karafka 2.1.4 → 2.1.6

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

@@ -49,14 +49,13 @@ module Karafka
           #
           # @param code [Integer]
           # @param opaque [Rdkafka::Opaque]
-          # @param consumer [Rdkafka::Consumer]
           # @param tpl [Rdkafka::Consumer::TopicPartitionList]
-          def trigger_callbacks(code, opaque, consumer, tpl)
+          def trigger_callbacks(code, opaque, tpl)
             case code
             when RB::RD_KAFKA_RESP_ERR__ASSIGN_PARTITIONS
-              opaque.call_on_partitions_assigned(consumer, tpl)
+              opaque.call_on_partitions_assigned(tpl)
             when RB::RD_KAFKA_RESP_ERR__REVOKE_PARTITIONS
-              opaque.call_on_partitions_revoked(consumer, tpl)
+              opaque.call_on_partitions_revoked(tpl)
             end
           rescue StandardError => e
             Karafka.monitor.instrument(
@@ -92,9 +91,8 @@ module Karafka
           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)
+          pr.trigger_callbacks(code, opaque, tpl)
         end
       end
     end

data/lib/karafka/pro/iterator/expander.rb

@@ -0,0 +1,95 @@
+# frozen_string_literal: true
+
+# This Karafka component is a Pro component under a commercial license.
+# This Karafka component is NOT licensed under LGPL.
+#
+# All of the commercial components are present in the lib/karafka/pro directory of this
+# repository and their usage requires commercial license agreement.
+#
+# Karafka has also commercial-friendly license, commercial support and commercial components.
+#
+# By sending a pull request to the pro components, you are agreeing to transfer the copyright of
+# your code to Maciej Mensfeld.
+
+module Karafka
+  module Pro
+    class Iterator
+      # There are various ways you can provide topics information for iterating.
+      #
+      # This mapper normalizes this data, resolves offsets and maps the time based offsets into
+      # appropriate once
+      #
+      # Following formats are accepted:
+      #
+      # - 'topic1' - just a string with one topic name
+      # - ['topic1', 'topic2'] - just the names
+      # - { 'topic1' => -100 } - names with negative lookup offset
+      # - { 'topic1' => { 0 => 5 } } - names with exact partitions offsets
+      # - { 'topic1' => { 0 => -5 }, 'topic2' => { 1 => 5 } } - with per partition negative offsets
+      # - { 'topic1' => 100 } - means we run all partitions from the offset 100
+      # - { 'topic1' => Time.now - 60 } - we run all partitions from the message from 60s ago
+      # - { 'topic1' => { 1 => Time.now - 60 } } - partition1 from message 60s ago
+      #
+      class Expander
+        # Expands topics to which we want to subscribe with partitions information in case this
+        # info is not provided.
+        #
+        # @param topics [Array, Hash, String] topics definitions
+        # @return [Hash] expanded and normalized requested topics and partitions data
+        def call(topics)
+          expanded = Hash.new { |h, k| h[k] = {} }
+
+          normalize_format(topics).map do |topic, details|
+            if details.is_a?(Hash)
+              details.each do |partition, offset|
+                expanded[topic][partition] = offset
+              end
+            else
+              partition_count(topic).times do |partition|
+                # If no offsets are provided, we just start from zero
+                expanded[topic][partition] = details || 0
+              end
+            end
+          end
+
+          expanded
+        end
+
+        private
+
+        # Input can be provided in multiple formats. Here we normalize it to one (hash).
+        #
+        # @param topics [Array, Hash, String] requested topics
+        # @return [Hash] normalized hash with topics data
+        def normalize_format(topics)
+          # Simplification for the single topic case
+          topics = [topics] if topics.is_a?(String)
+
+          # If we've got just array with topics, we need to convert that into a representation
+          # that we can expand with offsets
+          topics = topics.map { |name| [name, false] }.to_h if topics.is_a?(Array)
+          # We remap by creating new hash, just in case the hash came as the argument for this
+          # expanded. We do not want to modify user provided hash
+          topics.transform_keys(&:to_s)
+        end
+
+        # List of topics with their partition information for expansion
+        # We cache it so we do not have to run consecutive requests to obtain data about multiple
+        # topics
+        def topics
+          @topics ||= Admin.cluster_info.topics
+        end
+
+        # @param name [String] topic name
+        # @return [Integer] number of partitions of the topic we want to iterate over
+        def partition_count(name)
+          topics
+            .find { |topic| topic.fetch(:topic_name) == name }
+            .tap { |topic| topic || raise(Errors::TopicNotFoundError, name) }
+            .fetch(:partitions)
+            .count
+        end
+      end
+    end
+  end
+end

