ruby-kafka 0.7.8 → 1.1.0

data/lib/kafka/connection.rb

@@ -58,6 +58,9 @@ module Kafka
       @connect_timeout = connect_timeout || CONNECT_TIMEOUT
       @socket_timeout = socket_timeout || SOCKET_TIMEOUT
       @ssl_context = ssl_context
+
+      @socket = nil
+      @last_request = nil
     end
 
     def to_s

data/lib/kafka/consumer.rb

@@ -44,7 +44,7 @@ module Kafka
   #
   class Consumer
 
-    def initialize(cluster:, logger:, instrumenter:, group:, fetcher:, offset_manager:, session_timeout:, heartbeat:)
+    def initialize(cluster:, logger:, instrumenter:, group:, fetcher:, offset_manager:, session_timeout:, heartbeat:, refresh_topic_interval: 0)
       @cluster = cluster
       @logger = TaggedLogger.new(logger)
       @instrumenter = instrumenter
@@ -53,6 +53,7 @@ module Kafka
       @session_timeout = session_timeout
       @fetcher = fetcher
       @heartbeat = heartbeat
+      @refresh_topic_interval = refresh_topic_interval
 
       @pauses = Hash.new {|h, k|
         h[k] = Hash.new {|h2, k2|
@@ -73,6 +74,15 @@ module Kafka
       #   when user commits message other than last in a batch, this would make ruby-kafka refetch
       #   some already consumed messages
       @current_offsets = Hash.new { |h, k| h[k] = {} }
+
+      # Map storing subscribed topics with their configuration
+      @subscribed_topics = Concurrent::Map.new
+
+      # Set storing topics that matched topics in @subscribed_topics
+      @matched_topics = Set.new
+
+      # Whether join_group must be executed again because new topics are added
+      @join_group_for_new_topics = false
     end
 
     # Subscribes the consumer to a topic.
@@ -97,13 +107,12 @@ module Kafka
     def subscribe(topic_or_regex, default_offset: nil, start_from_beginning: true, max_bytes_per_partition: 1048576)
       default_offset ||= start_from_beginning ? :earliest : :latest
 
-      if topic_or_regex.is_a?(Regexp)
-        cluster_topics.select { |topic| topic =~ topic_or_regex }.each do |topic|
-          subscribe_to_topic(topic, default_offset, start_from_beginning, max_bytes_per_partition)
-        end
-      else
-        subscribe_to_topic(topic_or_regex, default_offset, start_from_beginning, max_bytes_per_partition)
-      end
+      @subscribed_topics[topic_or_regex] = {
+        default_offset: default_offset,
+        start_from_beginning: start_from_beginning,
+        max_bytes_per_partition: max_bytes_per_partition
+      }
+      scan_for_subscribing
 
       nil
     end
@@ -116,7 +125,6 @@ module Kafka
     def stop
       @running = false
       @fetcher.stop
-      @cluster.disconnect
     end
 
     # Pause processing of a specific topic partition.
@@ -308,6 +316,7 @@ module Kafka
               topic: batch.topic,
               partition: batch.partition,
               last_offset: batch.last_offset,
+              last_create_time: batch.messages.last.try(:create_time),
               offset_lag: batch.offset_lag,
               highwater_mark_offset: batch.highwater_mark_offset,
               message_count: batch.messages.count,
@@ -401,6 +410,7 @@ module Kafka
       while running?
         begin
           @instrumenter.instrument("loop.consumer") do
+            refresh_topic_list_if_enabled
             yield
           end
         rescue HeartbeatError
@@ -432,6 +442,7 @@ module Kafka
       # important that members explicitly tell Kafka when they're leaving.
       make_final_offsets_commit!
       @group.leave rescue nil
+      @cluster.disconnect
       @running = false
       @logger.pop_tags
     end
@@ -452,6 +463,8 @@ module Kafka
     end
 
     def join_group
+      @join_group_for_new_topics = false
+
       old_generation_id = @group.generation_id
 
       @group.join
@@ -513,11 +526,19 @@ module Kafka
       end
     end
 
+    def refresh_topic_list_if_enabled
+      return if @refresh_topic_interval <= 0
+      return if @refreshed_at && @refreshed_at + @refresh_topic_interval > Time.now
+
+      scan_for_subscribing
+      @refreshed_at = Time.now
+    end
+
     def fetch_batches
       # Return early if the consumer has been stopped.
       return [] if shutting_down?
 
