karafka 2.0.10 → 2.0.12

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: f78f7cb985880d9172961be96386a0ebd37735831915f0cb9b9b46c832d2e9a9
-  data.tar.gz: a7f5a27cb3a6f0fa5185f32e5df7615f8a6013ef0018c85fb16c87de346f9362
+  metadata.gz: 2c407cd113e41314102fc910cf0b35e8081c55bdadae66635afe59491b58e390
+  data.tar.gz: 336dad28cc65218e33a2b3bc42b69a166ff4e8e6a10c65a435e57f75eca5ac90
 SHA512:
-  metadata.gz: 5b6fed517f69a2bd84b16e82d266251dd4b7564d52f466745f3a93be9106f866ec658581c38b1a0913dd9e126a17cf4b1f5e85c888447c6e1e4b790453e7c87e
-  data.tar.gz: 25cfa154028dd12519d7d4b72606242a4058c9fabd1655ba727a9d73a905a942cd80374e2f7357f410a098b74b623966fa840b1c6725f71d630bf3ce92db187e
+  metadata.gz: 82ffee28acdaa1d126944426a7b31628380fc04aca24002ea55067f6c80a0b72a885d246a656913ea35e79fa82a9d741c35125e31b75abbbf7fb0022360e3b2b
+  data.tar.gz: b64f9dc4fb2461f3cbe3a774231a661365af5b1be88a6c95b1543521d19d8171514eb5111f88ceb53e309b5ab433f982b9535c2bbb4919e4c07850570424605a

data/.github/workflows/ci.yml

@@ -52,9 +52,14 @@ jobs:
       fail-fast: false
       matrix:
         ruby:
+          # We run it against the oldest and the newest of a given major to make sure, that there
+          # are no syntax-sugars that we would use that were introduced down the road
           - '3.1'
+          - '3.1.0'
           - '3.0'
+          - '3.0.0'
           - '2.7'
+          - '2.7.0'
         include:
           - ruby: '3.1'
             coverage: 'true'

data/CHANGELOG.md

@@ -1,5 +1,15 @@
 # Karafka framework changelog
 
+## 2.0.12 (2022-10-06)
+- Commit stored offsets upon rebalance revocation event to reduce number of messages that are re-processed.
+- Support cooperative-sticky rebalance strategy.
+- Replace offset commit after each batch with a per-rebalance commit.
+- User instrumentation to publish internal rebalance errors.
+
+## 2.0.11 (2022-09-29)
+- Report early on errors related to network and on max poll interval being exceeded to indicate critical problems that will be retries but may mean some underlying problems in the system.
+- Fix support of Ruby 2.7.0 to 2.7.2 (#1045)
+
 ## 2.0.10 (2022-09-23)
 - Improve error recovery by delegating the recovery to the existing `librdkafka` instance.
 

data/Gemfile.lock

@@ -1,7 +1,7 @@
 PATH
   remote: .
   specs:
-    karafka (2.0.10)
+    karafka (2.0.12)
       karafka-core (>= 2.0.2, < 3.0.0)
       rdkafka (>= 0.12)
       thor (>= 0.20)
@@ -61,14 +61,13 @@ GEM
     thor (1.2.1)
     tzinfo (2.0.5)
       concurrent-ruby (~> 1.0)
-    waterdrop (2.4.1)
+    waterdrop (2.4.2)
       karafka-core (>= 2.0.2, < 3.0.0)
       rdkafka (>= 0.10)
       zeitwerk (~> 2.3)
-    zeitwerk (2.6.0)
+    zeitwerk (2.6.1)
 
 PLATFORMS
-  arm64-darwin
   x86_64-linux
 
 DEPENDENCIES

data/lib/karafka/connection/client.rb

@@ -188,6 +188,7 @@ module Karafka
         tpl = topic_partition_list(topic, partition) || @paused_tpls[topic][partition]
 
         return unless tpl
+
         # If we did not have it, it means we never paused this partition, thus no resume should
         # happen in the first place
         return unless @paused_tpls[topic].delete(partition)
@@ -317,7 +318,7 @@ module Karafka
         Rdkafka::Consumer::TopicPartitionList.new({ topic => [rdkafka_partition] })
       end
 
