sidekiq 6.2.2 → 8.1.5

data/lib/sidekiq/manager.rb

@@ -1,9 +1,6 @@
 # frozen_string_literal: true
 
-require "sidekiq/util"
 require "sidekiq/processor"
-require "sidekiq/fetch"
-require "set"
 
 module Sidekiq
   ##
@@ -21,46 +18,38 @@ module Sidekiq
   # the shutdown process.  The other tasks are performed by other threads.
   #
   class Manager
-    include Util
+    include Sidekiq::Component
 
     attr_reader :workers
-    attr_reader :options
+    attr_reader :capsule
 
-    def initialize(options = {})
-      logger.debug { options.inspect }
-      @options = options
-      @count = options[:concurrency] || 10
+    def initialize(capsule)
+      @config = @capsule = capsule
+      @count = capsule.concurrency
       raise ArgumentError, "Concurrency of #{@count} is not supported" if @count < 1
 
       @done = false
       @workers = Set.new
+      @plock = Mutex.new
       @count.times do
-        @workers << Processor.new(self, options)
+        @workers << Processor.new(@config, &method(:processor_result))
       end
-      @plock = Mutex.new
     end
 
     def start
-      @workers.each do |x|
-        x.start
-      end
+      @workers.each(&:start)
     end
 
     def quiet
       return if @done
       @done = true
 
-      logger.info { "Terminating quiet workers" }
-      @workers.each { |x| x.terminate }
-      fire_event(:quiet, reverse: true)
+      logger.info { "Terminating quiet threads for #{capsule.name} capsule" }
+      @workers.each(&:terminate)
     end
 
-    # hack for quicker development / testing environment #2774
-    PAUSE_TIME = $stdout.tty? ? 0.1 : 0.5
-
     def stop(deadline)
       quiet
-      fire_event(:shutdown, reverse: true)
 
       # some of the shutdown events can be async,
       # we don't have any way to know when they're done but
@@ -68,29 +57,20 @@ module Sidekiq
       sleep PAUSE_TIME
       return if @workers.empty?
 
-      logger.info { "Pausing to allow workers to finish..." }
-      remaining = deadline - ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
-      while remaining > PAUSE_TIME
-        return if @workers.empty?
-        sleep PAUSE_TIME
-        remaining = deadline - ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
-      end
+      logger.info { "Pausing to allow jobs to finish..." }
+      wait_for(deadline) { @workers.empty? }
       return if @workers.empty?
 
       hard_shutdown
+    ensure
+      capsule.stop
     end
 
-    def processor_stopped(processor)
+    def processor_result(processor, reason = nil)
       @plock.synchronize do
         @workers.delete(processor)
-      end
-    end
-
-    def processor_died(processor, reason)
-      @plock.synchronize do
-        @workers.delete(processor)
-        unless @done
-          p = Processor.new(self, options)
+        if !@done && @count > @workers.size
+          p = Processor.new(@config, &method(:processor_result))
           @workers << p
           p.start
         end
@@ -104,7 +84,7 @@ module Sidekiq
     private
 
     def hard_shutdown
-      # We've reached the timeout and we still have busy workers.
+      # We've reached the timeout and we still have busy threads.
       # They must die but their jobs shall live on.
       cleanup = nil
       @plock.synchronize do
@@ -114,22 +94,40 @@ module Sidekiq
       if cleanup.size > 0
         jobs = cleanup.map { |p| p.job }.compact
 
-        logger.warn { "Terminating #{cleanup.size} busy worker threads" }
-        logger.warn { "Work still in progress #{jobs.inspect}" }
+        logger.warn { "Terminating #{cleanup.size} busy threads" }
+        logger.debug { "Jobs still in progress #{jobs.inspect}" }
 
