sidekiq 6.5.0 → 6.5.3

data/lib/sidekiq/job_retry.rb

@@ -91,7 +91,7 @@ module Sidekiq
 
       msg = Sidekiq.load_json(jobstr)
       if msg["retry"]
-        attempt_retry(nil, msg, queue, e)
+        process_retry(nil, msg, queue, e)
       else
         Sidekiq.death_handlers.each do |handler|
           handler.call(msg, e)
@@ -128,7 +128,7 @@ module Sidekiq
       end
 
       raise e unless msg["retry"]
-      attempt_retry(jobinst, msg, queue, e)
+      process_retry(jobinst, msg, queue, e)
       # We've handled this error associated with this job, don't
       # need to handle it at the global level
       raise Skip
@@ -139,7 +139,7 @@ module Sidekiq
     # Note that +jobinst+ can be nil here if an error is raised before we can
     # instantiate the job instance.  All access must be guarded and
     # best effort.
-    def attempt_retry(jobinst, msg, queue, exception)
+    def process_retry(jobinst, msg, queue, exception)
       max_retry_attempts = retry_attempts_from(msg["retry"], @max_retries)
 
       msg["queue"] = (msg["retry_queue"] || queue)
@@ -170,19 +170,50 @@ module Sidekiq
         msg["error_backtrace"] = compress_backtrace(lines)
       end
 
-      if count < max_retry_attempts
-        delay = delay_for(jobinst, count, exception)
-        # Logging here can break retries if the logging device raises ENOSPC #3979
-        # logger.debug { "Failure! Retry #{count} in #{delay} seconds" }
-        retry_at = Time.now.to_f + delay
-        payload = Sidekiq.dump_json(msg)
-        Sidekiq.redis do |conn|
-          conn.zadd("retry", retry_at.to_s, payload)
-        end
+      # Goodbye dear message, you (re)tried your best I'm sure.
+      return retries_exhausted(jobinst, msg, exception) if count >= max_retry_attempts
+
+      strategy, delay = delay_for(jobinst, count, exception)
+      case strategy
+      when :discard
+        return # poof!
+      when :kill
+        return retries_exhausted(jobinst, msg, exception)
+      end
+
+      # Logging here can break retries if the logging device raises ENOSPC #3979
+      # logger.debug { "Failure! Retry #{count} in #{delay} seconds" }
+      jitter = rand(10) * (count + 1)
+      retry_at = Time.now.to_f + delay + jitter
+      payload = Sidekiq.dump_json(msg)
+      redis do |conn|
+        conn.zadd("retry", retry_at.to_s, payload)
+      end
+    end
+
+    # returns (strategy, seconds)
+    def delay_for(jobinst, count, exception)
+      rv = begin
+        # sidekiq_retry_in can return two different things:
+        # 1. When to retry next, as an integer of seconds
+        # 2. A symbol which re-routes the job elsewhere, e.g. :discard, :kill, :default
+        jobinst&.sidekiq_retry_in_block&.call(count, exception)
+      rescue Exception => e
+        handle_exception(e, {context: "Failure scheduling retry using the defined `sidekiq_retry_in` in #{jobinst.class.name}, falling back to default"})
+        nil
+      end
+
+      delay = if Integer === rv && rv > 0
+        rv
+      elsif rv == :discard
+        return [:discard, nil] # do nothing, job goes poof
+      elsif rv == :kill
+        return [:kill, nil]
       else
-        # Goodbye dear message, you (re)tried your best I'm sure.
-        retries_exhausted(jobinst, msg, exception)
+        (count**4) + 15
       end
+
+      [:default, delay]
     end
 
     def retries_exhausted(jobinst, msg, exception)
@@ -195,7 +226,7 @@ module Sidekiq
 
       send_to_morgue(msg) unless msg["dead"] == false
 
-      Sidekiq.death_handlers.each do |handler|
+      config.death_handlers.each do |handler|
         handler.call(msg, exception)
       rescue => e
         handle_exception(e, {context: "Error calling death handler", job: msg})
@@ -216,22 +247,6 @@ module Sidekiq
       end
     end
 
-    def delay_for(jobinst, count, exception)
-      jitter = rand(10) * (count + 1)
-      if jobinst&.sidekiq_retry_in_block
-        custom_retry_in = retry_in(jobinst, count, exception).to_i
-        return custom_retry_in + jitter if custom_retry_in > 0
-      end
-      (count**4) + 15 + jitter
-    end
-
-    def retry_in(jobinst, count, exception)
-      jobinst.sidekiq_retry_in_block.call(count, exception)
-    rescue Exception => e
-      handle_exception(e, {context: "Failure scheduling retry using the defined `sidekiq_retry_in` in #{jobinst.class.name}, falling back to default"})
-      nil
-    end
-
     def exception_caused_by_shutdown?(e, checked_causes = [])
       return false unless e.cause
 

data/lib/sidekiq/launcher.rb

@@ -79,6 +79,8 @@ module Sidekiq
     end
 
     def clear_heartbeat
+      flush_stats
+
       # Remove record from Redis since we are shutting down.
       # Note we don't stop the heartbeat thread; if the process
       # doesn't actually exit, it'll reappear in the Web UI.
@@ -98,7 +100,7 @@ module Sidekiq
       ❤
     end
 