-      join_group unless @group.member?
+      join_group if !@group.member? || @join_group_for_new_topics
 
       trigger_heartbeat
 
@@ -525,7 +546,7 @@ module Kafka
 
       if !@fetcher.data?
         @logger.debug "No batches to process"
-        sleep 2
+        sleep(@fetcher.max_wait_time || 2)
         []
       else
         tag, message = @fetcher.poll
@@ -571,10 +592,34 @@ module Kafka
       end
     end
 
+    def scan_for_subscribing
+      @subscribed_topics.each do |topic_or_regex, config|
+        default_offset = config.fetch(:default_offset)
+        start_from_beginning = config.fetch(:start_from_beginning)
+        max_bytes_per_partition = config.fetch(:max_bytes_per_partition)
+        if topic_or_regex.is_a?(Regexp)
+          subscribe_to_regex(topic_or_regex, default_offset, start_from_beginning, max_bytes_per_partition)
+        else
+          subscribe_to_topic(topic_or_regex, default_offset, start_from_beginning, max_bytes_per_partition)
+        end
+      end
+    end
+
+    def subscribe_to_regex(topic_regex, default_offset, start_from_beginning, max_bytes_per_partition)
+      cluster_topics.select { |topic| topic =~ topic_regex }.each do |topic|
+        subscribe_to_topic(topic, default_offset, start_from_beginning, max_bytes_per_partition)
+      end
+    end
+
     def subscribe_to_topic(topic, default_offset, start_from_beginning, max_bytes_per_partition)
+      return if @matched_topics.include?(topic)
+      @matched_topics.add(topic)
+      @join_group_for_new_topics = true
+
       @group.subscribe(topic)
       @offset_manager.set_default_offset(topic, default_offset)
       @fetcher.subscribe(topic, max_bytes_per_partition: max_bytes_per_partition)
+      @cluster.mark_as_stale!
     end
 
     def cluster_topics

data/lib/kafka/consumer_group.rb

@@ -7,11 +7,12 @@ module Kafka
   class ConsumerGroup
     attr_reader :assigned_partitions, :generation_id, :group_id
 
-    def initialize(cluster:, logger:, group_id:, session_timeout:, retention_time:, instrumenter:)
+    def initialize(cluster:, logger:, group_id:, session_timeout:, rebalance_timeout:, retention_time:, instrumenter:)
       @cluster = cluster
       @logger = TaggedLogger.new(logger)
       @group_id = group_id
       @session_timeout = session_timeout
+      @rebalance_timeout = rebalance_timeout
       @instrumenter = instrumenter
       @member_id = ""
       @generation_id = nil
@@ -140,7 +141,9 @@ module Kafka
         response = coordinator.join_group(
           group_id: @group_id,
           session_timeout: @session_timeout,
+          rebalance_timeout: @rebalance_timeout,
           member_id: @member_id,
+          topics: @topics,
         )
 
         Protocol.handle_error(response.error_code)
@@ -158,6 +161,12 @@ module Kafka
       @member_id = ""
       sleep 1
 
+      retry
+    rescue CoordinatorLoadInProgress
+      @logger.error "Coordinator broker still loading, retrying in 1s..."
+
+      sleep 1
+
       retry
     end
 

data/lib/kafka/datadog.rb

@@ -31,7 +31,7 @@ module Kafka
 
     class << self
       def statsd
-        @statsd ||= ::Datadog::Statsd.new(host, port, namespace: namespace, tags: tags)
+        @statsd ||= ::Datadog::Statsd.new(host, port, namespace: namespace, tags: tags, socket_path: socket_path)
       end
 
