Class: Karafka::Pro::Processing::VirtualOffsetManager

Inherits:
Object
  • Object
show all
Defined in:
lib/karafka/pro/processing/virtual_offset_manager.rb

Overview

Note:

We still use the regular coordinator “real” offset management as we want to have them as separated as possible because the real seek offset management is also used for pausing, filtering and others and should not be impacted by the virtual one

Note:

This manager is not thread-safe by itself. It should operate from coordinator locked locations.

Manager that keeps track of our offsets with the virtualization layer that are local to given partition assignment. It allows for easier offset management for virtual virtual partition cases as it provides us ability to mark as consumed and move the real offset behind as expected.

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(topic, partition, offset_metadata_strategy) ⇒ VirtualOffsetManager

Note:

We need topic and partition because we use a seek message (virtual) for real offset management. We could keep real message reference but this can be memory consuming and not worth it.

Returns a new instance of VirtualOffsetManager.

Parameters:

  • topic (String)
  • partition (Integer)
  • offset_metadata_strategy (Symbol) —

    what metadata should we select. That is, should we use the most recent or one picked from the offset that is going to be committed



39
40
41
42
43
44
45
46
47
48
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 39

def initialize(topic, partition, offset_metadata_strategy)
  @topic = topic
  @partition = partition
  @groups = []
  @marked = {}
  @offsets_metadata = {}
  @real_offset = -1
  @offset_metadata_strategy = offset_metadata_strategy
  @current_offset_metadata = nil
end

Instance Attribute Details

#groups ⇒ Object (readonly)

Returns the value of attribute groups.



29
30
31
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 29

def groups
  @groups
end

Instance Method Details

#clear ⇒ Object

Clears the manager for a next collective operation



51
52
53
54
55
56
57
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 51

def clear
  @groups.clear
  @offsets_metadata.clear
  @current_offset_metadata = nil
  @marked.clear
  @real_offset = -1
end

#mark(message, offset_metadata) ⇒ Object

Marks given message as marked (virtually consumed). We mark given message offset and other earlier offsets from the same group as done and we can refresh our real offset representation based on that as it might have changed to a newer real offset.

Parameters:

  • message (Karafka::Messages::Message) —

    message coming from VP we want to mark

  • offset_metadata (String, nil) —

    offset metadata. nil if none



76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 76

def mark(message, offset_metadata)
  offset = message.offset

  # Store metadata when we materialize the most stable offset
  @offsets_metadata[offset] = offset_metadata
  @current_offset_metadata = offset_metadata

  group = @groups.find { |reg_group| reg_group.include?(offset) }

  # This case can happen when someone uses MoM and wants to mark message from a previous
  # batch as consumed. We can add it, since the real offset refresh will point to it
  unless group
    group = [offset]
    @groups << group
  end

  position = group.index(offset)

  # Mark all previous messages from the same group also as virtually consumed
  group[0..position].each do |markable_offset|
    # Set previous messages metadata offset as the offset of higher one for overwrites
    # unless a different metadata were set explicitely
    @offsets_metadata[markable_offset] ||= offset_metadata
    @marked[markable_offset] = true
  end

  # Recompute the real offset representation
  materialize_real_offset
end

#mark_until(message, offset_metadata) ⇒ Object

Mark all from all groups including the message. Useful when operating in a collapsed state for marking

Parameters:



110
111
112
113
114
115
116
117
118
119
120
121
122
123
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 110

def mark_until(message, offset_metadata)
  mark(message, offset_metadata)

  @groups.each do |group|
    group.each do |offset|
      next if offset > message.offset

      @offsets_metadata[offset] = offset_metadata
      @marked[offset] = true
    end
  end

  materialize_real_offset
end

#markable ⇒ Array<Messages::Seek, String>

Returns markable message for real offset marking and its associated metadata.

Returns:

  • (Array<Messages::Seek, String>) —

    markable message for real offset marking and its associated metadata

Raises:



138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 138

def markable
  raise Errors::InvalidRealOffsetUsageError unless markable?

  offset_metadata = case @offset_metadata_strategy
                    when :exact
                      @offsets_metadata.fetch(@real_offset)
                    when :current
                      @current_offset_metadata
                    else
                      raise Errors::UnsupportedCaseError, @offset_metadata_strategy
                    end

  [
    Messages::Seek.new(
      @topic,
      @partition,
      @real_offset
    ),
    offset_metadata
  ]
end

#markable? ⇒ Boolean

Is there a real offset we can mark as consumed

Returns:

  • (Boolean) 


132
133
134
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 132

def markable?
  !@real_offset.negative?
end

#marked ⇒ Array<Integer>

Returns Offsets of messages already marked as consumed virtually.

Returns:

  • (Array<Integer>) —

    Offsets of messages already marked as consumed virtually



126
127
128
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 126

def marked
  @marked.select { |_, status| status }.map(&:first).sort
end

#register(offsets_group) ⇒ Object

Registers an offset group coming from one virtual consumer. In order to move the real underlying offset accordingly, we need to make sure to track the virtual consumers offsets groups independently and only materialize the end result.

Parameters:

  • offsets_group (Array<Integer>) —

    offsets from one virtual consumer



64
65
66
67
68
 
# File 'lib/karafka/pro/processing/virtual_offset_manager.rb', line 64

def register(offsets_group)
  @groups << offsets_group

  offsets_group.each { |offset| @marked[offset] = false }
end