data/lib/karafka/pro/iterator/tpl_builder.rb

@@ -0,0 +1,145 @@
+# frozen_string_literal: true
+
+# This Karafka component is a Pro component under a commercial license.
+# This Karafka component is NOT licensed under LGPL.
+#
+# All of the commercial components are present in the lib/karafka/pro directory of this
+# repository and their usage requires commercial license agreement.
+#
+# Karafka has also commercial-friendly license, commercial support and commercial components.
+#
+# By sending a pull request to the pro components, you are agreeing to transfer the copyright of
+# your code to Maciej Mensfeld.
+
+module Karafka
+  module Pro
+    class Iterator
+      # Max time for a TPL request. We increase it to compensate for remote clusters latency
+      TPL_REQUEST_TIMEOUT = 2_000
+
+      private_constant :TPL_REQUEST_TIMEOUT
+
+      # Because we have various formats in which we can provide the offsets, before we can
+      # subscribe to them, there needs to be a bit of normalization.
+      #
+      # For some of the cases, we need to go to Kafka and get the real offsets or watermarks.
+      #
+      # This builder resolves that and builds a tpl to which we can safely subscribe the way
+      # we want it.
+      class TplBuilder
+        # @param consumer [::Rdkafka::Consumer] consumer instance needed to talk with Kafka
+        # @param expanded_topics [Hash] hash with expanded and normalized topics data
+        def initialize(consumer, expanded_topics)
+          @consumer = consumer
+          @expanded_topics = expanded_topics
+          @mapped_topics = Hash.new { |h, k| h[k] = {} }
+        end
+
+        # @return [Rdkafka::Consumer::TopicPartitionList] final tpl we can use to subscribe
+        def call
+          resolve_partitions_without_offsets
+          resolve_partitions_with_exact_offsets
+          resolve_partitions_with_negative_offsets
+          resolve_partitions_with_time_offsets
+
+          # Final tpl with all the data
+          tpl = Rdkafka::Consumer::TopicPartitionList.new
+
+          @mapped_topics.each do |name, partitions|
+            tpl.add_topic_and_partitions_with_offsets(name, partitions)
+          end
+
+          tpl
+        end
+
+        private
+
+        # First we expand on those partitions that do not have offsets defined.
+        # When we operate in case like this, we just start from beginning
+        def resolve_partitions_without_offsets
+          @expanded_topics.each do |name, partitions|
+            # We can here only about the case where we have partitions without offsets
+            next unless partitions.is_a?(Array) || partitions.is_a?(Range)
+
+            # When no offsets defined, we just start from zero
+            @mapped_topics[name] = partitions.map { |partition| [partition, 0] }.to_h
+          end
+        end
+
+        # If we get exact numeric offsets, we can just start from them without any extra work
+        def resolve_partitions_with_exact_offsets
+          @expanded_topics.each do |name, partitions|
+            next unless partitions.is_a?(Hash)
+
+            partitions.each do |partition, offset|
+              # Skip negative and time based offsets
+              next unless offset.is_a?(Integer) && offset >= 0
+
+              # Exact offsets can be used as they are
+              # No need for extra operations
+              @mapped_topics[name][partition] = offset
+            end
+          end
+        end
+
+        # If the offsets are negative, it means we want to fetch N last messages and we need to
+        # figure out the appropriate offsets
+        #
+        # We do it by getting the watermark offsets and just calculating it. This means that for
+        # heavily compacted topics, this may return less than the desired number but it is a
+        # limitation that is documented.
+        def resolve_partitions_with_negative_offsets
+          @expanded_topics.each do |name, partitions|
+            next unless partitions.is_a?(Hash)
+
+            partitions.each do |partition, offset|
+              # Care only about negative offsets (last n messages)
+              next unless offset.is_a?(Integer) && offset.negative?
+
+              _, high_watermark_offset = @consumer.query_watermark_offsets(name, partition)
+              # We add because this offset is negative
+              @mapped_topics[name][partition] = high_watermark_offset + offset
+            end
+          end
+        end
+
+        # For time based offsets we first need to aggregate them and request the proper offsets.
+        # We want to get all times in one go for all tpls defined with times, so we accumulate
+        # them here and we will make one sync request to kafka for all.
+        def resolve_partitions_with_time_offsets
+          time_tpl = Rdkafka::Consumer::TopicPartitionList.new
+
+          # First we need to collect the time based once
+          @expanded_topics.each do |name, partitions|
+            next unless partitions.is_a?(Hash)
+
+            time_based = {}
+
+            partitions.each do |partition, offset|
+              next unless offset.is_a?(Time)
+
+              time_based[partition] = offset
+            end
+
+            next if time_based.empty?
+
+            time_tpl.add_topic_and_partitions_with_offsets(name, time_based)
+          end
+
+          # If there were no time-based, no need to query Kafka
+          return if time_tpl.empty?
+
+          real_offsets = @consumer.offsets_for_times(time_tpl, TPL_REQUEST_TIMEOUT)
+
+          real_offsets.to_h.each do |name, results|
+            results.each do |result|
+              raise(Errors::InvalidTimeBasedOffsetError) unless result
+
+              @mapped_topics[name][result.partition] = result.offset
+            end
+          end
+        end
+      end
+    end
+  end
+end

