pulsar-ack-1| 3.23 | 3.21 | 4.23 | 4.73 | 5.89 | pulsar-ack-2| 3.64 | 3.42 | 4.47 | 10.37 | 32.74 | kafka-ack-1| 2.54 | 1.99 | 3.78 | 11.37 | 18.71 | kafka-ack-all| 9.84 | 5.71 | 19.7 | 58.83 | 164.20"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"为了更深入地了解延迟如何随时间变化，我们绘制了 Pulsar 和 Kafka 采取不同复制持久性设置时的 P99 延迟图。如图 11 所示，Pulsar 的延迟比较稳定（约 5 毫秒），但 Kafka 的延迟波动较大。对于关键任务服务（mission-critical services）来说，稳定、持续的低延迟至关重要。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/9f\/5b\/9f49d0eabfeb36910b49b74a75d8f85b.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图11 Pulsar 和 Kafka 的 P99 延迟"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"端到端延迟--同步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用两种复制持久性设置（分别为 ack-1和 ack-2）和同步本地持久性时，Pulsar 和 Kafka 端到端延迟的差异如图 12。表 3 为每种情况下的精确延迟数值。在异步复制持久性（ack-1）下，Pulsar P99 端到端延迟是 Kafka 的 1\/3，在同步复制持久性（ack-2）下是 Kafka 的 1\/5。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/7d\/3a\/7d46b026cc1d195e535a47d5f1f42e3a.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图12 1 个订阅时，Pulsar 和 Kafka 的端到端延迟（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 3 1 个订阅时，Pulsar 和 Kafka 的实际端到端延迟测试结果(数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
pulsar-ack-1| 4.32 | 4.41 | 5.35 | 5.86 | 7.65 | pulsar-ack-2| 4.72 | 4.65 | 5.60 | 11.64 | 33.90 | kafka-ack-1| 6.23 | 4.91 | 9.08 | 18.75 | 91.74 | kafka-ack-all| 12.89 | 7.53 | 23.07 | 64.62 | 169.83"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"发布延迟 - 异步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用两种复制持久性设置（分别为 ack-1 和 ack-2）和异步本地持久性时，Pulsar 和 Kafka 发布延迟的差异如图 13。表 4 为每种情况下的精确延迟数值。在异步复制持久性（ack-1）下，Kafka 性能更优，但 Pulsar P99 发布延迟更稳定（低于 5 毫秒），增加复制持久性保证（从 ack-1 到 ack-2）并未影响延迟。在同步复制持久性（ack-2）下，"},{"type":"text","marks":[{"type":"strong"}],"text":"Kafka P99 发布延迟是 Pulsar 的 2 倍"},{"type":"text","text":"。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/98\/2b\/98e9918b1b4aee849a6d4d169942792b.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图13 Pulsar 和 Kafka 的发布延迟(无数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 4 Pulsar 和 Kafka 的实际发布延迟测试结果（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
pulsar-ack-1| 2.81 | 2.79 | 3.8 | 4.21 | 5.45 | pulsar-ack-2| 2.99 | 2.97 | 3.96 | 4.47 | 6.19 | kafka-ack-1| 1.74 | 1.62 | 2.60 | 3.06 | 9.91 | kafka-ack-all| 3.01 | 2.77 | 3.97 | 8.47 | 15.57"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"为了更深入地了解发布延迟如何随时间变化，我们绘制了 Pulsar 和 Kafka 在不同复制持久性设置下的 P99 发布延迟图。如图 14 所示，Pulsar 的延迟始终保持在较低的水平（低于 5 毫秒），"},{"type":"text","marks":[{"type":"strong"}],"text":"而 Kafka 的延迟始终是 Pulsar 的 2 倍"},{"type":"text","text":"。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/86\/ae\/86b22dda18605492ce4344370ae7f1ae.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图14 不同复制持久性设置下，Pulsar 和 Kafka 的 P99 发布延迟"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"端到端延迟--异步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用两种复制持久性设置（分别为 ack-1 和 ack-2）和异步本地持久性时，Pulsar 和 Kafka 端到端延迟的差异如图 15。表 5 为每种情况下的精确延迟数值。在所有情况下，Pulsar 的性能都优于 Kafka。Pulsar P99 端到端延迟保持稳定（约 5 毫秒），改变复制持久性设置对其并无影响。对于 ack-1，"},{"type":"text","marks":[{"type":"strong"}],"text":"Kafka 的 P99 端到端延迟比 Pulsar 高；对于 ack-2，Kafka 的 P99 端到端延迟是 Pulsar 的 2 倍"},{"type":"text","text":"。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/4f\/2f\/4fe43f34203a939fd78b4f0e9a94f22f.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图15 1 个订阅时，Pulsar 和 Kafka 的端到端延迟（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 5 1 个订阅时，Pulsar 和 Kafka 的实际端到端延迟测试结果（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
