Takahiro Suzuki | 241c10e | 2020-12-17 20:17:57 +0900 | [diff] [blame] | 1 | package sarama |
| 2 | |
| 3 | import ( |
| 4 | "hash" |
| 5 | "hash/fnv" |
| 6 | "math/rand" |
| 7 | "time" |
| 8 | ) |
| 9 | |
| 10 | // Partitioner is anything that, given a Kafka message and a number of partitions indexed [0...numPartitions-1], |
| 11 | // decides to which partition to send the message. RandomPartitioner, RoundRobinPartitioner and HashPartitioner are provided |
| 12 | // as simple default implementations. |
| 13 | type Partitioner interface { |
| 14 | // Partition takes a message and partition count and chooses a partition |
| 15 | Partition(message *ProducerMessage, numPartitions int32) (int32, error) |
| 16 | |
| 17 | // RequiresConsistency indicates to the user of the partitioner whether the |
| 18 | // mapping of key->partition is consistent or not. Specifically, if a |
| 19 | // partitioner requires consistency then it must be allowed to choose from all |
| 20 | // partitions (even ones known to be unavailable), and its choice must be |
| 21 | // respected by the caller. The obvious example is the HashPartitioner. |
| 22 | RequiresConsistency() bool |
| 23 | } |
| 24 | |
| 25 | // DynamicConsistencyPartitioner can optionally be implemented by Partitioners |
| 26 | // in order to allow more flexibility than is originally allowed by the |
| 27 | // RequiresConsistency method in the Partitioner interface. This allows |
| 28 | // partitioners to require consistency sometimes, but not all times. It's useful |
| 29 | // for, e.g., the HashPartitioner, which does not require consistency if the |
| 30 | // message key is nil. |
| 31 | type DynamicConsistencyPartitioner interface { |
| 32 | Partitioner |
| 33 | |
| 34 | // MessageRequiresConsistency is similar to Partitioner.RequiresConsistency, |
| 35 | // but takes in the message being partitioned so that the partitioner can |
| 36 | // make a per-message determination. |
| 37 | MessageRequiresConsistency(message *ProducerMessage) bool |
| 38 | } |
| 39 | |
| 40 | // PartitionerConstructor is the type for a function capable of constructing new Partitioners. |
| 41 | type PartitionerConstructor func(topic string) Partitioner |
| 42 | |
| 43 | type manualPartitioner struct{} |
| 44 | |
| 45 | // HashPartitionOption lets you modify default values of the partitioner |
| 46 | type HashPartitionerOption func(*hashPartitioner) |
| 47 | |
| 48 | // WithAbsFirst means that the partitioner handles absolute values |
| 49 | // in the same way as the reference Java implementation |
| 50 | func WithAbsFirst() HashPartitionerOption { |
| 51 | return func(hp *hashPartitioner) { |
| 52 | hp.referenceAbs = true |
| 53 | } |
| 54 | } |
| 55 | |
| 56 | // WithCustomHashFunction lets you specify what hash function to use for the partitioning |
| 57 | func WithCustomHashFunction(hasher func() hash.Hash32) HashPartitionerOption { |
| 58 | return func(hp *hashPartitioner) { |
| 59 | hp.hasher = hasher() |
| 60 | } |
| 61 | } |
| 62 | |
| 63 | // WithCustomFallbackPartitioner lets you specify what HashPartitioner should be used in case a Distribution Key is empty |
| 64 | func WithCustomFallbackPartitioner(randomHP *hashPartitioner) HashPartitionerOption { |
| 65 | return func(hp *hashPartitioner) { |
| 66 | hp.random = hp |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | // NewManualPartitioner returns a Partitioner which uses the partition manually set in the provided |
| 71 | // ProducerMessage's Partition field as the partition to produce to. |
| 72 | func NewManualPartitioner(topic string) Partitioner { |
| 73 | return new(manualPartitioner) |
| 74 | } |
| 75 | |
| 76 | func (p *manualPartitioner) Partition(message *ProducerMessage, numPartitions int32) (int32, error) { |
| 77 | return message.Partition, nil |
| 78 | } |
| 79 | |
| 80 | func (p *manualPartitioner) RequiresConsistency() bool { |
| 81 | return true |
| 82 | } |
| 83 | |
| 84 | type randomPartitioner struct { |
| 85 | generator *rand.Rand |
| 86 | } |
| 87 | |
| 88 | // NewRandomPartitioner returns a Partitioner which chooses a random partition each time. |
| 89 | func NewRandomPartitioner(topic string) Partitioner { |
| 90 | p := new(randomPartitioner) |
| 91 | p.generator = rand.New(rand.NewSource(time.Now().UTC().UnixNano())) |
| 92 | return p |
| 93 | } |
| 94 | |
| 95 | func (p *randomPartitioner) Partition(message *ProducerMessage, numPartitions int32) (int32, error) { |
| 96 | return int32(p.generator.Intn(int(numPartitions))), nil |
| 97 | } |
| 98 | |
| 99 | func (p *randomPartitioner) RequiresConsistency() bool { |
| 100 | return false |
| 101 | } |
| 102 | |
| 103 | type roundRobinPartitioner struct { |
| 104 | partition int32 |
| 105 | } |
| 106 | |