       def statsd=(statsd)
@@ -40,7 +40,7 @@ module Kafka
       end
 
       def host
-        @host ||= default_host
+        @host
       end
 
       def host=(host)
@@ -49,7 +49,7 @@ module Kafka
       end
 
       def port
-        @port ||= default_port
+        @port
       end
 
       def port=(port)
@@ -57,6 +57,15 @@ module Kafka
         clear
       end
 
+      def socket_path
+        @socket_path
+      end
+
+      def socket_path=(socket_path)
+        @socket_path = socket_path
+        clear
+      end
+
       def namespace
         @namespace ||= STATSD_NAMESPACE
       end
@@ -77,14 +86,6 @@ module Kafka
 
       private
 
-      def default_host
-        ::Datadog::Statsd.const_defined?(:Connection) ? ::Datadog::Statsd::Connection::DEFAULT_HOST : ::Datadog::Statsd::DEFAULT_HOST
-      end
-
-      def default_port
-        ::Datadog::Statsd.const_defined?(:Connection) ? ::Datadog::Statsd::Connection::DEFAULT_PORT : ::Datadog::Statsd::DEFAULT_PORT
-      end
-
       def clear
         @statsd && @statsd.close
         @statsd = nil
@@ -168,6 +169,8 @@ module Kafka
       def process_batch(event)
         offset = event.payload.fetch(:last_offset)
         messages = event.payload.fetch(:message_count)
+        create_time = event.payload.fetch(:last_create_time)
+        time_lag = create_time && ((Time.now - create_time) * 1000).to_i
 
         tags = {
           client: event.payload.fetch(:client_id),
@@ -184,6 +187,10 @@ module Kafka
         end
 
         gauge("consumer.offset", offset, tags: tags)
+
+        if time_lag
+          gauge("consumer.time_lag", time_lag, tags: tags)
+        end
       end
 
       def fetch_batch(event)

data/lib/kafka/fetched_batch_generator.rb

@@ -48,7 +48,7 @@ module Kafka
               partition: @fetched_partition.partition
             )
           end
-        end
+        end.compact
       end
       FetchedBatch.new(
         topic: @topic,

data/lib/kafka/fetcher.rb

@@ -4,7 +4,7 @@ require "kafka/fetch_operation"
 
 module Kafka
   class Fetcher
-    attr_reader :queue
+    attr_reader :queue, :max_wait_time
 
     def initialize(cluster:, logger:, instrumenter:, max_queue_size:, group:)
       @cluster = cluster
@@ -17,6 +17,9 @@ module Kafka
       @commands = Queue.new
       @next_offsets = Hash.new { |h, k| h[k] = {} }
 
+      # We are only running when someone calls start.
+      @running = false
+
       # Long poll until at least this many bytes can be fetched.
       @min_bytes = 1
 
@@ -110,7 +113,7 @@ module Kafka
       elsif @queue.size < @max_queue_size
         step
       else
-        @logger.warn "Reached max fetcher queue size (#{@max_queue_size}), sleeping 1s"
+        @logger.info "Reached max fetcher queue size (#{@max_queue_size}), sleeping 1s"
         sleep 1
       end
     ensure

data/lib/kafka/offset_manager.rb

@@ -50,9 +50,20 @@ module Kafka
     # @param offset [Integer] the offset of the message that should be marked as processed.
     # @return [nil]
     def mark_as_processed(topic, partition, offset)
-      @uncommitted_offsets += 1
+      unless @group.assigned_to?(topic, partition)
+        @logger.debug "Not marking #{topic}/#{partition}:#{offset} as processed for partition not assigned to this consumer."
+        return
+      end
       @processed_offsets[topic] ||= {}
 
+      last_processed_offset = @processed_offsets[topic][partition] || -1
+      if last_processed_offset > offset + 1
+        @logger.debug "Not overwriting newer offset #{topic}/#{partition}:#{last_processed_offset - 1} with older #{offset}"
+        return
+      end
+
+      @uncommitted_offsets += 1
+
       # The committed offset should always be the offset of the next message that the
       # application will read, thus adding one to the last message processed.
       @processed_offsets[topic][partition] = offset + 1