data/lib/karafka/pro/iterator.rb

@@ -50,7 +50,7 @@ module Karafka
         settings: { 'auto.offset.reset': 'beginning' },
         yield_nil: false
       )
-        @topics_with_partitions = expand_topics_with_partitions(topics)
+        @topics_with_partitions = Expander.new.call(topics)
 
         @routing_topics = @topics_with_partitions.map do |name, _|
           [name, ::Karafka::Routing::Router.find_or_initialize_by_name(name)]
@@ -71,7 +71,7 @@ module Karafka
       # only eat up resources.
       def each
         Admin.with_consumer(@settings) do |consumer|
-          tpl = tpl_with_expanded_offsets(consumer)
+          tpl = TplBuilder.new(consumer, @topics_with_partitions).call
           consumer.assign(tpl)
 
           # We need this for self-referenced APIs like pausing
@@ -131,43 +131,6 @@ module Karafka
 
       private
 
-      # Expands topics to which we want to subscribe with partitions information in case this
-      # info is not provided. For our convenience we want to support 5 formats of defining
-      # the subscribed topics:
-      #
-      # - 'topic1' - just a string with one topic name
-      # - ['topic1', 'topic2'] - just the names
-      # - { 'topic1' => -100 } - names with negative lookup offset
-      # - { 'topic1' => { 0 => 5 } } - names with exact partitions offsets
-      # - { 'topic1' => { 0 => -5 }, 'topic2' => { 1 => 5 } } - with per partition negative offsets
-      #
-      # @param topics [Array, Hash] topics definitions
-      # @return [Hash] hash with topics containing partitions definitions
-      def expand_topics_with_partitions(topics)
-        # Simplification for the single topic case
-        topics = [topics] if topics.is_a?(String)
-        # If we've got just array with topics, we need to convert that into a representation
-        # that we can expand with offsets
-        topics = topics.map { |name| [name, false] }.to_h if topics.is_a?(Array)
-
-        expanded = Hash.new { |h, k| h[k] = {} }
-
-        topics.map do |topic, details|
-          if details.is_a?(Hash)
-            details.each do |partition, offset|
-              expanded[topic][partition] = offset
-            end
-          else
-            partition_count(topic.to_s).times do |partition|
-              # If no offsets are provided, we just start from zero
-              expanded[topic][partition] = details || 0
-            end
-          end
-        end
-
-        expanded
-      end
-
       # @param timeout [Integer] timeout in ms
       # @return [Rdkafka::Consumer::Message, nil] message or nil if nothing to do
       def poll(timeout)
@@ -200,54 +163,6 @@ module Karafka
       def done?
         @stopped_partitions >= @total_partitions
       end
-
-      # Builds the tpl representing all the subscriptions we want to run
-      #
-      # Additionally for negative offsets, does the watermark calculation where to start
-      #
-      # @param consumer [Rdkafka::Consumer] consumer we need in case of negative offsets as
-      #   negative are going to be used to do "give me last X". We use the already initialized
-      #   consumer instance, not to start another one again.
-      # @return [Rdkafka::Consumer::TopicPartitionList]
-      def tpl_with_expanded_offsets(consumer)
-        tpl = Rdkafka::Consumer::TopicPartitionList.new
-
-        @topics_with_partitions.each do |name, partitions|