pulsar-ack-1| 3.96 | 3.99 | 4.90 | 5.33 | 7.93 | pulsar-ack-2| 4.06 | 4.17 | 5.08 | 5.55 | 8.52 | kafka-ack-1| 4.26 | 4.10 | 5.39 | 6.94 | 17.24 | kafka-ack-all| 4.22 | 3.96 | 5.19 | 10.43 | 18.95"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"#2 100 个分区, 10 个订阅"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"了解 Pulsar 和 Kafka 只有一个订阅时的性能后，我们想知道更改订阅数量会如何影响发布延迟和端到端延迟。于是我们把订阅数量从 1 增加到 10，并为每个订阅分配了 2 个消费者。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从表 6 的数据中可以看出，我们的测试结果表明："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Pulsar P99 发布延迟和端到端延迟保持在 5 到 10 毫秒之间。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"增加订阅数量会对 Kafka 的 P99 发布和端到端延迟产生巨大影响，导致延迟增加到几秒。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 6 10 个订阅时，发布和端到端延迟的测试结果"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 发布延迟 | 端到端延迟 |
|---|
同步本地持久性| 测试结果 | 测试结果 | 异步本地持久性| 测试结果 | 测试结果"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"发布延迟 - 同步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用两种复制持久性设置（分别为 ack-1和 ack-2）和同步本地持久性时，Pulsar 和 Kafka 发布延迟的差异如图 16。表 7 为每种情况下的精确延迟数值。在异步复制持久性（ack-1）下，Pulsar P99 发布延迟仍然是 Kafka 的 1\/3。但在同步复制持久性（ack-2）下，Pulsar 的发布延迟是 Kafka 的 1\/160（1 个订阅时，Pulsar 的发布延迟是 Kafka 的 1\/5）。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/16\/d6\/16a6e45bdyyfe063240c59aea7c308d6.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图16 10 个订阅时，Pulsar 和 Kafka 的发布延迟（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 7 10 个订阅时，Pulsar 和 Kafka 的实际发布延迟测试结果（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
pulsar-ack-1| 3.24 | 3.20 | 4.26 | 4.89 | 10.31 | pulsar-ack-2| 3.67 | 3.47 | 4.56 | 9.94 | 31.31 | pafka-ack-1| 3.14 | 2.39 | 4.39 | 15.07 | 61.29 | kafka-ack-all| 290.51 | 176.82 | 724.26 | 1593.46 | 2686.41"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"端到端延迟--同步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用两种复制持久性设置（分别为 ack-1和 ack-2）和同步本地持久性时，Pulsar 和 Kafka 端到端延迟的差异如图 17。表 8 为每种情况下的精确延迟数值。在异步复制持久性（ack-1）下，Pulsar P99 延迟是 Kafka 的 1\/20，在同步复制持久性（ack-2）下是 Kafka 的 1\/110 。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/ff\/49\/ffba1f49278e1531c4f353c8ae934b49.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图17 10 个订阅时，Pulsar 和 Kafka 的端到端延迟（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 8 10 个订阅时，Pulsar 和 Kafka 的实际端到端延迟测试结果(数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