data/lib/kafka/producer.rb

@@ -188,11 +188,14 @@ module Kafka
     # @raise [BufferOverflow] if the maximum buffer size has been reached.
     # @return [nil]
     def produce(value, key: nil, headers: {}, topic:, partition: nil, partition_key: nil, create_time: Time.now)
+      # We want to fail fast if `topic` isn't a String
+      topic = topic.to_str
+
       message = PendingMessage.new(
         value: value && value.to_s,
         key: key && key.to_s,
         headers: headers,
-        topic: topic.to_s,
+        topic: topic,
         partition: partition && Integer(partition),
         partition_key: partition_key && partition_key.to_s,
         create_time: create_time

data/lib/kafka/prometheus.rb

@@ -0,0 +1,316 @@
+# frozen_string_literal: true
+
+#
+#  Subscriber to ruby_kafka to report metrics to prometheus
+#
+#  Usage:
+#     require "kafka/prometheus"
+#
+#  Once the file has been required, no further configuration is needed, all operational
+#  metrics are automatically emitted (Unless PROMETHEUS_NO_AUTO_START is set).
+#
+#  By Peter Mustel, T2 Data AB
+#
+begin
+  require 'prometheus/client'
+rescue LoadError
+  warn 'In order to report Kafka client metrics to Prometheus you need to install the `prometheus-client` gem.'
+  raise
+end
+
+require 'active_support/subscriber'
+
+module Kafka
+  module Prometheus
+    SIZE_BUCKETS = [1, 10, 100, 1000, 10_000, 100_000, 1_000_000].freeze
+    LATENCY_BUCKETS = [0.0001, 0.001, 0.01, 0.1, 1.0, 10, 100, 1000].freeze
+    DELAY_BUCKETS = [1, 3, 10, 30, 100, 300, 1000, 3000, 10_000, 30_000].freeze
+
+    class << self
+      attr_accessor :registry
+
+      def start(registry = ::Prometheus::Client.registry)
+        @registry = registry
+        ConnectionSubscriber.attach_to 'connection.kafka'
+        ConsumerSubscriber.attach_to 'consumer.kafka'
+        ProducerSubscriber.attach_to 'producer.kafka'
+        AsyncProducerSubscriber.attach_to 'async_producer.kafka'
+        FetcherSubscriber.attach_to 'fetcher.kafka'
+      end
+    end
+
+    class ConnectionSubscriber < ActiveSupport::Subscriber
+      def initialize
+        super
+        @api_calls = Prometheus.registry.counter(:api_calls, docstring: 'Total calls', labels: [:client, :api, :broker])
+        @api_latency = Prometheus.registry.histogram(:api_latency, docstring: 'Latency', buckets: LATENCY_BUCKETS, labels: [:client, :api, :broker])
+        @api_request_size = Prometheus.registry.histogram(:api_request_size, docstring: 'Request size', buckets: SIZE_BUCKETS, labels: [:client, :api, :broker])
+        @api_response_size = Prometheus.registry.histogram(:api_response_size, docstring: 'Response size', buckets: SIZE_BUCKETS, labels: [:client, :api, :broker])
+        @api_errors = Prometheus.registry.counter(:api_errors, docstring: 'Errors', labels: [:client, :api, :broker])
+      end
+
+      def request(event)
+        key = {
+          client: event.payload.fetch(:client_id),
+          api: event.payload.fetch(:api, 'unknown'),
+          broker: event.payload.fetch(:broker_host)
+        }
+        request_size = event.payload.fetch(:request_size, 0)
+        response_size = event.payload.fetch(:response_size, 0)
+
+        @api_calls.increment(labels: key)
+        @api_latency.observe(event.duration, labels: key)
+        @api_request_size.observe(request_size, labels: key)
+        @api_response_size.observe(response_size, labels: key)
+        @api_errors.increment(labels: key) if event.payload.key?(:exception)
+      end
+    end
+
+    class ConsumerSubscriber < ActiveSupport::Subscriber
+      def initialize
+        super
+        @process_messages = Prometheus.registry.counter(:consumer_process_messages, docstring: 'Total messages', labels: [:client, :group_id, :topic, :partition])
+        @process_message_errors = Prometheus.registry.counter(:consumer_process_message_errors, docstring: 'Total errors', labels: [:client, :group_id, :topic, :partition])
+        @process_message_latency =
