ikiwiki/doc/todo/mercurial.mdwn

* Is the code sufficiently robust? It just warns when mercurial fails.
* When rcs_commit is called with a $user that is an openid, it will be
  passed through to mercurial -u. Will mercurial choke on this?
 * Nope. Mercurial doesn't expect any particular format for the username, 
   though "Name <address@domain>" is standard. --[[bma]]
* The way `-u $user` is passed to `hg commit`, there's no way to tell
  if a given commit came in over the web or was done directly. So
  rcs_recentchanges hardcodes 'committype => "mercurial"'. See the monotone
  backend for an example of one that does this right.
* The rcs_commit implementation seems not to notice if the file has been
  changed since a web edit started. Unlike all the other frontends, which
  use the rcstoken to detect if the web commit started editing an earlier
  version of the file, and if so, merge the two sets of changes together.
  It seems that with the current mercurial commit code, it will always
  blindly overwrite the current file with the web edited version, losing
  any other changes.
* `rcs_commit_staged`, `rcs_rename`, `rcs_remove`, and `rcs_diff` are not
  implemented for mercurial, and so attachments, remove and rename plugins
  and recentchangesdiff cannot be used with it. (These should be fairly
  easy to add..)

Posthook: in `$srcdir/.hg/hgrc`, I have the following

    [hooks]
    incoming.update = hg up
    update.ikiwiki = ikiwiki --setup /path/to/ikiwiki.setup --refresh

This should update the working directory and run ikiwiki every time a change is recorded (someone who knows mercurial better than I do may be able to suggest a better way, but this works for me.)

> Try running it with --post-commit instead of --refresh. That should
> work better, handling both the case where the edit was made via the web
> and then committed, and the case where a commit was made directly.
> It can deadlock if the post-commit hook runs with --refresh in the
> former case. --[[Joey]]

The problem with --post-commit is that if you delete some pages in $SRC, ikiwiki --setup setupfile --post-commit will not delete them in $DEST. --[[users/weakish]]

> You should really be using a setup file that has `mercurial_wrapper`
> set, and running the wrapper generated by that from your hook.
> That will work. I think that the `--setup --post-commit` on the command
> line is currently broken and does the same expensive rebuild process as --setup
> alone (which doesn't delete files from $DEST either). Will fix that.
> (fixed)
> --[[Joey]] 

>> Mercurial doesn't support put hooks in .hg/hooks/* (like git). In Mercurial, the only way to run
>> your own hooks is specifying them in the hgrc file. (Or write a new extension.)
>> I guess use a very long command will work.
>> (e.g. ikiwiki --post-commit --a-lot-of-switches --set var=value $SRC $DEST)
>> (Fortunately ikiwiki supports --set var=value so without --setup works.)
>>
>> Alternative is always editing via cgi or pushing.  Never work on the $SRC/repo directly.
>> --[[users/weakish]]

>>> I don't see anything preventing you from using a setup file with
>>> `mercurial_wrapper => ".hg/ikiwiki-hook",` and then modifying the hgrc
>>> to run that wrapper. --[[Joey]] 

>> Thanks for pointing out this.  I have some stupid misunderstanding on the
>> usage of mercurial_wrapper before.  The wrapper works nicely! --[[weakish]]

I add the following to .hg/hgrc:(I use changegroup since I don't think we need refresh per changeset, please point out if I am wrong.)

    [hooks]
    changegroup = hg update >&2 && ikiwiki --setup path.to.setup.file --refresh
    post-commit = path.to.the.mercurial.wrapper

-----

I have no idea when the deadlock will happen.  --[[users/weakish]]

> For the deadlock to occur, a edit has to be made via the web.
> 
> Ikiwiki,
> running as a CGI, takes a lock on the wiki, and commits the edit,
> continuing to run in the background, with the lock still held.
> When the edit is committed, the hg hook runs, running `ikwiki --refresh`.
> Nearly the first thing that process does it try to lock the wiki..
> which is already locked. This lock is taken in a blocking manner,
> thus the deadlock -- the cgi is waiting for the commit to finish before
> dropping the lock, and the commit is blocked waiting for the lock to be
> released.
> 
> --post-commit avoids this problem by checking if the cgi is running
> and avoiding doing anything in that case. (While still handing the
> refresh if the commit was not made by the CGI.)
> So in that case, the commit finishes w/o ikiwiki doing anything,
> and the ikiwiki CGI handles the wiki refresh.
> --[[Joey]]  


***

I have a few notes on mercurial usage after trying it out for a while:

1. I have been using ikiwiki's `--post-commit` option without apparent problems. I'm the only current user of my wiki, though.

1. The `ikiwiki.setup` file included in ikiwiki works with mercurial's `hgserve`, which is not the preferred solution. Mercurial's `hgwebdir.cgi` is more flexible and doesn't require running a server. I have this in my .setup file:

        # Mercurial stuff.
        rcs => "mercurial",
        historyurl => "http://localhost/cgi-bin/hgwebdir.cgi/ikiwiki/log/tip/\[[file]]",
        diffurl => "http://localhost/cgi-bin/hgwebdir.cgi/ikiwiki/diff/tip/\[[file]]",

1. I have noticed that running `ikiwiki` after a change to the wiki adds files to a directory called `recentchanges` under `$srcdir`. I don't understand why such files are needed; worse, they are not added to mercurial's list of tracked files, so they polute the output of `hg log`. Is this a bug? Should mercurial's commit hook be modified to add these files before the commit?

--buo

> No, those files should not be added to revision control. --[[Joey]]

>> OK. I see two problems:

>> 1. If I clone my wiki, I won't get an exact copy of it: I will lose the recentchanges history. This could be an acceptable limitation but IMO this should be documented.

>>> The history is stored in mercurial. How will it be lost?

>> 2. The output of `hg status` is polluted. This could be solved trivially by adding a line containing `recentchanges` to `.hgignore`. Another alternative would be to store the `recentchanges` directory inside `$srdcir/.ikiwiki`.

>> I think the ideal solution would be to build `$destdir/recentchanges/*` directly from the output of `hg log`. --[[buo]]

>>>> That would be 100 times as slow, so I chose not to do that. --[[Joey]]