Avoid redundant metrics collection

judoscale-sidekiq/lib/judoscale/sidekiq/metrics_collector.rb

@@ -6,11 +6,17 @@
 module Judoscale
   module Sidekiq
     class MetricsCollector < Judoscale::JobMetricsCollector
+      RECENT = 1 # second
+      RECENT_KEY = "judoscale:sidekiq:recent"
+      RECENT_VALUE = "1"
+
       def self.adapter_config
         Judoscale::Config.instance.sidekiq
       end
 
       def collect
+        return [] if collected_recently?
+
         metrics = []
         queues_by_name = ::Sidekiq::Queue.all.each_with_object({}) do |queue, obj|
           obj[queue.name] = queue
@@ -45,6 +51,18 @@ def collect
         log_collection(metrics)
         metrics
       end
+
+      def forget_recent_collection!
+        # We need this for testing
+        ::Sidekiq.redis { |r| r.del RECENT_KEY }
+      end
+
+      private
+
+      def collected_recently?
+        # If another process has collected metrics recently, we don't need to.
+        !::Sidekiq.redis { |r| r.set RECENT_KEY, RECENT_VALUE, nx: true, ex: RECENT }
+      end
     end
   end
 end

judoscale-sidekiq/test/benchmarks/collect_with_many_reporters_benchmark.rb

@@ -0,0 +1,43 @@
+# frozen_string_literal: true
+
+require "test_helper"
+require "minitest/benchmark"
+require "judoscale/sidekiq/metrics_collector"
+
+class CollectWithManyReportersBenchmark < Minitest::Benchmark
+  BATCH_SIZE = 1_000
+  QUEUES = %w[one two three four five six seven eight nine ten]
+
+  # performance assertions will iterate over `bench_range`.
+  # The values here don't matter—we just want several iterations through the benchmark.
+  def self.bench_range
+    (0..4).to_a
+  end
+
+  def setup
+    # Enqueue jobs on several queues
+    sidekiq_args = BATCH_SIZE.times.map { [] }
+    QUEUES.each do |queue|
+      Sidekiq::Client.push_bulk "class" => "Foo", "args" => sidekiq_args, "queue" => queue
+    end
+
+    # Prepare a collector for each benchmark iteration
+    @collectors = {}
+    self.class.bench_range.each do |n|
+      @collectors[n] = Judoscale::Sidekiq::MetricsCollector.new
+    end
+  end
+
+  def bench_collect
+    validation = proc do |_, times|
+      # The first collector should take the longest, since the rest will be
+      # no-ops after checking for `collected_recently?`.
+      first, *rest = times
+      assert_operator first, :>, rest.max
+    end
+
+    assert_performance validation do |n|
+      @collectors.fetch(n).collect
+    end
+  end
+end

judoscale-sidekiq/test/metrics_collector_test.rb

@@ -10,9 +10,8 @@ module Judoscale
     subject { Sidekiq::MetricsCollector.new }
 
     describe "#collect" do
-      after {
-        subject.clear_queues
-      }
+      before { ::Sidekiq.redis { |r| r.flushdb } }
+      after { subject.clear_queues }
 
       it "collects latency for each queue" do
         queues = [
@@ -39,6 +38,22 @@ module Judoscale
         _(metrics[3].identifier).must_equal :qd
       end
 
+      it "avoids redundant collections" do
+        queues = [SidekiqQueueStub.new(name: "default", latency: 11, size: 1)]
+
+        metrics = ::Sidekiq::Queue.stub(:all, queues) {
+          subject.collect
+        }
+
+        _(metrics.size).must_equal 2
+
+        metrics = ::Sidekiq::Queue.stub(:all, queues) {
+          subject.collect
+        }
+
+        _(metrics.size).must_equal 0
+      end
+
       it "always collects for known queues" do
         queues = []
 
@@ -51,6 +66,7 @@ module Judoscale
         queues = [SidekiqQueueStub.new(name: "default", latency: 11, size: 1)]
 
         metrics = ::Sidekiq::Queue.stub(:all, queues) {
+          subject.forget_recent_collection!
           subject.collect
         }
 
@@ -63,6 +79,7 @@ module Judoscale
 
         metrics = ::Sidekiq::Queue.stub(:all, queues) {
           ::Sidekiq::Queue.stub(:new, ->(queue_name) { new_queues.fetch(queue_name) }) {
+            subject.forget_recent_collection!
             subject.collect
           }
         }