+          Prometheus.registry.histogram(:consumer_process_message_latency, docstring: 'Latency', buckets: LATENCY_BUCKETS, labels: [:client, :group_id, :topic, :partition])
+        @offset_lag = Prometheus.registry.gauge(:consumer_offset_lag, docstring: 'Offset lag', labels: [:client, :group_id, :topic, :partition])
+        @time_lag = Prometheus.registry.gauge(:consumer_time_lag, docstring: 'Time lag of message', labels: [:client, :group_id, :topic, :partition])
+        @process_batch_errors = Prometheus.registry.counter(:consumer_process_batch_errors, docstring: 'Total errors in batch', labels: [:client, :group_id, :topic, :partition])
+        @process_batch_latency =
+          Prometheus.registry.histogram(:consumer_process_batch_latency, docstring: 'Latency in batch', buckets: LATENCY_BUCKETS, labels: [:client, :group_id, :topic, :partition])
+        @batch_size = Prometheus.registry.histogram(:consumer_batch_size, docstring: 'Size of batch', buckets: SIZE_BUCKETS, labels: [:client, :group_id, :topic, :partition])
+        @join_group = Prometheus.registry.histogram(:consumer_join_group, docstring: 'Time to join group', buckets: DELAY_BUCKETS, labels: [:client, :group_id])
+        @join_group_errors = Prometheus.registry.counter(:consumer_join_group_errors, docstring: 'Total error in joining group', labels: [:client, :group_id])
+        @sync_group = Prometheus.registry.histogram(:consumer_sync_group, docstring: 'Time to sync group', buckets: DELAY_BUCKETS, labels: [:client, :group_id])
+        @sync_group_errors = Prometheus.registry.counter(:consumer_sync_group_errors, docstring: 'Total error in syncing group', labels: [:client, :group_id])
+        @leave_group = Prometheus.registry.histogram(:consumer_leave_group, docstring: 'Time to leave group', buckets: DELAY_BUCKETS, labels: [:client, :group_id])
+        @leave_group_errors = Prometheus.registry.counter(:consumer_leave_group_errors, docstring: 'Total error in leaving group', labels: [:client, :group_id])
+        @pause_duration = Prometheus.registry.gauge(:consumer_pause_duration, docstring: 'Pause duration', labels: [:client, :group_id, :topic, :partition])
+      end
+
+      def process_message(event)
+        key = {
+          client: event.payload.fetch(:client_id),
+          group_id: event.payload.fetch(:group_id),
+          topic: event.payload.fetch(:topic),
+          partition: event.payload.fetch(:partition)
+        }
+
+        offset_lag = event.payload.fetch(:offset_lag)
+        create_time = event.payload.fetch(:create_time)
+
+        time_lag = create_time && ((Time.now - create_time) * 1000).to_i
+
+        if event.payload.key?(:exception)
+          @process_message_errors.increment(labels: key)
+        else
+          @process_message_latency.observe(event.duration, labels: key)
+          @process_messages.increment(labels: key)
+        end
+
+        @offset_lag.set(offset_lag, labels: key)
+
+        # Not all messages have timestamps.
+        return unless time_lag
+
+        @time_lag.set(time_lag, labels: key)
+      end
+
+      def process_batch(event)
+        key = {
+          client: event.payload.fetch(:client_id),
+          group_id: event.payload.fetch(:group_id),
+          topic: event.payload.fetch(:topic),
+          partition: event.payload.fetch(:partition)
+        }
+        message_count = event.payload.fetch(:message_count)
+
+        if event.payload.key?(:exception)
+          @process_batch_errors.increment(labels: key)
+        else
+          @process_batch_latency.observe(event.duration, labels: key)
+          @process_messages.increment(by: message_count, labels: key)
+        end
+      end
+
+      def fetch_batch(event)
+        key = {
+          client: event.payload.fetch(:client_id),
+          group_id: event.payload.fetch(:group_id),
+          topic: event.payload.fetch(:topic),
+          partition: event.payload.fetch(:partition)
+        }
+        offset_lag = event.payload.fetch(:offset_lag)
+        batch_size = event.payload.fetch(:message_count)
+
+        @batch_size.observe(batch_size, labels: key)