         # Re-enqueue unfinished jobs
         # NOTE: You may notice that we may push a job back to redis before
-        # the worker thread is terminated. This is ok because Sidekiq's
+        # the thread is terminated. This is ok because Sidekiq's
         # contract says that jobs are run AT LEAST once. Process termination
         # is delayed until we're certain the jobs are back in Redis because
         # it is worse to lose a job than to run it twice.
-        strategy = @options[:fetch]
-        strategy.bulk_requeue(jobs, @options)
+        capsule.fetcher.bulk_requeue(jobs)
       end
 
       cleanup.each do |processor|
         processor.kill
       end
+
+      # when this method returns, we immediately call `exit` which may not give
+      # the remaining threads time to run `ensure` blocks, etc. We pause here up
+      # to 3 seconds to give threads a minimal amount of time to run `ensure` blocks.
+      deadline = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC) + 3
+      wait_for(deadline) { @workers.empty? }
+    end
+
+    # hack for quicker development / testing environment #2774
+    PAUSE_TIME = $stdout.tty? ? 0.1 : 0.5
+
+    # Wait for the orblock to be true or the deadline passed.
+    def wait_for(deadline, &condblock)
+      remaining = deadline - ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
+      while remaining > PAUSE_TIME
+        return if condblock.call
+        sleep PAUSE_TIME
+        remaining = deadline - ::Process.clock_gettime(::Process::CLOCK_MONOTONIC)
+      end
     end
   end
 end

