I believe the trail3-integrated and trail3-prebuild branches address Joey's review comments from IRC:

06-12-2011 19:01:07 <joeyh>: ok, light review finished. so, if you want
to make a branch with inline trail=yes, and perhaps also adding a hook
so you don't need to inject, I think I can merge it right away

I haven't published instructions for using this version as a standalone plugin, because it needs core and inline changes.

Commits up to 63bb8b42 make the trail plugin better-integrated, including [[!inline trail=yes]]. 63bb8b42 is the commit to merge if you don't like the design of my hooks.

Commit 24168b99 adds a build_affected hook, run at about the same time as render_backlinks, and uses it to render the extra pages. This removes the need for trail to inject anything. In principle, backlinks etc. could use this hook too, if they weren't core.

Commit d0dea308 on the trail3-prebuild branch adds a prebuild hook, which runs after everything has been scanned but before anything is rendered. This removes the need for trail to run its old prerender function in its render hooks (preprocess, pagetemplate etc.) to collate metadata before it renders anything. However, I'm not sure that this is really the right thing to do, which is why it's in its own branch: the prebuild hook is a lot like needsbuild (but later), so it's called even if no trail or trail member has actually been edited.

For it to be useful for trail, the prebuild hook has to run after both pagespecs and sorting work. The other use case I've seen for a similar hook was for Giuseppe Bilotta to sort an inline-of-inlines by mtime of newest post, but that can't be the same hook, because it has to run after pagespecs work, but before sorting.

--smcv

I've merged trail3-integrated, but not prebuild. I don't exactly dislike prebuild, but dunno that the hook prolieration is worth the minor cleanup it allows in trail. --Joey

Hmm, t/trail.t is failing several tests here. To reproduce, I build the debian package from a clean state, or rm -rf .t between test runs. --Joey

t/trail.t .................... 1/? 
#   Failed test at t/trail.t line 211.
#   Failed test at t/trail.t line 213.
#   Failed test at t/trail.t line 215.
#   Failed test at t/trail.t line 217.
#   Failed test at t/trail.t line 219.
#   Failed test at t/trail.t line 221.
#   Failed test at t/trail.t line 223.
#   Failed test at t/trail.t line 225.
#   Failed test at t/trail.t line 227.
#   Failed test at t/trail.t line 229.
#   Failed test at t/trail.t line 231.

Looking at the first of these, it expected "trail=sorting n=sorting/new p=" but gets: "trail=sorting n=sorting/ancient p=sorting/new"

Looking at the second failure, it expected "trail=sorting n=sorting/middle p=sorting/old$" but got: "trail=sorting n=sorting/old p=sorting/end"

Perhaps a legitimate bug? --Joey

I saw this while developing, but couldn't reproduce it, and assumed I'd failed to update blib before make test, or some such. In fact it's a race condition, I think.

The change and failure here is that sorting.mdwn is modified to sort its trail in reverse order of title. Previously, it was sorted by order of directives in the page, and secondarily by whatever sort order each directive specified (e.g. new, old and ancient were sorted by increasing age). old appearing between new and ancient, and new appearing between end and old, indicates that this re-sorting has not actually taken effect, and the old sort order is still used.

I believe this is because the system time (as an integer) remained the same for the entire test, and mtimes as used in ikiwiki only have a 1-second resolution. We can either fix this with utime or sleep; I chose utime, since sleeping for 1 second would slow down the test significantly. Please merge or cherry-pick smcv/trail-test (there's only one commit). --smcv


Available in a git repository branch.
Branch: smcv/ready/trail
Author: smcv

Some later changes to trail:

  • Display the trail links at beginning/end of default page.tmpl as suggested on IRC
  • Improve CSS, particularly in blueview and goldtype themes (example)
  • Fix a possible bug regarding state deletion

--smcv

Applied --Joey


Trail plugin creates unexpected interdependencies?

(ikiwiki master branch 2014-06-06 also tested with 3.20140228 release)

I noticed the problem when using the album plugin but a bit of testing revealed that the trail plugin, which is used by album may be the cause of the problem.

On a site with the following structure where all albumN.mdwn files have the [[!inline pages="page(./album01/*)" trail="yes"]] directive set. All albumN pages and imgN pages get rebuilt whenever any one of the albumN or imgN pages are changed and the command ikiwiki --setup wiki.setup --refresh --verbose is issued.

/index.mdwn                        Contains no links maps or inlines
|-album01.mdwn                 [[!inline   pages="page(./album01/*)" trail="yes"]]
|-album01/
| |-imgA.mdwn
| |-imgB.mdwn
|
|-album02.mdwn                 [[!inline   pages="page(./album02/*)" trail="yes"]]
|-album02/
| |-imgC.mdwn
| |-imgD.mdwn
|
|-album03.mdwn                 [[!inline   pages="page(./album03/*)" trail="yes"]]
|-album03/
| |-imgE.mdwn
| |-imgF.mdwn

Changing the index.mdwn page also triggers a full rebuild of all pages with trail directives. My sites tend to look like the above but with double digit numbers of files in at each level. Changing any file then means a full rebuild of a rather complex site which takes a long time.

My setup and test may very well have mistakes but perhaps someone using the trail plugin could check (using the --verbose flag) if all their trails get rebuild when changing only one. I also find it curious that changes to the parent index.mdwn page triggers the same behaviour.

I have removed a similar comment from the album discussion.

--kjs

I would expect changing imgE.mdwn to rebuild album03.mdwn (because album03 inlines imgE) and vice versa (because imgE uses album03's [[!meta title]]).

I would not expect changing imgE.mdwn or album03.mdwn to affect album02 or imgC.

I would also not expect changing index.mdwn to rebuild anything else unless there is a valid dependency reason to do so.

Can you reproduce this problem in a wiki that does not contain anything private, and publish its git repo somewhere? (I realise photo galleries tend to be more personal/private than typical wikis, so you don't necessarily want to link the real thing - that's why my album demos tend to use dummy data). --smcv

I was expecting the same depends pattern you describe. My photo wikis are mostly public so I've set up a publicly accessible repo (update-server-info type, git clone the first link below), a low-res copy of the underlay and a quick sanitized setup file.

It might be a bit unwieldly and the site itself at http://stockholm.kalleswork.net uses a few tweaks to the album templates and css, but I don't currently have access to the machine where I setup a cleaner debug wiki to test. (travelling atm). The images will likely be distorted due to the up scaling bug in the img plugin but other than that it should work.

Let me know if you need anything else. Would be great to hear it works as expected for everyone else ;) --kjs

Hmm. Investigating the indexdb:

perl -le 'use Storable; my $index = Storable::retrieve("stockholm/.ikiwiki/indexdb"); use Data::Dumper; print Dumper $index' |less

indicates that 20130504 depends on internal(*) and so does 20130505.

After adding some Carp::cluck calls to the bits of IkiWiki.pm that add dependencies, this turns out to be two similar issues, in album and trail: they each use pagespec_match_list with the pagespec internal(*) in order to apply a trivial filter (accept everything) to an existing list for its side-effect of sorting that list. Bug filed as trails depend on everything --smcv