测试环境虚拟机
CPU:2核
RAM:2G
Kafka Topic为1分区，1副本
Kafka生产者端发送延迟优化
batch.size
batch.size 单位为字节，为了方便这里都表示为kb
默认配置,batch.size=16kb
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic first --record-size 1024 --num-records 1000000 --throughput 50000 249892 records sent, 49978.4 records/sec (48.81 MB/sec), 153.6 ms avg latency, 537.0 ms max latency. 250193 records sent, 50038.6 records/sec (48.87 MB/sec), 1.4 ms avg latency, 12.0 ms max latency. 211747 records sent, 42349.4 records/sec (41.36 MB/sec), 194.3 ms avg latency, 1106.0 ms max latency. 1000000 records sent, 49972.515117 records/sec (48.80 MB/sec), 119.65 ms avg latency, 1106.00 ms max latency, 2 ms 50th, 488 ms 95th, 1043 ms 99th, 1102 ms 99.9th.
结果显示平均延迟有456.94 ms，最高延迟5308.00 ms
现在我要降低最高延迟数,batch.size的意思是ProducerBatch的内存区域充满后，消息就会被立即发送，那我们把值改小看看
batch.size=8kb
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic first --record-size 1024 --num-records 1000000 --throughput 50000 148553 records sent, 29710.6 records/sec (29.01 MB/sec), 812.4 ms avg latency, 1032.0 ms max latency. 195468 records sent, 39093.6 records/sec (38.18 MB/sec), 735.9 ms avg latency, 907.0 ms max latency. 189700 records sent, 37940.0 records/sec (37.05 MB/sec), 763.4 ms avg latency, 1053.0 ms max latency. 208418 records sent, 41683.6 records/sec (40.71 MB/sec), 689.7 ms avg latency, 923.0 ms max latency. 196504 records sent, 39300.8 records/sec (38.38 MB/sec), 718.1 ms avg latency, 1056.0 ms max latency. 1000000 records sent, 37608.123355 records/sec (36.73 MB/sec), 741.56 ms avg latency, 1056.00 ms max latency, 725 ms 50th, 937 ms 95th, 1029 ms 99th, 1051 ms 99.9th.
但经过测试发现，延迟反而很高，连设定的50000吞吐量都达不到，原因应该是这样：batch.size小了，消息很快就会充满，这样消息就会被立即发送的服务端，但这样的话发送的次数就变多了，但由于网络原因是不可控的，有时候网络发生抖动就会造成较高的延迟
那就改大看看。
batch.size=32kb
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic first --record-size 1024 --num-records 1000000 --throughput 50000 249852 records sent, 49970.4 records/sec (48.80 MB/sec), 88.8 ms avg latency, 492.0 ms max latency. 250143 records sent, 50028.6 records/sec (48.86 MB/sec), 1.2 ms avg latency, 15.0 ms max latency. 250007 records sent, 49991.4 records/sec (48.82 MB/sec), 1.2 ms avg latency, 17.0 ms max latency. 1000000 records sent, 49952.545082 records/sec (48.78 MB/sec), 31.07 ms avg latency, 492.00 ms max latency, 1 ms 50th, 305 ms 95th, 440 ms 99th, 486 ms 99.9th.
测试后，平均延迟，最高延迟都降下来很多，而且比默认值延迟都要小很多，那再改大延迟还会降低吗
batch.size=50kb
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic first --record-size 1024 --num-records 1000000 --throughput 50000 249902 records sent, 49970.4 records/sec (48.80 MB/sec), 27.3 ms avg latency, 219.0 ms max latency. 250200 records sent, 50030.0 records/sec (48.86 MB/sec), 1.2 ms avg latency, 8.0 ms max latency. 250098 records sent, 50019.6 records/sec (48.85 MB/sec), 18.6 ms avg latency, 288.0 ms max latency. 242327 records sent, 48407.3 records/sec (47.27 MB/sec), 121.3 ms avg latency, 920.0 ms max latency. 1000000 records sent, 49823.127896 records/sec (48.66 MB/sec), 41.98 ms avg latency, 920.00 ms max latency, 1 ms 50th, 221 ms 95th, 792 ms 99th, 910 ms 99.9th.
如上测试在不同的机器上结果会有不同，但总体的变化曲线是一样的，成U型变化
batch.size代码实现
Kafka客户端有一个RecordAccumulator类，叫做消息记录池，内部有一个BufferPool内存区域
RecordAccumulator(LogContext logContext, int batchSize, CompressionType compression, int lingerMs, long retryBackoffMs, int deliveryTimeoutMs, Metrics metrics, String metricGrpName, Time time, ApiVersions apiVersions, TransactionManager transactionManager, BufferPool bufferPool)
当该判断为true，消息就会被发送
if (result.batchIsFull || result.newBatchCreated) { log.trace("Waking up the sender since topic {} partition {} is either full or getting a new batch", record.topic(), partition); this.sender.wakeup(); }
PS: 6.27号 2:30
max.in.flight.requests.per.connection
该参数可以在一个connection中发送多个请求，叫作一个flight,这样可以减少开销，但是如果产生错误，可能会造成数据的发送顺序改变,默认5
在batch.size=100kb的基础上，增加该参数值到10，看看效果
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic two --record-size 1024 --num-records 1000000 --throughput 50000 249902 records sent, 49960.4 records/sec (48.79 MB/sec), 16.1 ms avg latency, 185.0 ms max latency. 250148 records sent, 50019.6 records/sec (48.85 MB/sec), 1.3 ms avg latency, 14.0 ms max latency. 239585 records sent, 47917.0 records/sec (46.79 MB/sec), 6.4 ms avg latency, 226.0 ms max latency. 1000000 records sent, 49960.031974 records/sec (48.79 MB/sec), 9.83 ms avg latency, 226.00 ms max latency, 1 ms 50th, 83 ms 95th, 182 ms 99th, 219 ms 99.9th.