pulsar-ack-1| 4.79 | 4.83 | 6.03 | 7.12 | 15.36 | pulsar-ack-2| 5.34 | 5.12 | 6.43 | 14.65 | 39.90 | kafka-ack-1| 11.36 | 6.65 | 17.12 | 145.10 | 914.19 | kafka-ack-all| 296.45 | 171.32 | 731.67 | 1599.79 | 2696.63"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"发布延迟 - 异步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用两种复制持久性设置（分别为 ack-1 和 ack-2）和异步本地持久性时，Pulsar 和 Kafka 发布延迟的差异如图 18。表 9 为每种情况下的精确延迟数值。Pulsar 的性能显著优于 Kafka。Pulsar 的平均发布延迟约为 3 毫秒，P99 延迟在 5 毫秒以内。Kafka 在异步复制持久性（ack-1）下性能令人满意，但在同步复制持久性（ack-2）下性能差了很多。在同步复制持久性下，Kafka 的 P99 发布延迟是 Pulsar 的 270 倍。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/9a\/2d\/9a63b1e6a8f914c8b23a00dca439872d.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图18 10 个订阅时，Pulsar 和 Kafka 的发布延迟（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 9 10 个订阅时，Pulsar 和 Kafka 的实际发布延迟测试结果(无数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
pulsar-ack-1| 2.86 | 2.82 | 3.86 | 4.46 | 11.18 | pulsar-ack-2| 3.05 | 3.00 | 4.03 | 4.73 | 10.39 | kafka-ack-1| 2.11 | 1.89 | 3.02 | 6.35 | 14.74 | kafka-ack-all| 158.04 | 17.63 | 526.91 | 1281.25 | 1956.71"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"端到端延迟--异步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用两种复制持久性设置（分别为 ack-1 和 ack-2）和异步本地持久性时，Pulsar 和 Kafka 端到端延迟的差异如图 19。表 10 为不同情况下的精确延迟数值。在所有情况下，Pulsar 的性能始终优于 Kafka。Pulsar 的端到端延迟始终保持在 4～7 毫秒之间，改变复制持久性设置对其并无影响。对于 ack-1，Kafka 的 P99 端到端延迟是 Pulsar 的 13 倍；对于ack-2，则是 187 倍。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/6a\/34\/6aaa289389c4a607949fbfdce178df34.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图19 10 个订阅时，Pulsar 和 Kafka 的端到端延迟（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 10 10 个订阅时，实际端到端延迟测试结果(无数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
pulsar-ack-1| 4.51 | 4.47 | 5.60 | 6.84 | 15.77 | pulsar-ack-2| 4.61 | 4.61 | 5.76 | 6.94 | 14.21 | kafka-ack-1| 8.01 | 5.90 | 9.38 | 89.80 | 532.68 | kafka-ack-all| 212.77 | 87.72 | 537.85 | 1295.78 | 1971.03"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"#3 100, 5000, 8000, 10000 个分区"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"了解更改订阅数量会如何影响 Pulsar 和 Kafka 的发布延迟后，我们想更改分区数量，看看会产生什么效果。于是，我们把分区数量从 100 个累加到 10000 个，观察发生了什么变化。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"测试数据见表 11，测试结果表明："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"分区数量增加时，Pulsar P99 发布延迟稳定在 5 毫秒内。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Kafka 的 P99 发布延迟受分区数量累加影响很大，延迟上升到几秒。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"分区数量超过 5000 后，Kafka 的消费者就跟不上发布吞吐量了。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 11 不同 ack 和不同持久性下的实际发布延迟测试结果"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | Ack = 1 | Ack = 2 |
|---|
| 同步本地持久性 | 测试结果 | 测试结果 | | 异步本地持久性 | 测试结果 | 测试结果"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Ack = 1，同步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用同步本地持久性和异步复制持久性 (ack = 1) 更改分区数量时，Pulsar 和 Kafka 发布和端到端延迟的差异分别如图 20 和图 21。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/0b\/e1\/0b6ec2e8afc745a2b96ab46f64c187e1.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图20 分区数量不同，ACK=1 时的发布延迟（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/8f\/11\/8f7669e4363a3476788ced67f9fd8511.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图21 分区数量不同，ACK=1 时的端到端延迟(数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"表 12 显示了分区数量不同，ACK=1 时的实际发布延迟测试结果。表 13 显示了分区数量不同，ACK=1 时的实际端到端延迟测试结果。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 12 分区数量不同，ACK=1 时的实际发布延迟测试结果（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