data/lib/sidekiq/metrics/query.rb

@@ -0,0 +1,184 @@
+# frozen_string_literal: true
+
+require "date"
+require "sidekiq"
+require "sidekiq/metrics/shared"
+
+module Sidekiq
+  module Metrics
+    # Allows caller to query for Sidekiq execution metrics within Redis.
+    # Caller sets a set of attributes to act as filters. {#fetch} will call
+    # Redis and return a Hash of results.
+    #
+    # NB: all metrics and times/dates are UTC only. We explicitly do not
+    # support timezones.
+    class Query
+      def initialize(pool: nil, now: Time.now)
+        @time = now.utc
+        @pool = pool || Sidekiq.default_configuration.redis_pool
+        @klass = nil
+      end
+
+      ROLLUPS = {
+        # minutely aggregates per minute
+        minutely: [60, ->(time) { time.strftime("j|%y%m%d|%-H:%M") }],
+        # hourly aggregates every 10 minutes so we'll have six data points per hour
+        hourly: [600, ->(time) {
+          m = time.min
+          mins = (m < 10) ? "0" : m.to_s[0]
+          time.strftime("j|%y%m%d|%-H:#{mins}")
+        }]
+      }
+
+      # Get metric data for all jobs from the last hour
+      #  +class_filter+: return only results for classes matching filter
+      #  +minutes+: the number of fine-grained minute buckets to retrieve
+      #  +hours+: the number of coarser-grained 10-minute buckets to retrieve, in hours
+      def top_jobs(class_filter: nil, minutes: nil, hours: nil)
+        time = @time
+        minutes = 60 unless minutes || hours
+
+        # DoS protection, sanity check
+        minutes = 60 if minutes && minutes > 480
+        hours = 72 if hours && hours > 72
+
+        granularity = hours ? :hourly : :minutely
+        result = Result.new(granularity)
+        result.ends_at = time
+        count = hours ? hours * 6 : minutes
+        stride, keyproc = ROLLUPS[granularity]
+
+        redis_results = @pool.with do |conn|
+          conn.pipelined do |pipe|
+            count.times do |idx|
+              key = keyproc.call(time)
+              pipe.hgetall key
+              time -= stride
+            end
+          end
+        end
+
+        result.starts_at = time
+        time = @time
+        redis_results.each do |hash|
+          hash.each do |k, v|
+            kls, metric = k.split("|")
+            next if class_filter && !class_filter.match?(kls)
+            result.job_results[kls].add_metric metric, time, v.to_i
+          end
+          time -= stride
+        end
+
+        result.marks = fetch_marks(result.starts_at..result.ends_at, granularity)
+        result
+      end
+
+      def for_job(klass, minutes: nil, hours: nil)
+        time = @time
+        minutes = 60 unless minutes || hours
+
+        # DoS protection, sanity check
+        minutes = 60 if minutes && minutes > 480
+        hours = 72 if hours && hours > 72
+
+        granularity = hours ? :hourly : :minutely
+        result = Result.new(granularity)
+        result.ends_at = time
+        count = hours ? hours * 6 : minutes
+        stride, keyproc = ROLLUPS[granularity]
+
+        redis_results = @pool.with do |conn|
+          conn.pipelined do |pipe|
+            count.times do |idx|
+              key = keyproc.call(time)
+              pipe.hmget key, "#{klass}|ms", "#{klass}|p", "#{klass}|f"
+              time -= stride
+            end
+          end
+        end
+
+        result.starts_at = time
+        time = @time
+        @pool.with do |conn|
+          redis_results.each do |(ms, p, f)|
+            result.job_results[klass].add_metric "ms", time, ms.to_i if ms
+            result.job_results[klass].add_metric "p", time, p.to_i if p
+            result.job_results[klass].add_metric "f", time, f.to_i if f
+            result.job_results[klass].add_hist time, Histogram.new(klass).fetch(conn, time).reverse if minutes
+            time -= stride
+          end
+        end
+
+        result.marks = fetch_marks(result.starts_at..result.ends_at, granularity)
+        result
+      end
+
+      class Result < Struct.new(:granularity, :starts_at, :ends_at, :size, :job_results, :marks)
+        def initialize(granularity = :minutely)
+          super
+          self.granularity = granularity
+          self.marks = []
+          self.job_results = Hash.new { |h, k| h[k] = JobResult.new(granularity) }
+        end
+      end
+
+      class JobResult < Struct.new(:granularity, :series, :hist, :totals)
+        def initialize(granularity = :minutely)
+          super
+          self.granularity = granularity
+          self.series = Hash.new { |h, k| h[k] = Hash.new(0) }
+          self.hist = Hash.new { |h, k| h[k] = [] }
+          self.totals = Hash.new(0)
+        end
+
+        def add_metric(metric, time, value)
+          totals[metric] += value
+          series[metric][Query.bkt_time_s(time, granularity)] += value
+
+          # Include timing measurements in seconds for convenience
+          add_metric("s", time, value / 1000.0) if metric == "ms"
+        end
+
+        def add_hist(time, hist_result)
+          hist[Query.bkt_time_s(time, granularity)] = hist_result
+        end
+
+        def total_avg(metric = "ms")
+          completed = totals["p"] - totals["f"]
+          return 0 if completed.zero?
+          totals[metric].to_f / completed
+        end
+
+        def series_avg(metric = "ms")
+          series[metric].each_with_object(Hash.new(0)) do |(bucket, value), result|
+            completed = series.dig("p", bucket) - series.dig("f", bucket)
+            result[bucket] = (completed == 0) ? 0 : value.to_f / completed
+          end
+        end
+      end
+
+      MarkResult = Struct.new(:time, :label, :bucket)
+
+      def self.bkt_time_s(time, granularity)
+        # truncate time to ten minutes ("8:40", not "8:43") or one minute
+        truncation = (granularity == :hourly) ? 600 : 60
+        Time.at(time.to_i - time.to_i % truncation).utc.iso8601
+      end
+
+      private
+
+      def fetch_marks(time_range, granularity)
+        [].tap do |result|
+          marks = @pool.with { |c| c.hgetall("#{@time.strftime("%Y%m%d")}-marks") }
+
+          marks.each do |timestamp, label|
+            time = Time.parse(timestamp)
+            if time_range.cover? time
+              result << MarkResult.new(time, label, Query.bkt_time_s(time, granularity))
+            end
+          end
+        end
+      end
+    end
+  end
+end