-    def self.flush_stats
+    def flush_stats
       fails = Processor::FAILURE.reset
       procd = Processor::PROCESSED.reset
       return if fails + procd == 0
@@ -122,7 +124,6 @@ module Sidekiq
         Sidekiq.logger.warn("Unable to flush stats: #{ex}")
       end
     end
-    at_exit(&method(:flush_stats))
 
     def ❤
       key = identity
@@ -179,6 +180,7 @@ module Sidekiq
 
         # first heartbeat or recovering from an outage and need to reestablish our heartbeat
         fire_event(:heartbeat) unless exists
+        fire_event(:beat, oneshot: false)
 
         return unless msg
 

data/lib/sidekiq/metrics/deploy.rb

@@ -0,0 +1,47 @@
+require "sidekiq"
+require "date"
+
+# This file is designed to be required within the user's
+# deployment script; it should need a bare minimum of dependencies.
+#
+#   require "sidekiq/metrics/deploy"
+#   gitdesc = `git log -1 --format="%h %s"`.strip
+#   d = Sidekiq::Metrics::Deploy.new
+#   d.mark(label: gitdesc)
+#
+# Note that you cannot mark more than once per minute. This is a feature, not a bug.
+module Sidekiq
+  module Metrics
+    class Deploy
+      MARK_TTL = 90 * 24 * 60 * 60 # 90 days
+
+      def initialize(pool = Sidekiq.redis_pool)
+        @pool = pool
+      end
+
+      def mark(at: Time.now, label: "")
+        # we need to round the timestamp so that we gracefully
+        # handle an excepted common error in marking deploys:
+        # having every process mark its deploy, leading
+        # to N marks for each deploy. Instead we round the time
+        # to the minute so that multple marks within that minute
+        # will all naturally rollup into one mark per minute.
+        whence = at.utc
+        floor = Time.utc(whence.year, whence.month, whence.mday, whence.hour, whence.min, 0)
+        datecode = floor.strftime("%Y%m%d")
+        key = "#{datecode}-marks"
+        @pool.with do |c|
+          c.pipelined do |pipe|
+            pipe.hsetnx(key, floor.rfc3339, label)
+            pipe.expire(key, MARK_TTL)
+          end
+        end
+      end
+
+      def fetch(date = Time.now.utc.to_date)
+        datecode = date.strftime("%Y%m%d")
+        @pool.with { |c| c.hgetall("#{datecode}-marks") }
+      end
+    end
+  end
+end

