karafka 1.4.13 → 2.0.0

data/lib/karafka/cli/server.rb

@@ -5,27 +5,18 @@ module Karafka
   class Cli < Thor
     # Server Karafka Cli action
     class Server < Base
-      # Server config settings contract
-      CONTRACT = Contracts::ServerCliOptions.new.freeze
-
-      private_constant :CONTRACT
+      include Helpers::Colorize
 
       desc 'Start the Karafka server (short-cut alias: "s")'
       option aliases: 's'
-      option :daemon, default: false, type: :boolean, aliases: :d
-      option :pid, default: 'tmp/pids/karafka', type: :string, aliases: :p
       option :consumer_groups, type: :array, default: nil, aliases: :g
 
       # Start the Karafka server
       def call
-        cli.info
-
-        validate!
+        # Print our banner and info in the dev mode
+        print_marketing_info if Karafka::App.env.development?
 
-        if cli.options[:daemon]
-          FileUtils.mkdir_p File.dirname(cli.options[:pid])
-          daemonize
-        end
+        Contracts::ServerCliOptions.new.validate!(cli.options)
 
         # We assign active topics on a server level, as only server is expected to listen on
         # part of the topics
@@ -36,35 +27,19 @@ module Karafka
 
       private
 
-      # Checks the server cli configuration
-      # options validations in terms of app setup (topics, pid existence, etc)
-      def validate!
-        result = CONTRACT.call(cli.options)
-        return if result.success?
-
-        raise Errors::InvalidConfigurationError, result.errors.to_h
-      end
-
-      # Detaches current process into background and writes its pidfile
-      def daemonize
-        ::Process.daemon(true)
-        File.open(
-          cli.options[:pid],
-          'w'
-        ) { |file| file.write(::Process.pid) }
-
-        # Remove pidfile on stop, just before the server instance is going to be GCed
-        # We want to delay the moment in which the pidfile is removed as much as we can,
-        # so instead of removing it after the server stops running, we rely on the gc moment
-        # when this object gets removed (it is a bit later), so it is closer to the actual
-        # system process end. We do that, so monitoring and deployment tools that rely on a pid
-        # won't alarm or start new system process up until the current one is finished
-        ObjectSpace.define_finalizer(self, proc { send(:clean) })
-      end
-
-      # Removes a pidfile (if exist)
-      def clean
-        FileUtils.rm_f(cli.options[:pid]) if cli.options[:pid]
+      # Prints marketing info
+      def print_marketing_info
+        Karafka.logger.info Info::BANNER
+
+        if Karafka.pro?
+          Karafka.logger.info(
+            green('Thank you for investing in the Karafka Pro subscription!')
+          )
+        else
+          Karafka.logger.info(
+            red('You like Karafka? Please consider getting a Pro version!')
+          )
+        end
       end
     end
   end

data/lib/karafka/cli.rb

@@ -2,6 +2,7 @@
 
 module Karafka
   # Karafka framework Cli
+  #
   # If you want to add/modify command that belongs to CLI, please review all commands
   # available in cli/ directory inside Karafka source code.
   #
@@ -10,24 +11,16 @@ module Karafka
   class Cli < Thor
     package_name 'Karafka'
 
-    default_task :missingno
-
     class << self
-      # Loads all Cli commands into Thor framework
+      # Loads all Cli commands into Thor framework.
       # This method should be executed before we run Karafka::Cli.start, otherwise we won't
-      # have any Cli commands available
+      # have any Cli commands available.
       def prepare
         cli_commands.each do |action|
           action.bind_to(self)
         end
       end
 
-      # When there is a CLI crash, exit
-      # @return [true]
-      def exit_on_failure?
-        true
-      end
-
       private
 
       # @return [Array<Class>] Array with Cli action classes that can be used as commands
@@ -42,7 +35,7 @@ module Karafka
   end
 end
 
