path: root/doc/todo/allow_plugins_to_add_sorting_methods.mdwn

[[!tag patch]]

The available [[ikiwiki/pagespec/sorting]] methods are currently hard-coded in
IkiWiki.pm, making it difficult to add any extra sorting mechanisms. I've
prepared a branch which adds 'sort' as a hook type and uses it to implement a
new `meta_title` sort type.

Someone could use this hook to make `\[[!inline sort=title]]` prefer the meta
title over the page name, but for compatibility, I'm not going to (I do wonder
whether it would be worth making sort=name an alias for the current sort=title,
and changing the meaning of sort=title in 4.0, though).

> What compatability concerns, exactly, are there that prevent making that
> change now? --[[Joey]] 

*[sort-hooks branch now withdrawn in favour of sort-package --s]*

I briefly tried to turn *all* the current sort types into hook functions, and
have some of them pre-registered, but decided that probably wasn't a good idea.
That earlier version of the branch is also available for comparison:

*[also withdrawn in favour of sort-package --s]*

>> I wonder if IkiWiki would benefit from the concept of a "sortspec", like a [[ikiwiki/PageSpec]] but dedicated to sorting lists of pages rather than defining lists of pages?  Rather than defining a sort-hook, define a SortSpec class, and enable people to add their own sort methods as functions defined inside that class, similarly to the way they can add their own pagespec definitions. --[[KathrynAndersen]]

>>> [[!template id=gitbranch branch=smcv/ready/sort-package author="[[Simon_McVittie|smcv]]"]]
>>> I'd be inclined to think that's overkill, but it wasn't very hard to
>>> implement, and in a way is more elegant. I set it up so sort mechanisms
>>> share the `IkiWiki::PageSpec` package, but with a `cmp_` prefix. Gitweb:
>>> <http://git.pseudorandom.co.uk/smcv/ikiwiki.git?a=shortlog;h=refs/heads/sort-package>

>>>> I agree it seems more elegant, so I have focused on it.
>>>>
>>>> I don't know about reusing `IkiWiki::PageSpec` for this.
>>>> --[[Joey]]

>>>>> Fair enough, `IkiWiki::SortSpec::cmp_foo` would be just
>>>>> as easy, or `IkiWiki::Sorting::cmp_foo` if you don't like
>>>>> introducing "sort spec" in the API. I took a cue from
>>>>> [[ikiwiki/pagespec/sorting]] being a subpage of
>>>>> [[ikiwiki/pagespec]], and decided that yes, sorting is
>>>>> a bit like a pagespec :-) Which name would you prefer? --s

>>>>>> `SortSpec` --[[Joey]] 

>>>>>>> [[Done]]. --s

>>>> I would be inclined to drop the `check_` stuff. --[[Joey]] 

>>>>> It basically exists to support `title_natural`, to avoid
>>>>> firing up the whole import mechanism on every `cmp`
>>>>> (although I suppose that could just be a call to a
>>>>> memoized helper function). It also lets sort specs that
>>>>> *must* have a parameter, like
>>>>> [[field|plugins/contrib/field/discussion]], fail early
>>>>> (again, not so valuable).
>>>>>
>>>>>> AFAIK, `use foo` has very low overhead when the module is already
>>>>>> loaded. There could be some evalation overhead in `eval q{use foo}`,
>>>>>> if so it would be worth addressing across the whole codebase.
>>>>>> --[[Joey]] 
>>>>>>
>>>>>>> check_cmp_foo now dropped. --s
>>>>>
>>>>> The former function could be achieved at a small
>>>>> compatibility cost by putting `title_natural` in a new
>>>>> `sortnatural` plugin (that fails to load if you don't
>>>>> have `title_natural`), if you'd prefer - that's what would
>>>>> have happened if `title_natural` was written after this
>>>>> code had been merged, I suspect. Would you prefer this? --s

>>>>>> Yes! (Assuming it does not make sense to support
>>>>>> natural order sort of other keys than the title, at least..)
>>>>>>  --[[Joey]]

>>>>>>> Done. I added some NEWS.Debian for it, too. --s

>>>> Wouldn't it make sense to have `meta(title)` instead
>>>> of `meta_title`? --[[Joey]]

>>>>> Yes, you're right. I added parameters to support `field`,
>>>>> and didn't think about making `meta` use them too.
>>>>> However, `title` does need a special case to make it
>>>>> default to the basename instead of the empty string.
>>>>>
>>>>> Another special case for `title` is to use `titlesort`
>>>>> first (the name `titlesort` is derived from Ogg/FLAC
>>>>> tags, which can have `titlesort` and `artistsort`).
>>>>> I could easily extend that to other metas, though;
>>>>> in fact, for e.g. book lists it would be nice for
>>>>> `field(bookauthor)` to behave similarly, so you can
>>>>> display "Douglas Adams" but sort by "Adams, Douglas".
>>>>>
>>>>> `meta_title` is also meant to be a prototype of how
>>>>> `sort=title` could behave in 4.0 or something - sorting
>>>>> by page name (which usually sorts in approximately the
>>>>> same place as the meta-title, but occasionally not), while
>>>>> displaying meta-titles, does look quite odd. --s