data/lib/sidekiq/metrics/query.rb

@@ -0,0 +1,124 @@
+require "sidekiq"
+require "date"
+require "set"
+
+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 specifically do not
+    # support timezones.
+    class Query
+      # :hour, :day, :month
+      attr_accessor :period
+
+      # a specific job class, e.g. "App::OrderJob"
+      attr_accessor :klass
+
+      # the date specific to the period
+      # for :day or :hour, something like Date.today or Date.new(2022, 7, 13)
+      # for :month, Date.new(2022, 7, 1)
+      attr_accessor :date
+
+      # for period = :hour, the specific hour, integer e.g. 1 or 18
+      # note that hours and minutes do not have a leading zero so minute-specific
+      # keys will look like "j|20220718|7:3" for data at 07:03.
+      attr_accessor :hour
+
+      def initialize(pool: Sidekiq.redis_pool, now: Time.now)
+        @time = now.utc
+        @pool = pool
+        @klass = nil
+      end
+
+      # Get metric data from the last hour and roll it up
+      # into top processed count and execution time based on class.
+      def top_jobs
+        resultset = {}
+        resultset[:date] = @time.to_date
+        resultset[:period] = :hour
+        resultset[:ends_at] = @time
+        time = @time
+
+        results = @pool.with do |conn|
+          conn.pipelined do |pipe|
+            resultset[:size] = 60
+            60.times do |idx|
+              key = "j|#{time.strftime("%Y%m%d")}|#{time.hour}:#{time.min}"
+              pipe.hgetall key
+              time -= 60
+            end
+            resultset[:starts_at] = time
+          end
+        end
+
+        t = Hash.new(0)
+        klsset = Set.new
+        # merge the per-minute data into a totals hash for the hour
+        results.each do |hash|
+          hash.each { |k, v| t[k] = t[k] + v.to_i }
+          klsset.merge(hash.keys.map { |k| k.split("|")[0] })
+        end
+        resultset[:job_classes] = klsset.delete_if { |item| item.size < 3 }
+        resultset[:totals] = t
+        top = t.each_with_object({}) do |(k, v), memo|
+          (kls, metric) = k.split("|")
+          memo[metric] ||= Hash.new(0)
+          memo[metric][kls] = v
+        end
+
+        sorted = {}
+        top.each_pair do |metric, hash|
+          sorted[metric] = hash.sort_by { |k, v| v }.reverse.to_h
+        end
+        resultset[:top_classes] = sorted
+        resultset
+      end
+
+      def for_job(klass)
+        resultset = {}
+        resultset[:date] = @time.to_date
+        resultset[:period] = :hour
+        resultset[:ends_at] = @time
+        marks = @pool.with { |c| c.hgetall("#{@time.strftime("%Y%m%d")}-marks") }
+
+        time = @time
+        initial = @pool.with do |conn|
+          conn.pipelined do |pipe|
+            resultset[:size] = 60
+            60.times do |idx|
+              key = "j|#{time.strftime("%Y%m%d|%-H:%-M")}"
+              pipe.hmget key, "#{klass}|ms", "#{klass}|p", "#{klass}|f"
+              time -= 60
+            end
+          end
+        end
+
+        time = @time
+        hist = Histogram.new(klass)
+        results = @pool.with do |conn|
+          initial.map do |(ms, p, f)|
+            tm = Time.utc(time.year, time.month, time.mday, time.hour, time.min, 0)
+            {
+              time: tm.iso8601,
+              epoch: tm.to_i,
+              ms: ms.to_i, p: p.to_i, f: f.to_i, hist: hist.fetch(conn, time)
+            }.tap { |x|
+              x[:mark] = marks[x[:time]] if marks[x[:time]]
+              time -= 60
+            }
+          end
+        end
+
+        resultset[:marks] = marks
+        resultset[:starts_at] = time
+        resultset[:data] = results
+        resultset
+      end
+    end
+  end
+end

data/lib/sidekiq/metrics/shared.rb

@@ -0,0 +1,94 @@
+require "concurrent"
+
+module Sidekiq
+  module Metrics
+    # TODO Support apps without concurrent-ruby
+    Counter = ::Concurrent::AtomicFixnum
+
+    # 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.
+    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,
+        Float::INFINITY # the "maybe your job is too long" bucket
+      ]
+      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"
+      ]
+
+      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
+
+      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 = "#{@klass}-#{window}"
+        conn.bitfield(key, *FETCH)
+      end
+
+      def persist(conn, now = Time.now)
+        buckets, @buckets = @buckets, []
+        window = now.utc.strftime("%d-%H:%-M")
+        key = "#{@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, 86400)
+        key
+      end
+    end
+  end
+end

data/lib/sidekiq/metrics/tracking.rb

@@ -0,0 +1,134 @@
+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 = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC, :millisecond)
+        time_ms = 0
+        begin
+          begin
+            yield
+          ensure
+            finish = ::Process.clock_gettime(::Process::CLOCK_MONOTONIC, :millisecond)
+            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.
+          @lock.synchronize {
+            @grams[klass].record_time(time_ms)
+            @jobs["#{klass}|ms"] += time_ms
+            @totals["ms"] += time_ms
+          }
+        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 = 7 * 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")
+        nowhour = now.strftime("%Y%m%d|%-H")
+        nowmin = now.strftime("%Y%m%d|%-H:%-M")
+        count = 0
+
+        redis do |conn|
+          if grams.size > 0
+            conn.pipelined do |pipe|
+              grams.each do |_, gram|
+                gram.persist(pipe, now)
+              end
+            end
+          end
+
+          [
+            ["j", jobs, nowdate, LONG_TERM],
+            ["j", jobs, nowhour, MID_TERM],
+            ["j", jobs, nowmin, SHORT_TERM]
+          ].each do |prefix, data, bucket, ttl|
+            # Quietly seed the new 7.0 stats format so migration is painless.
+            conn.pipelined do |xa|
+              stats = "#{prefix}|#{bucket}"
+              # logger.debug "Flushing metrics #{stats}"
+              data.each_pair do |key, value|
+                xa.hincrby stats, key, value
+                count += 1
+              end
+              xa.expire(stats, ttl)
+            end
+          end
+          logger.info "Flushed #{count} metrics"
+          count
+        end
+      end
+
+      private
+
+      def reset
+        @lock.synchronize {
+          array = [@totals, @jobs, @grams]
+          @totals = Hash.new(0)
+          @jobs = Hash.new(0)
+          @grams = Hash.new { |hash, key| hash[key] = Histogram.new(key) }
+          array
+        }
+      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
+
+if ENV["SIDEKIQ_METRICS_BETA"] == "1"
+  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
+  end
+end