-      # Performs a single poll operation.
+      # Performs a single poll operation and handles retries and error
       #
       # @param timeout [Integer] timeout for a single poll
       # @return [Rdkafka::Consumer::Message, nil] fetched message or nil if nothing polled
@@ -330,20 +331,39 @@ module Karafka
 
         @kafka.poll(timeout)
       rescue ::Rdkafka::RdkafkaError => e
-        # Most of the errors can be safely ignored as librdkafka will recover from them
-        # @see https://github.com/edenhill/librdkafka/issues/1987#issuecomment-422008750
-        # @see https://github.com/edenhill/librdkafka/wiki/Error-handling
-        if time_poll.attempts > MAX_POLL_RETRIES || !time_poll.retryable?
+        early_report = false
+
+        # There are retryable issues on which we want to report fast as they are source of
+        # problems and can mean some bigger system instabilities
+        # Those are mainly network issues and exceeding the max poll interval
+        # We want to report early on max poll interval exceeding because it may mean that the
+        # underlying processing is taking too much time and it is not LRJ
+        case e.code
+        when :max_poll_exceeded # -147
+          early_report = true
+        when :network_exception # 13
+          early_report = true
+        when :transport # -195
+          early_report = true
+        end
+
+        retryable = time_poll.attempts <= MAX_POLL_RETRIES && time_poll.retryable?
+
+        if early_report || !retryable
           Karafka.monitor.instrument(
             'error.occurred',
             caller: self,
             error: e,
             type: 'connection.client.poll.error'
           )
-
-          raise
         end
 
+        raise unless retryable
+
+        # Most of the errors can be safely ignored as librdkafka will recover from them
+        # @see https://github.com/edenhill/librdkafka/issues/1987#issuecomment-422008750
+        # @see https://github.com/edenhill/librdkafka/wiki/Error-handling
+
         time_poll.checkpoint
         time_poll.backoff
 

data/lib/karafka/connection/listener.rb

@@ -118,10 +118,6 @@ module Karafka
           build_and_schedule_consumption_jobs
 
           wait
-
-          # We don't use the `#commit_offsets!` here for performance reasons. This can be achieved
-          # if needed by using manual offset management.
-          @client.commit_offsets
         end
 
         # If we are stopping we will no longer schedule any jobs despite polling.

data/lib/karafka/connection/rebalance_manager.rb

@@ -14,6 +14,9 @@ module Karafka
     #   that are lost, are those that got revoked but did not get re-assigned back. We do not
     #   expose this concept outside and we normalize to have them revoked, as it is irrelevant
     #   from the rest of the code perspective as only those that are lost are truly revoked.
+    #
+    # @note For cooperative-sticky `#assigned_partitions` holds only the recently assigned
+    #   partitions, not all the partitions that are owned
     class RebalanceManager
       # Empty array for internal usage not to create new objects
       EMPTY_ARRAY = [].freeze