-# This is kinda trick - since we don't have a autoload and other magic stuff
+# This is kinda tricky - since we don't have an autoload and other magic stuff
 # like Rails does, so instead this method allows us to replace currently running
 # console with a new one via Kernel.exec. It will start console with new code loaded
 # Yes, we know that it is not turbo fast, however it is turbo convenient and small

data/lib/karafka/connection/client.rb

@@ -1,119 +1,414 @@
 # frozen_string_literal: true
 
 module Karafka
+  # Namespace for Kafka connection related logic
   module Connection
-    # Class used as a wrapper around Ruby-Kafka client to simplify additional
-    # features that we provide/might provide in future and to hide the internal implementation
+    # An abstraction layer on top of the rdkafka consumer.
+    #
+    # It is threadsafe and provides some security measures so we won't end up operating on a
+    # closed consumer instance as it causes Ruby VM process to crash.
     class Client
-      extend Forwardable
-
-      %i[
-        seek
-        trigger_heartbeat
-        trigger_heartbeat!
-      ].each do |delegated_method|
-        def_delegator :kafka_consumer, delegated_method
-      end
-
-      # Creates a queue consumer client that will pull the data from Kafka
-      # @param consumer_group [Karafka::Routing::ConsumerGroup] consumer group for which
-      #   we create a client
-      # @return [Karafka::Connection::Client] group consumer that can subscribe to
-      #   multiple topics
-      def initialize(consumer_group)
-        @consumer_group = consumer_group
-        Persistence::Client.write(self)
-      end
-
-      # Opens connection, gets messages and calls a block for each of the incoming messages
-      # @yieldparam [Array<Kafka::FetchedMessage>, Symbol] kafka response with an info about
-      #   the type of the fetcher that is being used
-      # @note This will yield with raw messages - no preprocessing or reformatting.
-      def fetch_loop
-        settings = ApiAdapter.consumption(consumer_group)
-
-        if consumer_group.batch_fetching
-          kafka_consumer.each_batch(**settings) { |batch| yield(batch, :batch) }
-        else
-          kafka_consumer.each_message(**settings) { |message| yield(message, :message) }
+      attr_reader :rebalance_manager
+
+      # @return [String] underlying consumer name
+      # @note Consumer name may change in case we regenerate it
+      attr_reader :name
+
+      # How many times should we retry polling in case of a failure
+      MAX_POLL_RETRIES = 10
+
+      private_constant :MAX_POLL_RETRIES
+
+      # Creates a new consumer instance.
+      #
+      # @param subscription_group [Karafka::Routing::SubscriptionGroup] subscription group
+      #   with all the configuration details needed for us to create a client
+      # @return [Karafka::Connection::Rdk::Consumer]
+      def initialize(subscription_group)
+        # Name is set when we build consumer
+        @name = ''
+        @mutex = Mutex.new
+        @closed = false
+        @subscription_group = subscription_group
+        @buffer = RawMessagesBuffer.new
+        @rebalance_manager = RebalanceManager.new
+        @kafka = build_consumer
+        # Marks if we need to offset. If we did not store offsets, we should not commit the offset
+        # position as it will crash rdkafka
+        @offsetting = false
+        # We need to keep track of what we have paused for resuming
+        # In case we loose partition, we still need to resume it, otherwise it won't be fetched
+        # again if we get reassigned to it later on. We need to keep them as after revocation we
+        # no longer may be able to fetch them from Kafka. We could build them but it is easier
+        # to just keep them here and use if needed when cannot be obtained
+        @paused_tpls = Hash.new { |h, k| h[k] = {} }
+      end
+
+      # Fetches messages within boundaries defined by the settings (time, size, topics, etc).
+      #
+      # @return [Karafka::Connection::MessagesBuffer] messages buffer that holds messages per topic
+      #   partition
+      # @note This method should not be executed from many threads at the same time
+      def batch_poll
+        time_poll = TimeTrackers::Poll.new(@subscription_group.max_wait_time)
+
+        @buffer.clear
+        @rebalance_manager.clear
+
+        loop do
+          time_poll.start
+
+          # Don't fetch more messages if we do not have any time left
+          break if time_poll.exceeded?
+          # Don't fetch more messages if we've fetched max as we've wanted
+          break if @buffer.size >= @subscription_group.max_messages
+
+          # Fetch message within our time boundaries
+          message = poll(time_poll.remaining)
+
+          # Put a message to the buffer if there is one
+          @buffer << message if message
+
+          # Upon polling rebalance manager might have been updated.
+          # If partition revocation happens, we need to remove messages from revoked partitions
+          # as well as ensure we do not have duplicated due to the offset reset for partitions
+          # that we got assigned
+          # We also do early break, so the information about rebalance is used as soon as possible
+          if @rebalance_manager.changed?
+            remove_revoked_and_duplicated_messages
+            break
+          end
+
+          # Track time spent on all of the processing and polling
+          time_poll.checkpoint
+
+          # Finally once we've (potentially) removed revoked, etc, if no messages were returned
+          # we can break.
+          # Worth keeping in mind, that the rebalance manager might have been updated despite no
+          # messages being returned during a poll
+          break unless message
         end
-      # @note We catch only the processing errors as any other are considered critical (exceptions)
-      #   and should require a client restart with a backoff
-      rescue Kafka::ProcessingError => e
-        # If there was an error during consumption, we have to log it, pause current partition
-        # and process other things
-        Karafka.monitor.instrument(
-          'connection.client.fetch_loop.error',
-          caller: self,
-          error: e.cause
-        )
-        pause(e.topic, e.partition)
-        retry
+
+        @buffer
       end
 