data/lib/sidekiq/metrics/shared.rb

@@ -0,0 +1,109 @@
+# frozen_string_literal: true
+
+module Sidekiq
+  module Metrics
+    class Counter
+      def initialize
+        @value = 0
+        @lock = Mutex.new
+      end
+
+      def increment
+        @lock.synchronize { @value += 1 }
+      end
+
+      def value
+        @lock.synchronize { @value }
+      end
+    end
+
+    # Implements space-efficient but statistically useful histogram storage.
+    # A precise time histogram stores every time. Instead we break times into a set of
+    # known buckets and increment counts of the associated time bucket. Even if we call
+    # the histogram a million times, we'll still only store 26 buckets.
+    # NB: needs to be thread-safe or resiliant to races.
+    #
+    # To store this data, we use Redis' BITFIELD command to store unsigned 16-bit counters
+    # per bucket per klass per minute. It's unlikely that most people will be executing more
+    # than 1000 job/sec for a full minute of a specific type (i.e. overflow 65,536).
+    #
+    # Histograms are only stored at the fine-grained level, they are not rolled up
+    # for longer-term buckets.
+    class Histogram
+      include Enumerable
+
+      # This number represents the maximum milliseconds for this bucket.
+      # 20 means all job executions up to 20ms, e.g. if a job takes
+      # 280ms, it'll increment bucket[7]. Note we can track job executions
+      # up to about 5.5 minutes. After that, it's assumed you're probably
+      # not too concerned with its performance.
+      BUCKET_INTERVALS = [
+        20, 30, 45, 65, 100,
+        150, 225, 335, 500, 750,
+        1100, 1700, 2500, 3800, 5750,
+        8500, 13000, 20000, 30000, 45000,
+        65000, 100000, 150000, 225000, 335000,
+        1e20 # the "maybe your job is too long" bucket
+      ].freeze
+      LABELS = [
+        "20ms", "30ms", "45ms", "65ms", "100ms",
+        "150ms", "225ms", "335ms", "500ms", "750ms",
+        "1.1s", "1.7s", "2.5s", "3.8s", "5.75s",
+        "8.5s", "13s", "20s", "30s", "45s",
+        "65s", "100s", "150s", "225s", "335s",
+        "Slow"
+      ].freeze
+      FETCH = "GET u16 #0 GET u16 #1 GET u16 #2 GET u16 #3 \
+        GET u16 #4 GET u16 #5 GET u16 #6 GET u16 #7 \
+        GET u16 #8 GET u16 #9 GET u16 #10 GET u16 #11 \
+        GET u16 #12 GET u16 #13 GET u16 #14 GET u16 #15 \
+        GET u16 #16 GET u16 #17 GET u16 #18 GET u16 #19 \
+        GET u16 #20 GET u16 #21 GET u16 #22 GET u16 #23 \
+        GET u16 #24 GET u16 #25".split
+      HISTOGRAM_TTL = 8 * 60 * 60
+
+      def each
+        buckets.each { |counter| yield counter.value }
+      end
+
+      def label(idx)
+        LABELS[idx]
+      end
+
+      attr_reader :buckets
+      def initialize(klass)
+        @klass = klass
+        @buckets = Array.new(BUCKET_INTERVALS.size) { Counter.new }
+      end
+
+      def record_time(ms)
+        index_to_use = BUCKET_INTERVALS.each_index do |idx|
+          break idx if ms < BUCKET_INTERVALS[idx]
+        end
+
+        @buckets[index_to_use].increment
+      end
+
+      def fetch(conn, now = Time.now)
+        window = now.utc.strftime("%-d-%-H:%-M")
+        key = "h|#{@klass}-#{window}"
+        conn.bitfield_ro(key, *FETCH)
+      end
+
+      def persist(conn, now = Time.now)
+        buckets, @buckets = @buckets, []
+        window = now.utc.strftime("%-d-%-H:%-M")
+        key = "h|#{@klass}-#{window}"
+        cmd = [key, "OVERFLOW", "SAT"]
+        buckets.each_with_index do |counter, idx|
+          val = counter.value
+          cmd << "INCRBY" << "u16" << "##{idx}" << val.to_s if val > 0
+        end
+
+        conn.bitfield(*cmd) if cmd.size > 3
+        conn.expire(key, HISTOGRAM_TTL)
+        key
+      end
+    end
+  end
+end