-          partitions_with_offsets = {}
-
-          # When no offsets defined, we just start from zero
-          if partitions.is_a?(Array) || partitions.is_a?(Range)
-            partitions_with_offsets = partitions.map { |partition| [partition, 0] }.to_h
-          else
-            # When offsets defined, we can either use them if positive or expand and move back
-            # in case of negative (-1000 means last 1000 messages, etc)
-            partitions.each do |partition, offset|
-              if offset.negative?
-                _, high_watermark_offset = consumer.query_watermark_offsets(name, partition)
-                # We add because this offset is negative
-                partitions_with_offsets[partition] = high_watermark_offset + offset
-              else
-                partitions_with_offsets[partition] = offset
-              end
-            end
-          end
-
-          tpl.add_topic_and_partitions_with_offsets(name, partitions_with_offsets)
-        end
-
-        tpl
-      end
-
-      # @param name [String] topic name
-      # @return [Integer] number of partitions of the topic we want to iterate over
-      def partition_count(name)
-        Admin
-          .cluster_info
-          .topics
-          .find { |topic| topic.fetch(:topic_name) == name }
-          .fetch(:partitions)
-          .count
-      end
     end
   end
 end

data/lib/karafka/pro/processing/filters_applier.rb

@@ -81,6 +81,7 @@ module Karafka
         # The first message we do need to get next time we poll. We use the minimum not to jump
         # accidentally by over any.
         # @return [Karafka::Messages::Message, nil] cursor message or nil if none
+        # @note Cursor message can also return the offset in the time format
         def cursor
           return nil unless active?
 

data/lib/karafka/pro/processing/strategies/aj/dlq_ftr_lrj_mom.rb

@@ -44,7 +44,9 @@ module Karafka
                   if coordinator.filtered? && !revoked?
                     handle_post_filtering
                   elsif !revoked?
-                    seek(coordinator.seek_offset)
+                    # no need to check for manual seek because AJ consumer is internal and
+                    # fully controlled by us
+                    seek(coordinator.seek_offset, false)
                     resume
                   else
                     resume

data/lib/karafka/pro/processing/strategies/aj/dlq_ftr_lrj_mom_vp.rb

@@ -50,7 +50,9 @@ module Karafka
                   if coordinator.filtered? && !revoked?
                     handle_post_filtering
                   elsif !revoked?
-                    seek(coordinator.seek_offset)
+                    # no need to check for manual seek because AJ consumer is internal and
+                    # fully controlled by us
+                    seek(coordinator.seek_offset, false)
                     resume
                   else
                     resume

data/lib/karafka/pro/processing/strategies/aj/dlq_lrj_mom.rb

@@ -42,7 +42,9 @@ module Karafka
                 if coordinator.success?
                   coordinator.pause_tracker.reset
 
-                  seek(coordinator.seek_offset) unless revoked?
+                  # no need to check for manual seek because AJ consumer is internal and
+                  # fully controlled by us
+                  seek(coordinator.seek_offset, false) unless revoked?
 