>>>>>> Agreed. --[[Joey]]

>>>>>>> I've implemented meta(title). meta(author) also has the
>>>>>>> `sortas` special case; meta(updated) and meta(date)
>>>>>>> should also work how you'd expect them to (but they're
>>>>>>> earliest-first, unlike age). --s

>>>> As I read the regexp in `cmpspec_translate`, the "command"
>>>> is required to have params. They should be optional, 
>>>> to match the documentation and because most sort methods
>>>> do not need parameters. --[[Joey]]

>>>>> No, `$2` is either `\w+\([^\)]*\)` or `[^\s]+` (with the
>>>>> latter causing an error later if it doesn't also match `\w+`).
>>>>> This branch doesn't add any parameterized sort methods,
>>>>> in fact, although I did provide one on
>>>>> [[field's_discussion_page|plugins/contrib/report/discussion]]. --s

>>>> I wonder if it would make sense to add some combining keywords, so
>>>> a sortspec reads like `sort="age then ascending title"`
>>>> In a way, this reduces the amount of syntax that needs to be learned.
>>>> I like the "then" (and it could allow other operations than
>>>> simple combination, if any others make sense). Not so sure about the
>>>> "ascending", which could be "reverse" instead, but "descending age" and
>>>> "ascending age" both seem useful to be able to explicitly specify.
>>>> --[[Joey]]

>>>>> Perhaps. I do like the simplicity of [[KathrynAndersen]]'s syntax
>>>>> from [[plugins/contrib/report]] (which I copied verbatim, except for
>>>>> turning sort-by-`field` into a parameterized spec).
>>>>>
>>>>> If we're getting into English-like (or at least SQL-like) queries,
>>>>> it might make sense to change the signature of the hook function
>>>>> so it's a function to return a key, e.g.
>>>>> `sub key_age { return -%pagemtime{$_[0]) }`. Then we could sort like
>>>>> this:
>>>>>
>>>>>     field(artistsort) or field(artist) or constant(Various Artists) then meta(titlesort) or meta(title) or title
>>>>>
>>>>> with "or" binding more closely than "then". Does this seem valuable?
>>>>> I think the implementation would be somewhat more difficult. and
>>>>> it's probably getting too complicated to be worthwhile, though?
>>>>> (The keys that actually benefit from this could just
>>>>> have smarter cmp functions, I think.)
>>>>>
>>>>> If the hooks return keys rather than cmp results, then we could even
>>>>> have "lowercase" as an adjective used like "ascending"... maybe.
>>>>> However, there are two types of adjective here: "lowercase"
>>>>> really applies to the keys, whereas "ascending" applies to the "cmp"
>>>>> result. Again, I think this is getting too complex, and could just
>>>>> be solved with smarter cmp functions.
>>>>>
>>>>>> I agree. (Also, I think returning keys may make it harder to write
>>>>>> smarter cmp functions.) --[[Joey]] 
>>>>>
>>>>> Unfortunately, `sort="ascending mtime"` actually sorts by *descending*
>>>>> timestamp (but`sort=age` is fine, because `age` could be defined as
>>>>> now minus `ctime`). `sort=freshness` isn't right either, because
>>>>> "sort by freshness" seems as though it ought to mean freshest first,
>>>>> but "sort by ascending freshness" means put the least fresh first. If
>>>>> we have ascending and descending keywords which are optional, I don't
>>>>> think we really want different sort types to have different default
>>>>> directions - it seems clearer to have `ascending` always be a no-op,
>>>>> and `descending` always negate.
>>>>>
>>>>>> I think you've convinced me that ascending/descending impose too
>>>>>> much semantics on it, so "-" is better. --[[Joey]]

>>>>>>> I've kept the semantics from `report` as-is, then:
>>>>>>> e.g. `sort="age -title"`. --s

>>>>> Perhaps we could borrow from `meta updated` and use `update_age`?
>>>>> `updateage` would perhaps be a more normal IkiWiki style - but that
>>>>> makes me think that updateage is a quantity analagous to tonnage or
>>>>> voltage, with more or less recently updated pages being said to have
>>>>> more or less updateage. I don't know whether that's good or bad :-)
>>>>>
>>>>> I'm sure there's a much better word, but I can't see it. Do you have
>>>>> a better idea? --s

[Regarding the `meta title=foo sort=bar` special case]

> I feel it sould be clearer to call that "sortas", since "sort=" is used
> to specify a sort method in other directives. --[[Joey]]
>> Done. --[[smcv]]

## speed

I notice the implementation does not use the magic `$a` and `$b` globals.
That nasty perl optimisation is still worthwhile:

	perl -e 'use warnings; use strict; use Benchmark; sub a { $a <=> $b } sub b ($$) { $_[0] <=> $_[1] }; my @list=reverse(1..9999); timethese(10000, {a => sub {my @f=sort a @list}, b => sub {my @f=sort b  @list}, c => => sub {my @f=sort { b($a,$b) } @list}})'
	Benchmark: timing 10000 iterations of a, b, c...
	         a: 80 wallclock secs (76.74 usr +  0.05 sys = 76.79 CPU) @ 130.23/s (n=10000)
	         b: 112 wallclock secs (106.14 usr +  0.20 sys = 106.34 CPU) @ 94.04/s (n=10000)
	         c: 330 wallclock secs (320.25 usr +  0.17 sys = 320.42 CPU) @ 31.21/s (n=10000)

