Use a Schwartzian transform with custom sorting

benchmark/schwartzian_transform.rb

@@ -0,0 +1,115 @@
+#!/usr/bin/env ruby
+# frozen_string_literal: true
+#
+# The Ruby documentation for #sort_by describes what's called a Schwartzian transform:
+#
+#  > A more efficient technique is to cache the sort keys (modification times in this case)
+#  > before the sort. Perl users often call this approach a Schwartzian transform, after
+#  > Randal Schwartz. We construct a temporary array, where each element is an array
+#  > containing our sort key along with the filename. We sort this array, and then extract
+#  > the filename from the result.
+#  > This is exactly what sort_by does internally.
+#
+# The well-documented efficiency of sort_by is a good reason to use it. However, when a property
+# does not exist on an item being sorted, it can cause issues (no nil's allowed!)
+# In Jekyll::Filters#sort_input, we extract the property in each iteration of #sort,
+# which is quite inefficient! How inefficient? This benchmark will tell you just how, and how much
+# it can be improved by using the Schwartzian transform. Thanks, Randall!
+
+require 'benchmark/ips'
+require 'minitest'
+require File.expand_path("../lib/jekyll", __dir__)
+
+def site
+  @site ||= Jekyll::Site.new(
+    Jekyll.configuration("source" => File.expand_path("../docs", __dir__))
+  ).tap(&:reset).tap(&:read)
+end
+
+def site_docs
+  site.collections["docs"].docs.dup
+end
+
+def sort_by_property_directly(docs, meta_key)
+  docs.sort! do |apple, orange|
+    apple_property = apple[meta_key]
+    orange_property = orange[meta_key]
+
+    if !apple_property.nil? && !orange_property.nil?
+      apple_property <=> orange_property
+    elsif !apple_property.nil? && orange_property.nil?
+      -1
+    elsif apple_property.nil? && !orange_property.nil?
+      1
+    else
+      apple <=> orange
+    end
+  end
+end
+
+def schwartzian_transform(docs, meta_key)
+  docs.collect! { |d|
+    [d[meta_key], d]
+  }.sort! { |apple, orange|
+    if !apple[0].nil? && !orange[0].nil?
+      apple.first <=> orange.first
+    elsif !apple[0].nil? && orange[0].nil?
+      -1
+    elsif apple[0].nil? && !orange[0].nil?
+      1
+    else
+      apple[-1] <=> orange[-1]
+    end
+  }.collect! { |d| d[-1] }
+end
+
+# Before we test efficiency, do they produce the same output?
+class Correctness
+  include Minitest::Assertions
+
+  require "pp"
+  define_method :mu_pp, &:pretty_inspect
+
+  attr_accessor :assertions
+
+  def initialize(docs, property)
+    @assertions = 0
+    @docs = docs
+    @property = property
+  end
+
+  def assert!
+    assert sort_by_property_directly(@docs, @property).is_a?(Array), "sort_by_property_directly must return an array"
+    assert schwartzian_transform(@docs, @property).is_a?(Array), "schwartzian_transform must return an array"
+    assert_equal sort_by_property_directly(@docs, @property),
+      schwartzian_transform(@docs, @property)
+    puts "Yeah, ok, correctness all checks out for property #{@property.inspect}"
+  end
+end
+
+Correctness.new(site_docs, "redirect_from".freeze).assert!
+Correctness.new(site_docs, "title".freeze).assert!
+
+# First, test with a property only a handful of documents have.
+Benchmark.ips do |x|
+  x.config(time: 10, warmup: 5)
+  x.report('sort_by_property_directly with sparse property') do
+    sort_by_property_directly(site_docs, "redirect_from".freeze)
+  end
+  x.report('schwartzian_transform with sparse property') do
+    schwartzian_transform(site_docs, "redirect_from".freeze)
+  end
+  x.compare!
+end
+
+# Next, test with a property they all have.
+Benchmark.ips do |x|
+  x.config(time: 10, warmup: 5)
+  x.report('sort_by_property_directly with non-sparse property') do
+    sort_by_property_directly(site_docs, "title".freeze)
+  end
+  x.report('schwartzian_transform with non-sparse property') do
+    schwartzian_transform(site_docs, "title".freeze)
+  end
+  x.compare!
+end

lib/jekyll/filters.rb

@@ -335,19 +335,26 @@ def inspect(input)
     end
 
     private
+
+    # Sort the input Enumerable by the given property.
+    # If the property doesn't exist, return the sort order respective of
+    # which item doesn't have the property.
+    # We also utilize the Schwartzian transform to make this more efficient.
     def sort_input(input, property, order)
-      input.sort do |apple, orange|
-        apple_property = item_property(apple, property)
-        orange_property = item_property(orange, property)
-
-        if !apple_property.nil? && orange_property.nil?
-          - order
-        elsif apple_property.nil? && !orange_property.nil?
-          + order
-        else
-          apple_property <=> orange_property
+      input.map { |item| [item_property(item, property), item] }
+        .sort! do |apple_info, orange_info|
+          apple_property = apple_info.first
+          orange_property = orange_info.first
+
+          if !apple_property.nil? && orange_property.nil?
+            - order
+          elsif apple_property.nil? && !orange_property.nil?
+            + order
+          else
+            apple_property <=> orange_property
+          end
         end
-      end
+        .map!(&:last)
     end
 
     private