Merge branch 'an-sidekiq-records-failure-durations' into 'master'

Record latencies for sidekiq failures

Closes gitlab-com/gl-infra/scalability#46

See merge request gitlab-org/gitlab!18909

changelogs/unreleased/an-sidekiq-records-failure-durations.yml

+---
+title: Record latencies for Sidekiq failures
+merge_request: 18909
+author:
+type: performance

lib/gitlab/sidekiq_middleware/metrics.rb

@@ -21,22 +21,28 @@ module Gitlab
           @metrics[:sidekiq_jobs_retried_total].increment(labels, 1)
         end
 
+        job_succeeded = false
+        monotonic_time_start = Gitlab::Metrics::System.monotonic_time
         job_thread_cputime_start = get_thread_cputime
-
-        realtime = Benchmark.realtime do
+        begin
           yield
-        end
-
+          job_succeeded = true
+        ensure
+          monotonic_time_end = Gitlab::Metrics::System.monotonic_time
           job_thread_cputime_end = get_thread_cputime
+
+          monotonic_time = monotonic_time_end - monotonic_time_start
           job_thread_cputime = job_thread_cputime_end - job_thread_cputime_start
-        @metrics[:sidekiq_jobs_cpu_seconds].observe(labels, job_thread_cputime)
 
-        @metrics[:sidekiq_jobs_completion_seconds].observe(labels, realtime)
-      rescue Exception # rubocop: disable Lint/RescueException
-        @metrics[:sidekiq_jobs_failed_total].increment(labels, 1)
-        raise
-      ensure
+          # sidekiq_running_jobs, sidekiq_jobs_failed_total should not include the job_status label
           @metrics[:sidekiq_running_jobs].increment(labels, -1)
+          @metrics[:sidekiq_jobs_failed_total].increment(labels, 1) unless job_succeeded
+
+          # job_status: done, fail match the job_status attribute in structured logging
+          labels[:job_status] = job_succeeded ? :done : :fail
+          @metrics[:sidekiq_jobs_cpu_seconds].observe(labels, job_thread_cputime)
+          @metrics[:sidekiq_jobs_completion_seconds].observe(labels, monotonic_time)
+        end
       end
 
       private

spec/lib/gitlab/sidekiq_middleware/metrics_spec.rb

@@ -44,12 +44,14 @@ describe Gitlab::SidekiqMiddleware::Metrics do
 
     it 'sets queue specific metrics' do
       labels = { queue: :test }
+      labels_with_job_status = { queue: :test, job_status: :done }
       allow(middleware).to receive(:get_thread_cputime).and_return(1, 3)
+      allow(Gitlab::Metrics::System).to receive(:monotonic_time).and_return(2, 3)
 
-      expect(user_execution_seconds_metric).to receive(:observe).with(labels, 2)
       expect(running_jobs_metric).to receive(:increment).with(labels, 1)
       expect(running_jobs_metric).to receive(:increment).with(labels, -1)
-      expect(completion_seconds_metric).to receive(:observe).with(labels, kind_of(Numeric))
+      expect(user_execution_seconds_metric).to receive(:observe).with(labels_with_job_status, 2)
+      expect(completion_seconds_metric).to receive(:observe).with(labels_with_job_status, 1)
 
       middleware.call(worker, {}, :test) { nil }
     end
@@ -74,8 +76,18 @@ describe Gitlab::SidekiqMiddleware::Metrics do
 
     context 'when error is raised' do
       it 'sets sidekiq_jobs_failed_total and reraises' do
-        expect(failed_total_metric).to receive(:increment)
-        expect { middleware.call(worker, {}, :test) { raise } }.to raise_error
+        labels = { queue: :test }
+        labels_with_job_status = { queue: :test, job_status: :fail }
+        allow(middleware).to receive(:get_thread_cputime).and_return(1, 4)
+        allow(Gitlab::Metrics::System).to receive(:monotonic_time).and_return(2, 6)
+
+        expect(running_jobs_metric).to receive(:increment).with(labels, 1)
+        expect(running_jobs_metric).to receive(:increment).with(labels, -1)
+        expect(failed_total_metric).to receive(:increment).with(labels, 1)
+        expect(user_execution_seconds_metric).to receive(:observe).with(labels_with_job_status, 3)
+        expect(completion_seconds_metric).to receive(:observe).with(labels_with_job_status, 4)
+
+        expect { middleware.call(worker, {}, :test) { raise StandardError, "Failed" } }.to raise_error(StandardError, "Failed")
       end
     end
   end