| 107 | // NewRoundRobinPartitioner returns a Partitioner which walks through the available partitions one at a time. |
| 108 | func NewRoundRobinPartitioner(topic string) Partitioner { |
| 109 | return &roundRobinPartitioner{} |
| 110 | } |
| 111 | |
| 112 | func (p *roundRobinPartitioner) Partition(message *ProducerMessage, numPartitions int32) (int32, error) { |
| 113 | if p.partition >= numPartitions { |
| 114 | p.partition = 0 |
| 115 | } |
| 116 | ret := p.partition |
| 117 | p.partition++ |
| 118 | return ret, nil |
| 119 | } |
| 120 | |
| 121 | func (p *roundRobinPartitioner) RequiresConsistency() bool { |
| 122 | return false |
| 123 | } |
| 124 | |
| 125 | type hashPartitioner struct { |
| 126 | random Partitioner |
| 127 | hasher hash.Hash32 |
| 128 | referenceAbs bool |
| 129 | } |
| 130 | |
| 131 | // NewCustomHashPartitioner is a wrapper around NewHashPartitioner, allowing the use of custom hasher. |
| 132 | // The argument is a function providing the instance, implementing the hash.Hash32 interface. This is to ensure that |
| 133 | // each partition dispatcher gets its own hasher, to avoid concurrency issues by sharing an instance. |
| 134 | func NewCustomHashPartitioner(hasher func() hash.Hash32) PartitionerConstructor { |
| 135 | return func(topic string) Partitioner { |
| 136 | p := new(hashPartitioner) |
| 137 | p.random = NewRandomPartitioner(topic) |
| 138 | p.hasher = hasher() |
| 139 | p.referenceAbs = false |
| 140 | return p |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | // NewCustomPartitioner creates a default Partitioner but lets you specify the behavior of each component via options |
| 145 | func NewCustomPartitioner(options ...HashPartitionerOption) PartitionerConstructor { |
| 146 | return func(topic string) Partitioner { |
| 147 | p := new(hashPartitioner) |
| 148 | p.random = NewRandomPartitioner(topic) |
| 149 | p.hasher = fnv.New32a() |
| 150 | p.referenceAbs = false |
| 151 | for _, option := range options { |
| 152 | option(p) |
| 153 | } |
| 154 | return p |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | // NewHashPartitioner returns a Partitioner which behaves as follows. If the message's key is nil then a |
| 159 | // random partition is chosen. Otherwise the FNV-1a hash of the encoded bytes of the message key is used, |
| 160 | // modulus the number of partitions. This ensures that messages with the same key always end up on the |
| 161 | // same partition. |
| 162 | func NewHashPartitioner(topic string) Partitioner { |
| 163 | p := new(hashPartitioner) |
| 164 | p.random = NewRandomPartitioner(topic) |
| 165 | p.hasher = fnv.New32a() |
| 166 | p.referenceAbs = false |
| 167 | return p |
| 168 | } |
| 169 | |
| 170 | // NewReferenceHashPartitioner is like NewHashPartitioner except that it handles absolute values |
| 171 | // in the same way as the reference Java implementation. NewHashPartitioner was supposed to do |
| 172 | // that but it had a mistake and now there are people depending on both behaviours. This will |
| 173 | // all go away on the next major version bump. |
| 174 | func NewReferenceHashPartitioner(topic string) Partitioner { |
| 175 | p := new(hashPartitioner) |
| 176 | p.random = NewRandomPartitioner(topic) |
| 177 | p.hasher = fnv.New32a() |
| 178 | p.referenceAbs = true |
| 179 | return p |
| 180 | } |
| 181 | |
| 182 | func (p *hashPartitioner) Partition(message *ProducerMessage, numPartitions int32) (int32, error) { |
| 183 | if message.Key == nil { |
| 184 | return p.random.Partition(message, numPartitions) |
| 185 | } |
| 186 | bytes, err := message.Key.Encode() |
| 187 | if err != nil { |
| 188 | return -1, err |
| 189 | } |
| 190 | p.hasher.Reset() |
| 191 | _, err = p.hasher.Write(bytes) |
| 192 | if err != nil { |
| 193 | return -1, err |
| 194 | } |
| 195 | var partition int32 |
| 196 | // Turns out we were doing our absolute value in a subtly different way from the upstream |
| 197 | // implementation, but now we need to maintain backwards compat for people who started using |
| 198 | // the old version; if referenceAbs is set we are compatible with the reference java client |
| 199 | // but not past Sarama versions |
| 200 | if p.referenceAbs { |
| 201 | partition = (int32(p.hasher.Sum32()) & 0x7fffffff) % numPartitions |
| 202 | } else { |
| 203 | partition = int32(p.hasher.Sum32()) % numPartitions |
| 204 | if partition < 0 { |
| 205 | partition = -partition |
| 206 | } |
| 207 | } |
| 208 | return partition, nil |
| 209 | } |
| 210 | |
| 211 | func (p *hashPartitioner) RequiresConsistency() bool { |
| 212 | return true |
| 213 | } |
| 214 | |
| 215 | func (p *hashPartitioner) MessageRequiresConsistency(message *ProducerMessage) bool { |
| 216 | return message.Key != nil |
| 217 | } |