+        @offset_lag.set(offset_lag, labels: key)
+      end
+
+      def join_group(event)
+        key = { client: event.payload.fetch(:client_id), group_id: event.payload.fetch(:group_id) }
+        @join_group.observe(event.duration, labels: key)
+
+        @join_group_errors.increment(labels: key) if event.payload.key?(:exception)
+      end
+
+      def sync_group(event)
+        key = { client: event.payload.fetch(:client_id), group_id: event.payload.fetch(:group_id) }
+        @sync_group.observe(event.duration, labels: key)
+
+        @sync_group_errors.increment(labels: key) if event.payload.key?(:exception)
+      end
+
+      def leave_group(event)
+        key = { client: event.payload.fetch(:client_id), group_id: event.payload.fetch(:group_id) }
+        @leave_group.observe(event.duration, labels: key)
+
+        @leave_group_errors.increment(labels: key) if event.payload.key?(:exception)
+      end
+
+      def pause_status(event)
+        key = {
+          client: event.payload.fetch(:client_id),
+          group_id: event.payload.fetch(:group_id),
+          topic: event.payload.fetch(:topic),
+          partition: event.payload.fetch(:partition)
+        }
+
+        duration = event.payload.fetch(:duration)
+        @pause_duration.set(duration, labels: key)
+      end
+    end
+
+    class ProducerSubscriber < ActiveSupport::Subscriber
+      def initialize
+        super
+        @produce_messages = Prometheus.registry.counter(:producer_produced_messages, docstring: 'Produced messages total', labels: [:client, :topic])
+        @produce_message_size =
+          Prometheus.registry.histogram(:producer_message_size, docstring: 'Message size', buckets: SIZE_BUCKETS, labels: [:client, :topic])
+        @buffer_size = Prometheus.registry.histogram(:producer_buffer_size, docstring: 'Buffer size', buckets: SIZE_BUCKETS, labels: [:client])
+        @buffer_fill_ratio = Prometheus.registry.histogram(:producer_buffer_fill_ratio, docstring: 'Buffer fill ratio', labels: [:client])
+        @buffer_fill_percentage = Prometheus.registry.histogram(:producer_buffer_fill_percentage, docstring: 'Buffer fill percentage', labels: [:client])
+        @produce_errors = Prometheus.registry.counter(:producer_produce_errors, docstring: 'Produce errors', labels: [:client, :topic])
+        @deliver_errors = Prometheus.registry.counter(:producer_deliver_errors, docstring: 'Deliver error', labels: [:client])
+        @deliver_latency =
+          Prometheus.registry.histogram(:producer_deliver_latency, docstring: 'Delivery latency', buckets: LATENCY_BUCKETS, labels: [:client])
+        @deliver_messages = Prometheus.registry.counter(:producer_deliver_messages, docstring: 'Total count of delivered messages', labels: [:client])
+        @deliver_attempts = Prometheus.registry.histogram(:producer_deliver_attempts, docstring: 'Delivery attempts', labels: [:client])
+        @ack_messages = Prometheus.registry.counter(:producer_ack_messages, docstring: 'Ack', labels: [:client, :topic])
+        @ack_delay = Prometheus.registry.histogram(:producer_ack_delay, docstring: 'Ack delay', buckets: LATENCY_BUCKETS, labels: [:client, :topic])
+        @ack_errors = Prometheus.registry.counter(:producer_ack_errors, docstring: 'Ack errors', labels: [:client, :topic])
+      end
+
+      def produce_message(event)
+        client = event.payload.fetch(:client_id)
+        key = { client: client, topic: event.payload.fetch(:topic) }
+
+        message_size = event.payload.fetch(:message_size)
+        buffer_size = event.payload.fetch(:buffer_size)
+        max_buffer_size = event.payload.fetch(:max_buffer_size)
+        buffer_fill_ratio = buffer_size.to_f / max_buffer_size.to_f
+        buffer_fill_percentage = buffer_fill_ratio * 100.0
+
+        # This gets us the write rate.
+        @produce_messages.increment(labels: key)
+        @produce_message_size.observe(message_size, labels: key)
+
+        # This gets us the avg/max buffer size per producer.