kafka-100| 2.54 | 1.99 | 3.78 | 11.37 | 18.71 | kafka-5000| 3.50 | 4.41 | 9.21 | 29.26 | 44.39 | kafka-8000| 8.37 | 5.11 | 20.18 | 40.70 | 69.72 | kafka-10000| 11.14 | 6.57 | 26.81 | 52.24 | 82.07 | pulsar-100| 3.23 | 3.21 | 4.23 | 4.73 | 5.89 | pulsar-5000| 3.35 | 3.30 | 4.34 | 5.03 | 13.96 | pulsar-8000| 3.67 | 3.67 | 4.89 | 5.61 | 16.07 | pulsar-10000| 3.42 | 3.37 | 4.48 | 5.36 | 19.20"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 13 分区数量不同，ACK=1 时的实际端到端延迟测试结果（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
kafka-100| 6.23 | 4.91 | 9.08 | 18.75 | 91.74 | pulsar-100| 4.32 | 4.41 | 5.35 | 5.86 | 7.65 | pulsar-5000| 4.52 | 4.53 | 5.55 | 6.26 | 17.78 | pulsar-8000| 4.89 | 4.99 | 6.11 | 6.86 | 23.83 | pulsar-10000| 4.49 | 4.62 | 5.70 | 6.67 | 27.25"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"图 22 给出了分区数量不同，ACK=1 时 Pulsar 端到端的延迟。图 23 给出了分区数量不同，ACK=1 时 Kafka 端到端的延迟。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/dd\/8e\/dd6fc4a2169b5f2c687dd677df95388e.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图22 分区数量不同，ACK=1 时 Pulsar 的端到端延迟（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/f6\/92\/f63a2195bc40c587529aa19b313f5a92.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图23 分区数量不同，ACK=1 时 Kafka 的端到端延迟(数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从以上图表可以看出："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Pulsar P99 发布延迟稳定在 5 毫秒左右，更改分区数量对其并无影响。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Pulsar P99 端到端延迟稳定在 6 毫秒左右，更改分区数量对其并无影响。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Kafka 的 P99 发布延迟随着分区数量增加而逐渐拉长，在 10000 个分区时，其延迟增加了 5 倍（与 100 个分区时相比），比 Pulsar 的延迟高出 10 倍。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Kafka 的 P99 端到端延迟随着分区数量增加而逐渐拉长，在 10000 个分区时，其延迟高出 10000 倍（与 100 个分区时相比），增加到 180 秒，是 Pulsar 的 280000倍。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Ack = 2，同步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用同步本地持久性和同步复制持久性（ack = 2）更改分区数量时，Pulsar 和 Kafka 发布延迟的差异如图 24。表 14 为每种情况下的精切延迟数值。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/fc\/4e\/fc16d7dbb7df1dec268914a7aa68c34e.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图24 分区数量不同，ack=2\/all 时的发布延迟（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 14 分区数量不同， ack=2\/all 时的实际发布延迟测试结果（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
kafka-100| 9.84 | 5.71 | 19.7 | 58.83 | 164.20 | kafka-5000| 154.50 | 36.86 | 468.75 | 1259.82 | 2017.15 | kafka-8000| 283.50 | 124.82 | 784.69 | 1742.27 | 2729.79 | kafka-10000| 259.50 | 96.15 | 731.85 | 1718.09 | 2684.28 | pulsar-100| 3.64 | 3.42 | 4.47 | 10.37 | 32.74 | pulsar-5000| 3.84 | 3.57 | 4.69 | 12.81 | 42.83 | pulsar-8000| 4.15 | 3.97 | 5.23 | 11.90 | 42.19"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"图 25 和图 26 分别显示了更改分区数量如何影响 Pulsar 和 Kafka 端到端的延迟。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/59\/45\/59cfbe12e99465c501510b4e5428e945.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图25 分区数量不同，ack=2 时 Pulsar 的端到端延迟（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/f4\/40\/f45a915c2473d850fb3c775939ff1c40.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图26 分区数量不同，ack=2 时 Kafka 的端到端延迟（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从以上图表可以看出："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Pulsar P99 发布延迟稳定在约 10 毫秒，增加分区数量对其并无影响。