On 10/13/2016 05:39 PM, Pierre-Yves David wrote:
On 10/13/2016 05:34 PM, Yuya Nishihara wrote:
On Thu, 13 Oct 2016 17:28:02 +0200, Mads Kiilerich wrote:
On 10/13/2016 05:07 PM, Pierre-Yves David wrote:
On 10/13/2016 03:21 PM, Mads Kiilerich wrote:
On 10/13/2016 01:53 PM, Pierre-Yves David wrote:
# HG changeset patch
# User Pierre-Yves David <pierre-yves.da...@ens-lyon.org>
# Date 1476359131 -7200
# Thu Oct 13 13:45:31 2016 +0200
# Node ID 88cc944830d0c1895e527d6ca13687f1d5e1c785
# Parent 747e546c561fbf34d07cd30013eaf42b0190bb3b
eol: do not wait on lack when writing cache
The cache writing process is properly catching and handling the case
lock is unavailable. However, it fails to specify the lock can failed
acquired when requesting it. This is now fixed.
*If* the user has write access to the repo and *could* lock the repo,
then it seems reasonable that it waits for the lock and does the right
thing. It would be unfortunate to bail out early and happily
expose the less optimal state that read only users might have to deal
The change introduced by this changeset make the cache in line with
how most of other cache works in Mercurial.
A part of the problem here might be that it is unclear to me what
happens with wait=False. I don't remember the details: will it continue
without locking or will it raise? It would be nice to get that clarified
in the localrepo docstrings.
LockHeld would be raised. So we'll need to catch LockError if wait=False.
And we actually only catch LockUnavailable, lets drop patch 1 for now
(patch 2 is still valid)
So looking more into this it appears that:
- This function is always run for every reposetup(), (irk)
- If the cache seems up to date the function end early, (okay)
- If the content within the locking scope is not run, the repository
have an inconsistent state and can misbehave (irk)
My current strategy is to slowly back away from this code and leave it
in its current state.
Mercurial-devel mailing list