Problems with pretxnchangegroup hook

[Matt Mackall]

On Wed, 2010-06-23 at 17:47 +0200, Douglas Simon wrote:
> The documentation for 'pretxnchangegroup' at http://hgbook.red-bean.com/read/handling-repository-events-with-hooks.html#sec:hook:pretxnchangegroup includes the following:
> 
>    "...if other Mercurial processes access this repository, they will be able to see the almost-added changesets as if they are permanent."
> 
> Somewhere between Mercurial 1.1 and 1.5.4, this no longer appears to be true. I use a pretxnchangegroup hook in a gate repo to launch a master script to run a bunch of tests on the pending changes. This master script uses a number of slave/remote scripts on other machines to actually do the testing (so that testing is done on the different platforms targeted by the code). The remote scripts obtain a copy of the code by cloning the gate repo (via ssh). If the above statement holds, then the remote clones should include the pending changesets. While I was using Mercurial 1.1, this indeed was the case.
> 
> For unrelated reasons, I recently upgraded to Mercurial 1.5.4 on the gate repo machine (and all remote testing machines). After this change, the remote clones no longer include the pending changesets. To reproduce this in a terminal, create and cd to an empty directory and then cut and paste the following commands (or put them in a script and execute it): 
> 
> hg init gate
> echo -e "[hooks]\npretxnchangegroup = sleep 2" > gate/.hg/hgrc
> hg clone gate dev
> echo 'a new file' > dev/file1
> hg -R dev add dev/file1
> hg -R dev ci -m 'first file'
> hg -R dev push &
> hg -R gate log >log.1
> sleep 3
> hg -R gate log >log.2
> diff log.1 log.2
> 
> Unless I'm misinterpreting the documentation, the final diff command should report that log.1 and log.2 are identical. However, log.1 is empty where as log.2 is not. Here's the terminal transcript when I run these commands:
> 
> ~/bug$ hg init gate
> ~/bug$ echo -e "[hooks]\npretxnchangegroup = sleep 2" > gate/.hg/hgrc
> ~/bug$ hg clone gate dev
> updating to branch default
> 0 files updated, 0 files merged, 0 files removed, 0 files unresolved
> ~/bug$ echo 'a new file' > dev/file1
> ~/bug$ hg -R dev add dev/file1
> ~/bug$ hg -R dev ci -m 'first file'
> ~/bug$ hg -R dev push &
> [1] 16798
> ~/bug$ hg -R gate log >log.1
> ~/bug$ sleep 3
> pushing to /Users/dsimon/bug/gate
> searching for changes
> adding changesets
> adding manifests
> adding file changes
> added 1 changesets with 1 changes to 1 files
> [1]+  Done                    hg -R dev push
> ~/bug$ hg -R gate log >log.2
> ~/bug$ diff log.1 log.2
> 0a1,6
> > changeset:   0:4891bc5180af
> > tag:         tip
> > user:        Doug Simon <doug.simon at oracle.com>
> > date:        Wed Jun 23 17:43:42 2010 +0200
> > summary:     first file
> > 
> ~/bug$
> 
> Am I missing something or did the semantics of pretxnchangegroup change at some point? If the latter, what are the current semantics?

It's easy to find this change with:

$ hg log -k pretxn -k hook

Or equivalently:

http://www.selenic.com/hg/log?rev=pretxn+hook 
(type 'pretxn hook' in the search box)

It changed here:

changeset:   7787:b8d750daadde
user:        Matt Mackall <mpm at selenic.com>
date:        Mon Feb 16 19:35:07 2009 -0600
summary:     Introduce HG_PREPEND to solve pretxn races

According to (1.6 feature):

$ hg log -r '7787:: and tagged()'

it was first released in 1.2.

After this fix, transactions in progress are only visible to Mercurial
commands in the hook's subshell (which have the magic HG_PREPEND flag in
the environment). This means, for instance, someone can't accidentally
start pulling transactions that aren't completed and might get rolled
back in the middle of the pull.

-- 
Mathematics is the supreme nostalgia of our time.