                   resume
                 elsif coordinator.pause_tracker.attempt <= topic.dead_letter_queue.max_retries

data/lib/karafka/pro/processing/strategies/aj/dlq_lrj_mom_vp.rb

@@ -46,7 +46,9 @@ module Karafka
                   # Since we have VP here we do not commit intermediate offsets and need to commit
                   # them here. We do commit in collapsed mode but this is generalized.
                   mark_as_consumed(last_group_message) unless revoked?
-                  seek(coordinator.seek_offset) unless revoked?
+                  # no need to check for manual seek because AJ consumer is internal and
+                  # fully controlled by us
+                  seek(coordinator.seek_offset, false) unless revoked?
 
                   resume
                 elsif coordinator.pause_tracker.attempt <= topic.dead_letter_queue.max_retries

data/lib/karafka/pro/processing/strategies/aj/ftr_lrj_mom_vp.rb

@@ -46,7 +46,9 @@ module Karafka
                   if coordinator.filtered? && !revoked?
                     handle_post_filtering
                   elsif !revoked?
-                    seek(coordinator.seek_offset)
+                    # no need to check for manual seek because AJ consumer is internal and
+                    # fully controlled by us
+                    seek(coordinator.seek_offset, false)
                     resume
                   else
                     resume

data/lib/karafka/pro/processing/strategies/aj/lrj_mom_vp.rb

@@ -48,7 +48,10 @@ module Karafka
                   coordinator.pause_tracker.reset
 
                   mark_as_consumed(last_group_message) unless revoked?
-                  seek(coordinator.seek_offset) unless revoked?
+
+                  # no need to check for manual seek because AJ consumer is internal and
+                  # fully controlled by us
+                  seek(coordinator.seek_offset, false) unless revoked?
 
                   resume
                 else

data/lib/karafka/pro/processing/strategies/dlq/ftr_lrj.rb

@@ -47,8 +47,8 @@ module Karafka
 
                   if coordinator.filtered? && !revoked?
                     handle_post_filtering
-                  elsif !revoked?
-                    seek(coordinator.seek_offset)
+                  elsif !revoked? && !coordinator.manual_seek?
+                    seek(coordinator.seek_offset, false)
                     resume
                   else
                     resume

data/lib/karafka/pro/processing/strategies/dlq/ftr_lrj_mom.rb

@@ -42,8 +42,8 @@ module Karafka
 
                   if coordinator.filtered? && !revoked?
                     handle_post_filtering
-                  elsif !revoked?
-                    seek(last_group_message.offset + 1)
+                  elsif !revoked? && !coordinator.manual_seek?
+                    seek(last_group_message.offset + 1, false)
                     resume
                   else
                     resume

data/lib/karafka/pro/processing/strategies/dlq/lrj.rb

@@ -38,7 +38,8 @@ module Karafka
                   return if coordinator.manual_pause?
 
                   mark_as_consumed(last_group_message) unless revoked?
-                  seek(coordinator.seek_offset) unless revoked?
+                  # We should not overwrite user manual seel request with our seek
+                  seek(coordinator.seek_offset, false) unless revoked? || coordinator.manual_seek?
 
                   resume
                 elsif coordinator.pause_tracker.attempt <= topic.dead_letter_queue.max_retries

data/lib/karafka/pro/processing/strategies/dlq/lrj_mom.rb

@@ -37,7 +37,9 @@ module Karafka
 
                   return if coordinator.manual_pause?
 
-                  seek(last_group_message.offset + 1) unless revoked?
+                  unless revoked? || coordinator.manual_seek?
+                    seek(last_group_message.offset + 1, false)
+                  end
 
                   resume
                 elsif coordinator.pause_tracker.attempt <= topic.dead_letter_queue.max_retries

data/lib/karafka/pro/processing/strategies/ftr/default.rb

@@ -70,6 +70,10 @@ module Karafka
               when :skip
                 nil
               when :seek
+                # User direct actions take priority over automatic operations
+                # If we've already seeked we can just resume operations, nothing extra needed
+                return resume if coordinator.manual_seek?
+
                 throttle_message = filter.cursor
 