-      # Gracefully stops topic consumption
+      # Stores offset for a given partition of a given topic based on the provided message.
+      #
+      # @param message [Karafka::Messages::Message]
+      def store_offset(message)
+        @mutex.synchronize do
+          internal_store_offset(message)
+        end
+      end
+
+      # Commits the offset on a current consumer in a non-blocking or blocking way.
+      # Ignoring a case where there would not be an offset (for example when rebalance occurs).
+      #
+      # @param async [Boolean] should the commit happen async or sync (async by default)
+      # @return [Boolean] did committing was successful. It may be not, when we no longer own
+      #   given partition.
+      #
+      # @note This will commit all the offsets for the whole consumer. In order to achieve
+      #   granular control over where the offset should be for particular topic partitions, the
+      #   store_offset should be used to only store new offset when we want to to be flushed
+      def commit_offsets(async: true)
+        @mutex.lock
+
+        internal_commit_offsets(async: async)
+      ensure
+        @mutex.unlock
+      end
+
+      # Commits offset in a synchronous way.
+      #
+      # @see `#commit_offset` for more details
+      def commit_offsets!
+        commit_offsets(async: false)
+      end
+
+      # Seek to a particular message. The next poll on the topic/partition will return the
+      # message at the given offset.
+      #
+      # @param message [Messages::Message, Messages::Seek] message to which we want to seek to
+      def seek(message)
+        @mutex.lock
+
+        @kafka.seek(message)
+      ensure
+        @mutex.unlock
+      end
+
+      # Pauses given partition and moves back to last successful offset processed.
+      #
+      # @param topic [String] topic name
+      # @param partition [Integer] partition
+      # @param offset [Integer] offset of the message on which we want to pause (this message will
+      #   be reprocessed after getting back to processing)
+      # @note This will pause indefinitely and requires manual `#resume`
+      def pause(topic, partition, offset)
+        @mutex.lock
+
+        # Do not pause if the client got closed, would not change anything
+        return if @closed
+
+        pause_msg = Messages::Seek.new(topic, partition, offset)
+
+        internal_commit_offsets(async: false)
+
+        # Here we do not use our cached tpls because we should not try to pause something we do
+        # not own anymore.
+        tpl = topic_partition_list(topic, partition)
+
+        return unless tpl
+
+        @paused_tpls[topic][partition] = tpl
+
+        @kafka.pause(tpl)
+
+        @kafka.seek(pause_msg)
+      ensure
+        @mutex.unlock
+      end
+
+      # Resumes processing of a give topic partition after it was paused.
+      #
+      # @param topic [String] topic name
+      # @param partition [Integer] partition
+      def resume(topic, partition)
+        @mutex.lock
+
+        return if @closed
+
+        # Always commit synchronously offsets if any when we resume
+        # This prevents resuming without offset in case it would not be committed prior
+        # We can skip performance penalty since resuming should not happen too often
+        internal_commit_offsets(async: false)
+
+        # If we were not able, let's try to reuse the one we have (if we have)
+        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)
+
+        @kafka.resume(tpl)
+      ensure
+        @mutex.unlock
+      end
+
+      # Gracefully stops topic consumption.
+      #
       # @note Stopping running consumers without a really important reason is not recommended
       #   as until all the consumers are stopped, the server will keep running serving only
       #   part of the messages
       def stop
-        @kafka_consumer&.stop
-        @kafka_consumer = nil
+        close
       end
 