+        @buffer_size.observe(buffer_size, labels: { client: client })
+
+        # This gets us the avg/max buffer fill ratio per producer.
+        @buffer_fill_ratio.observe(buffer_fill_ratio, labels: { client: client })
+        @buffer_fill_percentage.observe(buffer_fill_percentage, labels: { client: client })
+      end
+
+      def buffer_overflow(event)
+        key = { client: event.payload.fetch(:client_id), topic: event.payload.fetch(:topic) }
+        @produce_errors.increment(labels: key)
+      end
+
+      def deliver_messages(event)
+        key = { client: event.payload.fetch(:client_id) }
+        message_count = event.payload.fetch(:delivered_message_count)
+        attempts = event.payload.fetch(:attempts)
+
+        @deliver_errors.increment(labels: key) if event.payload.key?(:exception)
+        @deliver_latency.observe(event.duration, labels: key)
+
+        # Messages delivered to Kafka:
+        @deliver_messages.increment(by: message_count, labels: key)
+
+        # Number of attempts to deliver messages:
+        @deliver_attempts.observe(attempts, labels: key)
+      end
+
+      def ack_message(event)
+        key = { client: event.payload.fetch(:client_id), topic: event.payload.fetch(:topic) }
+
+        # Number of messages ACK'd for the topic.
+        @ack_messages.increment(labels: key)
+
+        # Histogram of delay between a message being produced and it being ACK'd.
+        @ack_delay.observe(event.payload.fetch(:delay), labels: key)
+      end
+
+      def topic_error(event)
+        key = { client: event.payload.fetch(:client_id), topic: event.payload.fetch(:topic) }
+
+        @ack_errors.increment(labels: key)
+      end
+    end
+
+    class AsyncProducerSubscriber < ActiveSupport::Subscriber
+      def initialize
+        super
+        @queue_size = Prometheus.registry.histogram(:async_producer_queue_size, docstring: 'Queue size', buckets: SIZE_BUCKETS, labels: [:client, :topic])
+        @queue_fill_ratio = Prometheus.registry.histogram(:async_producer_queue_fill_ratio, docstring: 'Queue fill ratio', labels: [:client, :topic])
+        @produce_errors = Prometheus.registry.counter(:async_producer_produce_errors, docstring: 'Producer errors', labels: [:client, :topic])
+        @dropped_messages = Prometheus.registry.counter(:async_producer_dropped_messages, docstring: 'Dropped messages', labels: [:client])
+      end
+
+      def enqueue_message(event)
+        key = { client: event.payload.fetch(:client_id), topic: event.payload.fetch(:topic) }
+
+        queue_size = event.payload.fetch(:queue_size)
+        max_queue_size = event.payload.fetch(:max_queue_size)
+        queue_fill_ratio = queue_size.to_f / max_queue_size.to_f
+
+        # This gets us the avg/max queue size per producer.
+        @queue_size.observe(queue_size, labels: key)
+
+        # This gets us the avg/max queue fill ratio per producer.
+        @queue_fill_ratio.observe(queue_fill_ratio, labels: key)
+      end
+
+      def buffer_overflow(event)
+        key = { client: event.payload.fetch(:client_id), topic: event.payload.fetch(:topic) }
+        @produce_errors.increment(labels: key)
+      end
+
+      def drop_messages(event)
+        key = { client: event.payload.fetch(:client_id) }
+        message_count = event.payload.fetch(:message_count)
+        @dropped_messages.increment(by: message_count, labels: key)
+      end
+    end
+
+    class FetcherSubscriber < ActiveSupport::Subscriber
+      def initialize
+        super
+        @queue_size = Prometheus.registry.gauge(:fetcher_queue_size, docstring: 'Queue size', labels: [:client, :group_id])
+      end
+
+      def loop(event)
+        queue_size = event.payload.fetch(:queue_size)
+        client = event.payload.fetch(:client_id)
+        group_id = event.payload.fetch(:group_id)
+
+        @queue_size.set(queue_size, labels: { client: client, group_id: group_id })
+      end
+    end
+  end
+end
+
+# To enable testability, it is possible to skip the start until test time
+Kafka::Prometheus.start unless defined?(PROMETHEUS_NO_AUTO_START)