data/lib/karafka/patches/rdkafka/bindings.rb

@@ -0,0 +1,132 @@
+# frozen_string_literal: true
+
+module Karafka
+  module Patches
+    module Rdkafka
+      # Binding patches that slightly change how rdkafka operates in certain places
+      module Bindings
+        include ::Rdkafka::Bindings
+
+        # Alias internally
+        RB = ::Rdkafka::Bindings
+
+        class << self
+          # Handle assignments on cooperative rebalance
+          #
+          # @param client_ptr [FFI::Pointer]
+          # @param code [Integer]
+          # @param partitions_ptr [FFI::Pointer]
+          def on_cooperative_rebalance(client_ptr, code, partitions_ptr)
+            case code
+            when RB::RD_KAFKA_RESP_ERR__ASSIGN_PARTITIONS
+              RB.rd_kafka_incremental_assign(client_ptr, partitions_ptr)
+            when RB::RD_KAFKA_RESP_ERR__REVOKE_PARTITIONS
+              RB.rd_kafka_commit(client_ptr, nil, false)
+              RB.rd_kafka_incremental_unassign(client_ptr, partitions_ptr)
+            else
+              RB.rd_kafka_assign(client_ptr, FFI::Pointer::NULL)
+            end
+          end
+
+          # Handle assignments on a eager rebalance
+          #
+          # @param client_ptr [FFI::Pointer]
+          # @param code [Integer]
+          # @param partitions_ptr [FFI::Pointer]
+          def on_eager_rebalance(client_ptr, code, partitions_ptr)
+            case code
+            when RB::RD_KAFKA_RESP_ERR__ASSIGN_PARTITIONS
+              RB.rd_kafka_assign(client_ptr, partitions_ptr)
+            when RB::RD_KAFKA_RESP_ERR__REVOKE_PARTITIONS
+              RB.rd_kafka_commit(client_ptr, nil, false)
+              RB.rd_kafka_assign(client_ptr, FFI::Pointer::NULL)
+            else
+              RB.rd_kafka_assign(client_ptr, FFI::Pointer::NULL)
+            end
+          end
+
+          # Trigger Karafka callbacks
+          #
+          # @param code [Integer]
+          # @param opaque [Rdkafka::Opaque]
+          # @param consumer [Rdkafka::Consumer]
+          # @param tpl [Rdkafka::Consumer::TopicPartitionList]
+          def trigger_callbacks(code, opaque, consumer, tpl)
+            case code
+            when RB::RD_KAFKA_RESP_ERR__ASSIGN_PARTITIONS
+              opaque.call_on_partitions_assigned(consumer, tpl)
+            when RB::RD_KAFKA_RESP_ERR__REVOKE_PARTITIONS
+              opaque.call_on_partitions_revoked(consumer, tpl)
+            end
+          rescue StandardError => e
+            Karafka.monitor.instrument(
+              'error.occurred',
+              caller: self,
+              error: e,
+              type: 'connection.client.rebalance_callback.error'
+            )
+          end
+        end
+
+        # This patch changes few things:
+        # - it commits offsets (if any) upon partition revocation, so less jobs need to be
+        #   reprocessed if they are assigned to a different process
+        # - reports callback errors into the errors instrumentation instead of the logger
+        # - catches only StandardError instead of Exception as we fully control the directly
+        #   executed callbacks
+        #
+        # @see https://docs.confluent.io/2.0.0/clients/librdkafka/classRdKafka_1_1RebalanceCb.html
+        RebalanceCallback = FFI::Function.new(
+          :void, %i[pointer int pointer pointer]
+        ) do |client_ptr, code, partitions_ptr, opaque_ptr|
+          # Patch reference
+          pr = ::Karafka::Patches::Rdkafka::Bindings
+
+          if RB.rd_kafka_rebalance_protocol(client_ptr) == 'COOPERATIVE'
+            pr.on_cooperative_rebalance(client_ptr, code, partitions_ptr)
+          else
+            pr.on_eager_rebalance(client_ptr, code, partitions_ptr)
+          end
+
+          opaque = ::Rdkafka::Config.opaques[opaque_ptr.to_i]
+          return unless opaque
+
+          tpl = ::Rdkafka::Consumer::TopicPartitionList.from_native_tpl(partitions_ptr).freeze
+          consumer = ::Rdkafka::Consumer.new(client_ptr)
+
+          pr.trigger_callbacks(code, opaque, consumer, tpl)
+        end
+      end
+    end
+  end
+end
+
+# We need to replace the original callback with ours.
+# At the moment there is no API in rdkafka-ruby to do so
+::Rdkafka::Bindings.send(
+  :remove_const,
+  'RebalanceCallback'
+)
+
+::Rdkafka::Bindings.const_set(
+  'RebalanceCallback',
+  Karafka::Patches::Rdkafka::Bindings::RebalanceCallback
+)
+
+::Rdkafka::Bindings.attach_function(
+  :rd_kafka_rebalance_protocol,
+  %i[pointer],
+  :string
+)
+
+::Rdkafka::Bindings.attach_function(
+  :rd_kafka_incremental_assign,
+  %i[pointer pointer],
+  :string
+)
+
+::Rdkafka::Bindings.attach_function(
+  :rd_kafka_incremental_unassign,
+  %i[pointer pointer],
+  :string
+)