-      # Pauses fetching and consumption of a given topic partition
-      # @param topic [String] topic that we want to pause
-      # @param partition [Integer] number partition that we want to pause
-      def pause(topic, partition)
-        args, kwargs = ApiAdapter.pause(topic, partition, consumer_group).values_at(:args, :kwargs)
-        kafka_consumer.pause(*args, **kwargs)
+      # Marks given message as consumed.
+      #
+      # @param [Karafka::Messages::Message] message that we want to mark as processed
+      # @return [Boolean] true if successful. False if we no longer own given partition
+      # @note This method won't trigger automatic offsets commits, rather relying on the offset
+      #   check-pointing trigger that happens with each batch processed
+      def mark_as_consumed(message)
+        store_offset(message)
       end
 
-      # Marks given message as consumed
-      # @param [Karafka::Params::Params] params message that we want to mark as processed
-      # @note This method won't trigger automatic offsets commits, rather relying on the ruby-kafka
-      #   offsets time-interval based committing
-      def mark_as_consumed(params)
-        kafka_consumer.mark_message_as_processed(
-          *ApiAdapter.mark_message_as_processed(params)
-        )
+      # Marks a given message as consumed and commits the offsets in a blocking way.
+      #
+      # @param [Karafka::Messages::Message] message that we want to mark as processed
+      # @return [Boolean] true if successful. False if we no longer own given partition
+      def mark_as_consumed!(message)
+        return false unless mark_as_consumed(message)
+
+        commit_offsets!
       end
 
-      # Marks a given message as consumed and commit the offsets in a blocking way
-      # @param [Karafka::Params::Params] params message that we want to mark as processed
-      # @note This method commits the offset for each manual marking to be sure
-      #   that offset commit happen asap in case of a crash
-      def mark_as_consumed!(params)
-        mark_as_consumed(params)
-        # Trigger an immediate, blocking offset commit in order to minimize the risk of crashing
-        # before the automatic triggers have kicked in.
-        kafka_consumer.commit_offsets
+      # Closes and resets the client completely.
+      def reset
+        close
+
+        @mutex.synchronize do
+          @closed = false
+          @offsetting = false
+          @paused_tpls.clear
+          @kafka = build_consumer
+        end
       end
 
       private
 
-      attr_reader :consumer_group
+      # When we cannot store an offset, it means we no longer own the partition
+      #
+      # Non thread-safe offset storing method
+      # @param message [Karafka::Messages::Message]
+      # @return [Boolean] true if we could store the offset (if we still own the partition)
+      def internal_store_offset(message)
+        @offsetting = true
+        @kafka.store_offset(message)
+        true
+      rescue Rdkafka::RdkafkaError => e
+        return false if e.code == :assignment_lost
+        return false if e.code == :state
+
+        raise e
+      end
+
+      # Non thread-safe message committing method
+      # @param async [Boolean] should the commit happen async or sync (async by default)
+      # @return [Boolean] true if offset commit worked, false if we've lost the assignment
+      def internal_commit_offsets(async: true)
+        return true unless @offsetting
+
+        @kafka.commit(nil, async)
+        @offsetting = false
+
+        true
+      rescue Rdkafka::RdkafkaError => e
+        return false if e.code == :assignment_lost
+        return true if e.code == :no_offset
+
+        raise e
+      end
+
+      # Commits the stored offsets in a sync way and closes the consumer.
+      def close
+        # Once client is closed, we should not close it again
+        # This could only happen in case of a race-condition when forceful shutdown happens
+        # and triggers this from a different thread
+        return if @closed
+
+        @mutex.synchronize do
+          internal_commit_offsets(async: false)
+
+          @closed = true
+
+          # Remove callbacks runners that were registered
+          ::Karafka::Instrumentation.statistics_callbacks.delete(@subscription_group.id)
+          ::Karafka::Instrumentation.error_callbacks.delete(@subscription_group.id)
+
+          @kafka.close
+          @buffer.clear
+          # @note We do not clear rebalance manager here as we may still have revocation info here
+          # that we want to consider valid prior to running another reconnection
+        end
+      end
+
+      # @param topic [String]
+      # @param partition [Integer]
+      # @return [Rdkafka::Consumer::TopicPartitionList]
+      def topic_partition_list(topic, partition)
+        rdkafka_partition = @kafka
+                            .assignment
+                            .to_h[topic]
+                            &.detect { |part| part.partition == partition }
+
+        return unless rdkafka_partition
+
+        Rdkafka::Consumer::TopicPartitionList.new({ topic => [rdkafka_partition] })
+      end
+
+      # Performs a single poll operation.
+      #
+      # @param timeout [Integer] timeout for a single poll
+      # @return [Rdkafka::Consumer::Message, nil] fetched message or nil if nothing polled
+      def poll(timeout)
+        time_poll ||= TimeTrackers::Poll.new(timeout)
+
+        return nil if time_poll.exceeded?
 