                 Karafka.monitor.instrument(
@@ -77,11 +81,14 @@ module Karafka
                   caller: self,
                   message: throttle_message
                 ) do
-                  seek(throttle_message.offset)
+                  seek(throttle_message.offset, false)
                 end
 
                 resume
               when :pause
+                # User direct actions take priority over automatic operations
+                return nil if coordinator.manual_pause?
+
                 throttle_message = filter.cursor
 
                 Karafka.monitor.instrument(

data/lib/karafka/pro/processing/strategies/lrj/default.rb

@@ -53,7 +53,7 @@ module Karafka
                   return if coordinator.manual_pause?
 
                   mark_as_consumed(last_group_message) unless revoked?
-                  seek(coordinator.seek_offset) unless revoked?
+                  seek(coordinator.seek_offset, false) unless revoked? || coordinator.manual_seek?
 
                   resume
                 else

data/lib/karafka/pro/processing/strategies/lrj/ftr.rb

@@ -45,10 +45,10 @@ module Karafka
                   # If still not revoked and was throttled, we need to apply throttling logic
                   if coordinator.filtered? && !revoked?
                     handle_post_filtering
-                  elsif !revoked?
+                  elsif !revoked? && !coordinator.manual_seek?
                     # If not revoked and not throttled, we move to where we were suppose to and
                     # resume
-                    seek(coordinator.seek_offset)
+                    seek(coordinator.seek_offset, false)
                     resume
                   else
                     resume

data/lib/karafka/pro/processing/strategies/lrj/ftr_mom.rb

@@ -43,10 +43,10 @@ module Karafka
                   # If still not revoked and was throttled, we need to apply filtering logic
                   if coordinator.filtered? && !revoked?
                     handle_post_filtering
-                  elsif !revoked?
+                  elsif !revoked? && !coordinator.manual_seek?
                     # If not revoked and not throttled, we move to where we were suppose to and
                     # resume
-                    seek(last_group_message.offset + 1)
+                    seek(last_group_message.offset + 1, false)
                     resume
                   else
                     resume

data/lib/karafka/pro/processing/strategies/lrj/mom.rb

@@ -52,7 +52,9 @@ module Karafka
 
                   return if coordinator.manual_pause?
 
-                  seek(last_group_message.offset + 1) unless revoked?
+                  unless revoked? || coordinator.manual_seek?
+                    seek(last_group_message.offset + 1, false)
+                  end
 
                   resume
                 else

data/lib/karafka/pro/processing/virtual_offset_manager.rb

@@ -118,7 +118,7 @@ module Karafka
 
         # @return [Messages::Seek] markable message for real offset marking
         def markable
-          raise Errors::InvalidRealOffsetUsage unless markable?
+          raise Errors::InvalidRealOffsetUsageError unless markable?
 
           Messages::Seek.new(
             @topic,

data/lib/karafka/processing/coordinator.rb

@@ -23,6 +23,7 @@ module Karafka
         @consumptions = {}
         @running_jobs = 0
         @manual_pause = false
+        @manual_seek = false
         @mutex = Mutex.new
         @marked = false
         @failure = false
@@ -41,6 +42,9 @@ module Karafka
         # When starting to run, no pause is expected and no manual pause as well
         @manual_pause = false
 
+        # No user invoked seeks on a new run
+        @manual_seek = false
+
         # We set it on the first encounter and never again, because then the offset setting
         # should be up to the consumers logic (our or the end user)
         # Seek offset needs to be always initialized as for case where manual offset management
@@ -148,6 +152,16 @@ module Karafka
         @pause_tracker.paused? && @manual_pause
       end
 
+      # Marks seek as manual for coordination purposes
+      def manual_seek
+        @manual_seek = true
+      end
+
+      # @return [Boolean] did a user invoke seek in the current operations scope
+      def manual_seek?
+        @manual_seek
+      end
+
       # Allows to run synchronized (locked) code that can operate in between virtual partitions
       # @param block [Proc] code we want to run in the synchronized mode
       def synchronize(&block)