多次测试结果延迟都比原来降低了10倍多，效果还是很明显的
但物极必反，如果你再调大后，效果就不明显了，最终延迟反而变高，这个batch.size道理是一样的
compression.type
指定消息的压缩方式，默认不压缩
在原来batch.size=100kb,max.in.flight.requests.per.connection=10的基础上，设置compression.type=gzip 看看延迟是否还可以降低
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic two --record-size 1024 --num-records 1000000 --throughput 50000 249785 records sent, 49957.0 records/sec (48.79 MB/sec), 2.5 ms avg latency, 199.0 ms max latency. 250091 records sent, 50008.2 records/sec (48.84 MB/sec), 1.9 ms avg latency, 17.0 ms max latency. 250123 records sent, 50024.6 records/sec (48.85 MB/sec), 1.5 ms avg latency, 18.0 ms max latency. 1000000 records sent, 49960.031974 records/sec (48.79 MB/sec), 1.89 ms avg latency, 199.00 ms max latency, 2 ms 50th, 4 ms 95th, 6 ms 99th, 18 ms 99.9th.
测试结果发现延迟又降低了，是不是感觉很强大😁
acks
指定分区中必须有多少个副本收到这条消息，才算消息发送成功，默认值1
如果配置acks=0还能降低一点点延迟，就是不等待broker返回是否成功，发出去就完了
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic two --record-size 1024 --num-records 1000000 --throughput 50000 249919 records sent, 49963.8 records/sec (48.79 MB/sec), 1.4 ms avg latency, 179.0 ms max latency. 250157 records sent, 50021.4 records/sec (48.85 MB/sec), 1.2 ms avg latency, 10.0 ms max latency. 250228 records sent, 50015.6 records/sec (48.84 MB/sec), 0.9 ms avg latency, 8.0 ms max latency. 1000000 records sent, 49967.521111 records/sec (48.80 MB/sec), 1.09 ms avg latency, 179.00 ms max latency, 1 ms 50th, 3 ms 95th, 4 ms 99th, 6 ms 99.9th.
通过测试上面几个参数，如果只配置其中一个，compression.type=gzip 效果是最好的
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic two --record-size 1024 --num-records 1000000 --throughput 50000 249882 records sent, 49956.4 records/sec (48.79 MB/sec), 11.9 ms avg latency, 191.0 ms max latency. 248708 records sent, 49731.7 records/sec (48.57 MB/sec), 2.9 ms avg latency, 92.0 ms max latency. 251380 records sent, 50276.0 records/sec (49.10 MB/sec), 2.0 ms avg latency, 23.0 ms max latency. 249980 records sent, 49996.0 records/sec (48.82 MB/sec), 1.5 ms avg latency, 18.0 ms max latency. 1000000 records sent, 49960.031974 records/sec (48.79 MB/sec), 4.55 ms avg latency, 191.00 ms max latency, 2 ms 50th, 12 ms 95th, 88 ms 99th, 163 ms 99.9th.
在当前环境下，平均延迟能只有4.55ms,最大延迟191ms
如上测试是在单机1分区，1副本的情况下的，为了能看到效果，延迟只是一个指标，但实际中并不是一味追求某个指标，还需要综合考虑，比如低延迟下，还要提高吞吐量，这就会要牺牲一部分的低延迟。不同的优化点，需要调整不同的参数，具体参数可以见 https://dwz.cn/Sl5L3zoq
另外：
如果Topic是多分区，也有显著效果，如果还需要降低延迟，可以再通过如上的参数进行优化
比如在当前环境下，我现在要达到10w的吞吐量，默认配置下是达不到的
[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic two --record-size 1024 --num-records 1000000 --throughput 100000 1 records sent, 0.1 records/sec (0.00 MB/sec), 7194.0 ms avg latency, 7194.0 ms max latency. 91167 records sent, 3306.3 records/sec (3.23 MB/sec), 519.4 ms avg latency, 26096.0 ms max latency. 330075 records sent, 66015.0 records/sec (64.47 MB/sec), 2843.5 ms avg latency, 26106.0 ms max latency. 227535 records sent, 45507.0 records/sec (44.44 MB/sec), 556.2 ms avg latency, 2306.0 ms max latency. 236940 records sent, 38577.0 records/sec (37.67 MB/sec), 522.0 ms avg latency, 3439.0 ms max latency. 1000000 records sent, 18762.078088 records/sec (18.32 MB/sec), 1402.18 ms avg latency, 26106.00 ms max latency, 443 ms 50th, 4018 ms 95th, 26073 ms 99th, 26095 ms 99.9th.
通过这几个配置
batch.size=204800 compression.type=gzip
就近乎达到了10w的吞吐量
C[root@10 kafka_2.11-2.2.0]# ./bin/kafka-producer-perf-test.sh --producer.config config/me.properties --topic tw --record-size 1024 --num-records 2000000 --throughput 100000 397998 records sent, 79599.6 records/sec (77.73 MB/sec), 3.4 ms avg latency, 193.0 ms max latency. 489610 records sent, 97922.0 records/sec (95.63 MB/sec), 2.5 ms avg latency, 24.0 ms max latency. 522791 records sent, 104558.2 records/sec (102.11 MB/sec), 1.8 ms avg latency, 29.0 ms max latency. 485255 records sent, 96973.4 records/sec (94.70 MB/sec), 1.8 ms avg latency, 26.0 ms max latency. 2000000 records sent, 94665.593790 records/sec (92.45 MB/sec), 2.31 ms avg latency, 193.00 ms max latency, 2