data/lib/karafka/routing/consumer_group.rb

@@ -21,7 +21,7 @@ module Karafka
       #   We add it to make a multi-system development easier for people that don't use
       #   kafka and don't understand the concept of consumer groups.
       def initialize(name)
-        @name = name
+        @name = name.to_s
         @id = Karafka::App.config.consumer_mapper.call(name)
         @topics = Topics.new([])
       end
@@ -52,7 +52,7 @@ module Karafka
       def subscription_group=(name, &block)
         # We cast it here, so the routing supports symbol based but that's anyhow later on
         # validated as a string
-        self.current_subscription_group_name = name.to_s
+        self.current_subscription_group_name = name
 
         Proxy.new(self, &block)
 

data/lib/karafka/routing/proxy.rb

@@ -14,17 +14,22 @@ module Karafka
         instance_eval(&block)
       end
 
-      # Translates the no "=" DSL of routing into elements assignments on target
-      # @param method_name [Symbol] name of the missing method
-      def method_missing(method_name, ...)
-        return super unless respond_to_missing?(method_name)
+      # Ruby 2.7.0 to 2.7.2 do not have arg forwarding, so we fallback to the old way
+      arg_forwarding = RUBY_VERSION < '3.0' ? '*args, &block' : '...'
+
+      class_eval <<~RUBY, __FILE__, __LINE__ + 1
+        # Translates the no "=" DSL of routing into elements assignments on target
+        # @param method_name [Symbol] name of the missing method
+        def method_missing(method_name, #{arg_forwarding})
+          return super unless respond_to_missing?(method_name)
 
-        if @target.respond_to?(:"#{method_name}=")
-          @target.public_send(:"#{method_name}=", ...)
-        else
-          @target.public_send(method_name, ...)
+          if @target.respond_to?(:"\#{method_name}=")
+            @target.public_send(:"\#{method_name}=", #{arg_forwarding})
+          else
+            @target.public_send(method_name, #{arg_forwarding})
+          end
         end
-      end
+      RUBY
 
       # Tells whether or not a given element exists on the target
       # @param method_name [Symbol] name of the missing method

data/lib/karafka/serialization/json/deserializer.rb

@@ -10,6 +10,7 @@ module Karafka
         # @param message [Karafka::Messages::Message] Message object that we want to deserialize
         # @return [Hash] hash with deserialized JSON data
         def call(message)
+          # nil payload can be present for example for tombstone messages
           message.raw_payload.nil? ? nil : ::JSON.parse(message.raw_payload)
         end
       end

data/lib/karafka/version.rb

@@ -3,5 +3,5 @@
 # Main module namespace
 module Karafka
   # Current Karafka version
-  VERSION = '2.0.10'
+  VERSION = '2.0.12'
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: karafka
 version: !ruby/object:Gem::Version
-  version: 2.0.10
+  version: 2.0.12
 platform: ruby
 authors:
 - Maciej Mensfeld
@@ -35,7 +35,7 @@ cert_chain:
   Qf04B9ceLUaC4fPVEz10FyobjaFoY4i32xRto3XnrzeAgfEe4swLq8bQsR3w/EF3
   MGU0FeSV2Yj7Xc2x/7BzLK8xQn5l7Yy75iPF+KP3vVmDHnNl
   -----END CERTIFICATE-----
-date: 2022-09-23 00:00:00.000000000 Z
+date: 2022-10-06 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: karafka-core
@@ -212,6 +212,7 @@ files:
 - lib/karafka/messages/messages.rb
 - lib/karafka/messages/metadata.rb
 - lib/karafka/messages/seek.rb
+- lib/karafka/patches/rdkafka/bindings.rb
 - lib/karafka/patches/rdkafka/consumer.rb
 - lib/karafka/pro.rb
 - lib/karafka/pro/active_job/consumer.rb