data/lib/sidekiq/metrics/tracking.rb

@@ -0,0 +1,153 @@
+# frozen_string_literal: true
+
+require "time"
+require "sidekiq"
+require "sidekiq/metrics/shared"
+
+# This file contains the components which track execution metrics within Sidekiq.
+module Sidekiq
+  module Metrics
+    class ExecutionTracker
+      include Sidekiq::Component
+
+      def initialize(config)
+        @config = config
+        @jobs = Hash.new(0)
+        @totals = Hash.new(0)
+        @grams = Hash.new { |hash, key| hash[key] = Histogram.new(key) }
+        @lock = Mutex.new
+      end
+
+      def track(queue, klass)
+        start = mono_ms
+        time_ms = 0
+        begin
+          begin
+            yield
+          ensure
+            finish = mono_ms
+            time_ms = finish - start
+          end
+          # We don't track time for failed jobs as they can have very unpredictable
+          # execution times. more important to know average time for successful jobs so we
+          # can better recognize when a perf regression is introduced.
+          track_time(klass, time_ms)
+        rescue JobRetry::Skip
+          # This is raised when iterable job is interrupted.
+          track_time(klass, time_ms)
+          raise
+        rescue Exception
+          @lock.synchronize {
+            @jobs["#{klass}|f"] += 1
+            @totals["f"] += 1
+          }
+          raise
+        ensure
+          @lock.synchronize {
+            @jobs["#{klass}|p"] += 1
+            @totals["p"] += 1
+          }
+        end
+      end
+
+      # LONG_TERM = 90 * 24 * 60 * 60
+      MID_TERM = 3 * 24 * 60 * 60
+      SHORT_TERM = 8 * 60 * 60
+
+      def flush(time = Time.now)
+        totals, jobs, grams = reset
+        procd = totals["p"]
+        fails = totals["f"]
+        return if procd == 0 && fails == 0
+
+        now = time.utc
+        # nowdate = now.strftime("%Y%m%d")
+        # "250214|8:4" is the 10 minute bucket for Feb 14 2025, 08:43
+        nowmid = now.strftime("%y%m%d|%-H:%M")[0..-2]
+        # "250214|8:43" is the 1 minute bucket for Feb 14 2025, 08:43
+        nowshort = now.strftime("%y%m%d|%-H:%M")
+        count = 0
+
+        redis do |conn|
+          # persist fine-grained histogram data
+          if grams.size > 0
+            conn.pipelined do |pipe|
+              grams.each do |_, gram|
+                gram.persist(pipe, now)
+              end
+            end
+          end
+
+          # persist coarse grained execution count + execution millis.
+          # note as of today we don't use or do anything with the
+          # daily or hourly rollups.
+          [
+            # ["j", jobs, nowdate, LONG_TERM],
+            ["j", jobs, nowmid, MID_TERM],
+            ["j", jobs, nowshort, SHORT_TERM]
+          ].each do |prefix, data, bucket, ttl|
+            conn.pipelined do |xa|
+              stats = "#{prefix}|#{bucket}"
+              data.each_pair do |key, value|
+                xa.hincrby stats, key, value
+                count += 1
+              end
+              xa.expire(stats, ttl)
+            end
+          end
+          logger.debug "Flushed #{count} metrics"
+          count
+        end
+      end
+
+      private
+
+      def track_time(klass, time_ms)
+        @lock.synchronize {
+          @grams[klass].record_time(time_ms)
+          @jobs["#{klass}|ms"] += time_ms
+          @totals["ms"] += time_ms
+        }
+      end
+
+      def reset
+        @lock.synchronize {
+          array = [@totals, @jobs, @grams]
+          reset_instance_variables
+          array
+        }
+      end
+
+      def reset_instance_variables
+        @totals = Hash.new(0)
+        @jobs = Hash.new(0)
+        @grams = Hash.new { |hash, key| hash[key] = Histogram.new(key) }
+      end
+    end
+
+    class Middleware
+      include Sidekiq::ServerMiddleware
+
+      def initialize(options)
+        @exec = options
+      end
+
+      def call(_instance, hash, queue, &block)
+        @exec.track(queue, hash["wrapped"] || hash["class"], &block)
+      end
+    end
+  end
+end
+
+Sidekiq.configure_server do |config|
+  exec = Sidekiq::Metrics::ExecutionTracker.new(config)
+  config.server_middleware do |chain|
+    chain.add Sidekiq::Metrics::Middleware, exec
+  end
+  config.on(:beat) do
+    exec.flush
+  end
+  config.on(:exit) do
+    exec.flush
+  end
+end