ikiwiki/doc/todo/smarter_sorting.mdwn

I benchmarked a build of a large wiki (my home wiki), and it was spending
quite a lot of time sorting; `CORE::sort` was called only 1138 times, but
still flagged as the #1 time sink. (I'm not sure I trust NYTProf fully
about that FWIW, since it also said 27238263 calls to `cmp_age` were
the #3 timesink, and I suspect it may not entirely accurately measure
the overhead of so many short function calls.)

`pagespec_match_list` currently always sorts *all* pages first, and then
finds the top M that match the pagespec. That's innefficient when M is
small (as for example in a typical blog, where only 20 posts are shown,
out of maybe thousands).

As [[smcv]] noted, It could be flipped, so the pagespec is applied first,
and then sort the smaller matching set. But, checking pagespecs is likely
more expensive than sorting. (Also, influence calculation complicates
doing that.)

Another option, when there is a limit on M pages to return, might be to
cull the M top pages without sorting the rest.

> The patch below implements this.
> 
> But, I have not thought enough about influence calculation.
> I need to figure out which pagespec matches influences need to be
> accumulated for in order to determine all possible influences of a
> pagespec are known.
> 
> The old code accumulates influences from matching all successful pages
> up to the num cutoff, as well as influences from an arbitrary (sometimes
> zero) number of failed matches. New code does not accumulate influences
> from all the top successful matches, only an arbitrary group of
> successes and some failures.
> 
> Also, by the time I finished this, it was not measuarably faster than 
> the old method. At least not with a few thousand pages; it
> might be worth revisiting this sometime for many more pages? [[done]]
> --[[Joey]] 

<pre>
diff --git a/IkiWiki.pm b/IkiWiki.pm
index 1730e47..bc8b23d 100644
--- a/IkiWiki.pm
+++ b/IkiWiki.pm
@@ -2122,36 +2122,54 @@ sub pagespec_match_list ($$;@) {
 	my $num=$params{num};
 	delete @params{qw{num deptype reverse sort filter list}};
 	
-	# when only the top matches will be returned, it's efficient to
-	# sort before matching to pagespec,
-	if (defined $num && defined $sort) {
-		@candidates=IkiWiki::SortSpec::sort_pages(
-			$sort, @candidates);
-	}
-	
+	# Find the first num matches (or all), before sorting.
 	my @matches;
-	my $firstfail;
 	my $count=0;
 	my $accum=IkiWiki::SuccessReason->new();
-	foreach my $p (@candidates) {
-		my $r=$sub->($p, %params, location => $page);
+	my $i;
+	for ($i=0; $i < @candidates; $i++) {
+		my $r=$sub->($candidates[$i], %params, location => $page);
 		error(sprintf(gettext("cannot match pages: %s"), $r))
 			if $r->isa("IkiWiki::ErrorReason");
 		$accum |= $r;
 		if ($r) {
-			push @matches, $p;
+			push @matches, $candidates[$i];
 			last if defined $num && ++$count == $num;
 		}
 	}
 
+	# We have num natches, but they may not be the best.
+	# Efficiently find and add the rest, without sorting the full list of
+	# candidates.
+	if (defined $num && defined $sort) {
+		@matches=IkiWiki::SortSpec::sort_pages($sort, @matches);
+
+		for ($i++; $i < @candidates; $i++) {
+			# Comparing candidate with lowest match is cheaper,
+			# so it's done before testing against pagespec.
+			if (IkiWiki::SortSpec::cmptwo($candidates[$i], $matches[-1], $sort) < 0 &&
+			    $sub->($candidates[$i], %params, location => $page)
+			) {
+				# this could be done less expensively
+				# using a binary search
+				for (my $j=0; $j < @matches; $j++) {
+					if (IkiWiki::SortSpec::cmptwo($candidates[$i], $matches[$j], $sort) < 0) {
+						splice @matches, $j, $#matches-$j+1, $candidates[$i],
+							@matches[$j..$#matches-1];
+						last;
+					}
+				}
+			}
+		}
+	}
+
 	# Add simple dependencies for accumulated influences.
-	my $i=$accum->influences;
-	foreach my $k (keys %$i) {
-		$depends_simple{$page}{lc $k} |= $i->{$k};
+	my $inf=$accum->influences;
+	foreach my $k (keys %$inf) {
+		$depends_simple{$page}{lc $k} |= $inf->{$k};
 	}
 
-	# when all matches will be returned, it's efficient to
-	# sort after matching
+	# Sort if we didn't already.
 	if (! defined $num && defined $sort) {
 		return IkiWiki::SortSpec::sort_pages(
 			$sort, @matches);
@@ -2455,6 +2473,12 @@ sub sort_pages {
 	sort $f @_
 }
 
+sub cmptwo {
+	$a=$_[0];
+	$b=$_[1];
+	$_[2]->();
+}
+
 sub cmp_title {
 	IkiWiki::pagetitle(IkiWiki::basename($a))
 	cmp
</pre>

This would be bad when M is very large, and particularly, of course, when
there is no limit and all pages are being matched on. (For example, an
archive page shows all pages that match a pagespec specifying a creation
date range.) Well, in this case, it *does* make sense to flip it, limit by
pagespe first, and do a (quick)sort second. (No influence complications,
either.)

> Flipping when there's no limit implemented, and it knocked 1/3 off
> the rebuild time of my blog's archive pages. --[[Joey]] 

Adding these special cases will be more complicated, but I think the best
of both worlds. --[[Joey]]