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"随着分区数量增加，Kafka 的 P99 发布延迟逐渐拉长；有 10000 个分区时，其延迟是 100 个分区的 30 倍，增加到 1.7 秒，是 Pulsar 的 126 倍。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Pulsar P99 端到端延迟稳定在约 10 毫秒，增加分区数量仅对 Pulsar P99 端到端延迟产生轻微影响。但即使有 10000 个分区，延迟仍相对较低，约为 50 毫秒。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"随着分区数量增加，Kafka 的 P99 端到端延迟逐渐拉长。有 10000 个分区时，Kafka P99 端到端延迟增加到 200 秒，是 Pulsar 的 14771 倍。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Ack = 1, Async 本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用异步本地持久性和异步复制持久性（ack = 1）更改分区数量时，Pulsar 和 Kafka 发布延迟的差异如图 27。表 15 为不同情况下的精切延迟数值。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/0c\/d6\/0c02b4488399e82ce8e6f1674f0038d6.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图27 分区数量不同，ack=1 时的发布延迟（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 15 分区数量不同，ack=1 时的实际发布延迟测试结果（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
kafka-100| 1.74 | 1.62 | 2.60 | 3.06 | 9.91 | kafka-5000| 4.13 | 3.05 | 5.20 | 25.32 | 33.86 | kafka-8000| 6.84 | 4.32 | 16.51 | 34.19 | 46.34 | kafka-10000| 8.95 | 5.32 | 22.75 | 41.83 | 59.32 | pulsar-100| 2.86 | 2.79 | 3.8 | 4.21 | 5.45 | pulsar-5000| 2.98 | 2.95 | 3.89 | 4.51 | 7.85 | pulsar-8000| 3.26 | 3.27 | 4.50 | 5.14 | 15.21 | pulsar-10000| 3.06 | 2.97 | 4.15 | 5.23 | 19.47"}}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"图 28 和图 29 分别显示了更改分区数量如何影响 Pulsar 和 Kafka 端到端的延迟。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/32\/ac\/329f3651ef8bfe94c210c402702101ac.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图28 分区数量不同，ack=1 时，Pulsar 端到端的延迟（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/96\/98\/96e667ac0efd732ffa0yyc0a0448d598.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图29 分区数量不同，ack=1 时，Kafka 端到端的延迟(无数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从以上图表可以看出："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Pulsar P99 发布延迟稳定在约 4～5 毫秒之间，增加分区数量对其并无影响。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Kafka P99 发布延迟随着分区数量增加而逐渐拉长，有 10000 个分区时，延迟增加到 41 毫秒，是有 100 个分区时的 13 倍，是 Pulsar 的 8 倍。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Pulsar P99 端到端延迟稳定在约 4～6 毫秒之间，增加分区数量仅对 Pulsar P99 端到端延迟产生轻微影响，但即使有 10000 个分区，它也保持在相对较低的水平（24 毫秒以内）。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"随着分区数量增加，Kafka P99 端到端延迟逐渐拉长。有 10000 个分区时，Kafka P99 端到端延迟高达 180 秒，是 Pulsar 的 34416 倍。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Ack = 2，异步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用异步本地持久性和同步复制持久性 (ack = 2)更改分区数量时，Pulsar 和 Kafka 发布延迟的差异如图 30。表 16 为每种情况下的精确延迟数值。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/a7\/e9\/a7a1f13c29956babf847bdb99b6cf0e9.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图30 分区数量不同，ack=all\/2 的发布延迟（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 16 分区数量不同，ack=all\/2 的实际发布延迟测试结果（无数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":" | 平均值 | P50 | P90 | P99 | P999 |
|---|
kafka-100| 3.01 | 2.77 | 3.97 | 8.47 | 15.57 | kafka-5000| 19.33 | 10.16 | 40.40 | 121.40 | 336.21 | kafka-8000| 138.19 | 42.52 | 385.86 | 1164.90 | 2008.28 | kafka-10000| 266.66 | 102.55 | 752.95 | 1717.83 | 2797.51 | pulsar-100| 2.99 | 2.97 | 3.96 | 4.47 | 6.19 | | pulsar-5000 | | | | | |