Unfortunatly, I think that c is closest to the new implementation.
--[[Joey]]

> Unfortunately, `$a` isn't always `$main::a` - it's `$Package::a` where
> `Package` is the call site of the sort call. This was a showstopper when
> `sort` was a hook implemented in many packages, but now that it's a
> `SortSpec`, I may be able to fix this by putting a `sort` wrapper in the
> `SortSpec` namespace, so it's like this:
>
>     sub sort ($@)
>     {
>         my $cmp = shift;
>         return sort $cmp @_;
>     }
>
> which would mean that the comparison used `$IkiWiki::SortSpec::a`.
> --s

>> I've now done this. On a wiki with many [[plugins/contrib/album]]s
>> (a full rebuild takes half an hour!), I tested a refresh after
>> `touch tags/*.mdwn` (my tag pages contain inlines of the form
>> `tagged(foo)` sorted by date, so they exercise sorting).
>> I also tried removing sorting from `pagespec_match_list`
>> altogether, as an upper bound for how fast we can possibly make it.
>>
>> * `master` at branch point: 63.72user 0.29system
>> * `master` at branch point: 63.91user 0.37system
>> * my branch, with `@_`: 65.28user 0.29system
>> * my branch, with `@_`: 65.21user 0.28system
>> * my branch, with `$a`: 64.09user 0.28system
>> * my branch, with `$a`: 63.83user 0.36system
>> * not sorted at all: 58.99user 0.29system
>> * not sorted at all: 58.92user 0.29system
>>
>> --s

> I do notice that `pagespec_match_list` performs the sort before the
> filter by pagespec. Is this a deliberate design choice, or
> coincidence? I can see that when `limit` is used, this could be
> used to only run the pagespec match function until `limit` pages
> have been selected, but the cost is that every page in the wiki
> is sorted. Or, it might be useful to do the filtering first, then
> sort the sub-list thus produced, then finally apply the limit? --s

>> Yes, it was deliberate, pagespec matching can be expensive enough that
>> needing to sort a lot of pages seems likely to be less work. (I don't
>> remember what benchmarking was done though.) --[[Joey]]

>>> We discussed this on IRC and Joey pointed out that this also affects
>>> dependency calculation, so I'm not going to get into this now... --s

Joey pointed out on IRC that the `titlesort` feature duplicates all the
meta titles. I did that in order to sort by the unescaped version, but
I've now changed the branch to only store that if it makes a difference.
--s

## Documentation from sort-package branch

### advanced sort orders (conditionally added to [[ikiwiki/pagespec/sorting]])

* `title_natural` - Orders by title, but numbers in the title are treated
  as such, ("1 2 9 10 20" instead of "1 10 2 20 9")
* `meta(title)` - Order according to the `\[[!meta title="foo" sortas="bar"]]`
  or `\[[!meta title="foo"]]` [[ikiwiki/directive]], or the page name if no
  full title was set. `meta(author)`, `meta(date)`, `meta(updated)`, etc.
  also work.

### Multiple sort orders (added to [[ikiwiki/pagespec/sorting]])

In addition, you can combine several sort orders and/or reverse the order of
sorting, with a string like `age -title` (which would sort by age, then by
title in reverse order if two pages have the same age).

### meta sortas parameter (added to [[ikiwiki/directive/meta]])

[in title]

An optional `sort` parameter will be used preferentially when
[[ikiwiki/pagespec/sorting]] by `meta(title)`:

       \[[!meta title="The Beatles" sort="Beatles, The"]]

       \[[!meta title="David Bowie" sort="Bowie, David"]]

[in author]

  An optional `sortas` parameter will be used preferentially when
  [[ikiwiki/pagespec/sorting]] by `meta(author)`:

        \[[!meta author="Joey Hess" sortas="Hess, Joey"]]

### Sorting plugins (added to [[plugins/write]])

Similarly, it's possible to write plugins that add new functions as
[[ikiwiki/pagespec/sorting]] methods. To achieve this, add a function to
the IkiWiki::SortSpec package named `cmp_foo`, which will be used when sorting
by `foo` or `foo(...)` is requested.

The names of pages to be compared are in the global variables `$a` and `$b`
in the IkiWiki::SortSpec package. The function should return the same thing
as Perl's `cmp` and `<=>` operators: negative if `$a` is less than `$b`,
positive if `$a` is greater, or zero if they are considered equal. It may
also raise an error using `error`, for instance if it needs a parameter but
one isn't provided.

The function will also be passed one or more parameters. The first is
`undef` if invoked as `foo`, or the parameter `"bar"` if invoked as `foo(bar)`;
it may also be passed additional, named parameters.