-      # @return [Kafka::Consumer] returns a ready to consume Kafka consumer
-      #   that is set up to consume from topics of a given consumer group
-      def kafka_consumer
-        # @note We don't cache the connection internally because we cache kafka_consumer that uses
-        #   kafka client object instance
-        @kafka_consumer ||= Builder.call(consumer_group).consumer(
-          **ApiAdapter.consumer(consumer_group)
-        ).tap do |consumer|
-          consumer_group.topics.each do |topic|
-            settings = ApiAdapter.subscribe(topic)
+        time_poll.start
 
-            consumer.subscribe(settings[0], **settings[1])
+        @kafka.poll(timeout)
+      rescue ::Rdkafka::RdkafkaError => e
+        # We return nil, so we do not restart until running the whole loop
+        # This allows us to run revocation jobs and other things and we will pick up new work
+        # next time after dispatching all the things that are needed
+        #
+        # If we would retry here, the client reset would become transparent and we would not have
+        # a chance to take any actions
+        case e.code
+        when :max_poll_exceeded # -147
+          reset
+          return nil
+        when :transport # -195
+          reset
+          return nil
+        when :rebalance_in_progress # -27
+          reset
+          return nil
+        when :not_coordinator # 16
+          reset
+          return nil
+        when :network_exception # 13
+          reset
+          return nil
+        when :unknown_topic_or_part
+          # This is expected and temporary until rdkafka catches up with metadata
+          return nil
+        end
+
+        raise if time_poll.attempts > MAX_POLL_RETRIES
+        raise unless time_poll.retryable?
+
+        time_poll.checkpoint
+        time_poll.backoff
+
+        # On unknown errors we do our best to retry and handle them before raising
+        retry
+      end
+
+      # Builds a new rdkafka consumer instance based on the subscription group configuration
+      # @return [Rdkafka::Consumer]
+      def build_consumer
+        ::Rdkafka::Config.logger = ::Karafka::App.config.logger
+        config = ::Rdkafka::Config.new(@subscription_group.kafka)
+        config.consumer_rebalance_listener = @rebalance_manager
+        consumer = config.consumer
+        @name = consumer.name
+
+        # Register statistics runner for this particular type of callbacks
+        ::Karafka::Instrumentation.statistics_callbacks.add(
+          @subscription_group.id,
+          Instrumentation::Callbacks::Statistics.new(
+            @subscription_group.id,
+            @subscription_group.consumer_group_id,
+            @name,
+            ::Karafka::App.config.monitor
+          )
+        )
+
+        # Register error tracking callback
+        ::Karafka::Instrumentation.error_callbacks.add(
+          @subscription_group.id,
+          Instrumentation::Callbacks::Error.new(
+            @subscription_group.id,
+            @subscription_group.consumer_group_id,
+            @name,
+            ::Karafka::App.config.monitor
+          )
+        )
+
+        # Subscription needs to happen after we assigned the rebalance callbacks just in case of
+        # a race condition
+        consumer.subscribe(*@subscription_group.topics.map(&:name))
+        consumer
+      end
+
+      # We may have a case where in the middle of data polling, we've lost a partition.
+      # In a case like this we should remove all the pre-buffered messages from list partitions as
+      # we are no longer responsible in a given process for processing those messages and they
+      # should have been picked up by a different process.
+      def remove_revoked_and_duplicated_messages
+        @rebalance_manager.lost_partitions.each do |topic, partitions|
+          partitions.each do |partition|
+            @buffer.delete(topic, partition)
           end
         end
-      rescue Kafka::ConnectionError
-        # If we would not wait it will spam log file with failed
-        # attempts if Kafka is down
-        sleep(consumer_group.reconnect_timeout)
-        # We don't log and just re-raise - this will be logged
-        # down the road
-        raise
+
+        @buffer.uniq!
       end
     end
   end