| | | | | | | | |
| | | | | | | |
| | | | | |
| | | | | | | | |
|
| | | | |
| | | | |
| | | | |
| | | | |
| | |
| | | | |
| | | | |
| | | | |
| | | | |
{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"为了更全面地了解 Pulsar 和 Kafka，我们“复现”了 Confluent 对 Pulsar 和 Kafka 基准测试。重复这一基准测试的原因有两个，一是 Confluent 的测试方法存在一些问题；二是 Confluent 的测试范围和测试场景不够全面。为了更准确地对比 Pulsar 和 Kafka，我们在测试中不仅修复了 Confluent 测试中的问题，还扩大了测试范围，纳入更多性能衡量标准，模拟更多实际场景。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"和 Confluent 的测试相比，我们的测试主要有三项改进："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":1,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"包含 Pulsar 和 Kafka 支持的所有持久性级别。在同等持久性级别下，对比二者的吞吐量和延迟。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"引入影响性能的其他因素和测试条件，如分区数量、订阅数量、客户端数量等。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"测试的混合负载同时包含写入、追赶读和追尾读，模拟实际使用场景。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们进行了最大吞吐量测试、发布和端到端延迟测试、追赶读测试和混合工作负载测试（"},{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/article\/HS4gtkQjop4q42z13uEv","title":"xxx","type":null},"content":[{"type":"text","text":"详见上篇"}]},{"type":"text","text":"）。由于篇幅有限，本篇主要介绍发布和端到端延迟测试详情。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"发布和端到端延迟测试"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"测试目标：观测在处理发布和追尾读工作负载时 Pulsar 和 Kafka 可实现的最低延迟。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"测试设置：通过调整订阅数量和分区数量，观测订阅数量和分区数量对发布延迟和端到端延迟的影响。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"测试策略："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将所有消息都复制三次，确保容错；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"改变 ack 数量，测试在不同持久性保证下，Pulsar 和 Kafka 吞吐量的差异；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"改变订阅数量（从 1 到 10），分别测试延迟；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"改变分区数量（从 100 到 10000），分别测试延迟；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"消息大小为 1KB；"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"生产者以 200000\/s（约 200MB\/s）的固定速度发送消息，追尾读消费者在生产者持续发送消息时处理消息。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"各项测试结果如下。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"#1 100 个分区, 1 个订阅"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们从 100 个分区、1 个订阅开始，在所有不同持久性保证下，对 Pulsar 和 Kafka 能实现的最低延迟进行了基准测试。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"测试证明，在所有持久性级别上，Pulsar 的发布延迟和端到端延迟都比 Kafka 低，是 Kafka 的 1\/5 到 1\/2。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"italic"}],"text":"发布延迟 - 同步本地持久性"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"使用两种复制持久性设置（分别为 ack-1 和 ack-2）和同步本地持久性时，Pulsar 和 Kafka 发布延迟的差异如图 10。表 2 为每种情况下的精确延迟数值。在异步复制持久性（ack-1）下，Pulsar P99 延迟是 Kafka 的 1\/3; 在同步复制持久性（ack-2）下是 Kafka 的 1\/5。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/c8\/1c\/c8828bd53b91aca9a765f5291d20621c.png","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"图10 Pulsar 和 Kafka 的发布延迟(数据同步)"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"表 2 Pulsar 和 Kafka 的实际发布延迟测试结果（数据同步）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"embedcomp","attrs":{"type":"table","data":{"content":"