I relax the checks to only require left-hand parent inventories then this fetch succeeds.

On Tue, 2009-04-28 at 09:18 +0000, Andrew Bennetts wrote:
> So basically ghosts are tripping up the new sanity checks in
> StreamSink.insert_stream. If I relax the checks to only require left-
> hand parent inventories then this fetch succeeds.

ghosts can occur on the lefthand side too.

Basically we need to differentiate between ghosts and nonghosts in these
cases.

The requested inventories should be right in all cases; an inventory we
can't send for which the rev is missing too is a ghost, I think.

Rob

Summary of where we are:

 * the problem is that missing parent inventories aren't necessarily a problem, so the check introduced in r4307 is too strict. Ghosts cause missing parents, but can do so harmlessly. The actual problem the code wants to catch is when we have an inventory but not all the texts it introduces. i.e. for every text version in an inventory in a possibly stacked repo, we require that either a parent inventory also references that version (so we know the current inventory didn't introduce it) or that that text version is present in this repository.
 * I have a branch with a tentative fix. It certainly copes with the ghosts in bzr.dev, but obviously I want to make sure it's robust for all cases...

https://code.launchpad.net/~spiv/bzr/all-referenced-texts-check has a fix for this, and has been submitted for review.

Merged in bzr.dev in revno 4353. There is no NEWS entry though :-/

There's no NEWS entry because, IIRC, this bug wasn't in 1.14.