Only initialise metrics for the most common feature categories

Most feature categories aren't used at all yet, even on GitLab.com. This
meant we generated a huge amount of series that would never change. As a
compromise between initialising nothing (and having gaps, plus
unpredictable memory usage in Prometheus) and initialising
everything (explosion in number of series), we only initialise for the
top five categories as of today, plus the default category of `unknown`.

The top five categories do include `not_owned`, but that's probably OK.

lib/gitlab/metrics/requests_rack_middleware.rb

@@ -18,6 +18,15 @@ module Gitlab
       FEATURE_CATEGORY_HEADER = 'X-Gitlab-Feature-Category'
       FEATURE_CATEGORY_DEFAULT = 'unknown'
 
+      # These were the top 5 categories at a point in time, chosen as a
+      # reasonable default. If we initialize every category we'll end up
+      # with an explosion in unused metric combinations, but we want the
+      # most common ones to be always present.
+      FEATURE_CATEGORIES_TO_INITIALIZE = ['authentication_and_authorization',
+                                          'code_review', 'continuous_integration',
+                                          'not_owned', 'source_code_management',
+                                          FEATURE_CATEGORY_DEFAULT].freeze
+
       def initialize(app)
         @app = app
       end
@@ -46,16 +55,10 @@ module Gitlab
         #
         # For example `rate(http_requests_total{status="500"}[1m])` would return
         # no data until the first 500 error would occur.
-
-        # The list of feature categories is currently not needed by the application
-        # anywhere else. So no need to keep these in memory forever.
-        # Doing it here, means we're only reading the file on boot.
-        feature_categories = YAML.load_file(Rails.root.join('config', 'feature_categories.yml')).map(&:strip).uniq << FEATURE_CATEGORY_DEFAULT
-
         HTTP_METHODS.each do |method, statuses|
           http_request_duration_seconds.get({ method: method })
 
-          statuses.product(feature_categories) do |status, feature_category|
+          statuses.product(FEATURE_CATEGORIES_TO_INITIALIZE) do |status, feature_category|
             http_requests_total.get({ method: method, status: status, feature_category: feature_category })
           end
         end

spec/lib/gitlab/metrics/requests_rack_middleware_spec.rb

@@ -129,12 +129,11 @@ RSpec.describe Gitlab::Metrics::RequestsRackMiddleware, :aggregate_failures do
 
     describe '.initialize_metrics', :prometheus do
       it "sets labels for http_requests_total" do
-        feature_categories = YAML.load_file(Rails.root.join('config', 'feature_categories.yml')).map(&:strip) << described_class::FEATURE_CATEGORY_DEFAULT
         expected_labels = []
 
         described_class::HTTP_METHODS.each do |method, statuses|
           statuses.each do |status|
-            feature_categories.each do |feature_category|
+            described_class::FEATURE_CATEGORIES_TO_INITIALIZE.each do |feature_category|
               expected_labels << { method: method.to_s, status: status.to_s, feature_category: feature_category.to_s }
             end
           end
@@ -152,6 +151,13 @@ RSpec.describe Gitlab::Metrics::RequestsRackMiddleware, :aggregate_failures do
 
         expect(described_class.http_request_duration_seconds.values.keys).to include(*expected_labels)
       end
+
+      it 'has every label in config/feature_categories.yml' do
+        defaults = [described_class::FEATURE_CATEGORY_DEFAULT, 'not_owned']
+        feature_categories = YAML.load_file(Rails.root.join('config', 'feature_categories.yml')).map(&:strip) + defaults
+
+        expect(described_class::FEATURE_CATEGORIES_TO_INITIALIZE).to all(be_in(feature_categories))
+      end
     end
   end
 end