tidb hash join优化版是什么意思

residentevil · 2025 年1 月 17 日 06:01

【 TiDB 使用环境】生产环境
【 TiDB 版本】

Kongdom · 2025 年1 月 17 日 06:26

感觉就是做了优化，但不确定有没有bug，所以加了个开关，默认不开启优化，但是可以通过修改变动启用优化，以验证优化效果。

相关issues
https://github.com/pingcap/tidb/issues/55153
https://github.com/pingcap/tidb/issues/53127

小龙虾爱大龙虾 · 2025 年1 月 17 日 06:49

就是对 hash join 算法实现做了优化，这里留了开关

as波哥 · 2025 年1 月 18 日 03:48

对 hash join 算法实现做了优化，这里留了开关

residentevil · 2025 年1 月 18 日 05:34

官方没有对这个优化做更详细的讲解，哈哈

YuchongXU · 2025 年1 月 18 日 07:43

这个只能看代码了

zhaokede · 2025 年1 月 18 日 09:02

可以尝试评估开启和关闭两者的性能差

有猫万事足 · 2025 年1 月 18 日 11:21

github.com

pingcap/tidb/blob/master/docs/design/2024-05-11-hash-join-v2.md

# Proposal: Hash Join V2

- Author(s): [windtalker](https://github.com/windtalker)
- Tracking Issue: https://github.com/pingcap/tidb/issues/53127

## Introduction 
Hash join is a widely used join algorithm. TiDB supports hash join since its 1.0 version, however, the current implementation 
of hash join has some shortcomings:
* At build stage, it does not support concurrent build, which may lead some performance issues when the build side is large. 
* At probe stage, the interface is not well-designed, which may cause redundant calculations in some cases: https://github.com/pingcap/tidb/issues/47424
* In current implementation, there are some concepts that are actually not found in other database's implementation. For example, beside left side/right side and build side/right side, there is inner side/outer side , and `useOuterToBuild`, `outerIsRight`, `tryToMatchInners`, `tryToMatchOuters` are introduced to handle the extra complex. This makes the current code too complex to understand and more error-prone when we try to fix bug.

Taking into account the above factors, we decided to do a complete refactoring of hash join.

## Problems

The basic idea of hash join is to divide the join into build stage and probe stage. On the build stage, a hash table is built 
based on the join key, on the probe stage, a lookup is made in the hash table using the join key, and the join result is 
generated based on the lookup result and the join type. The problems faced in the build stage mainly include the design of 
the hash table, the data organization on the build side and the concurrent build of the hash table. The problems faced in

This file has been truncated. show original

设计文档在这里

TiDBer_小杰 · 2025 年1 月 18 日 12:15

有哪个大佬做个这个hash join 算发开关测试吗

residentevil · 2025 年1 月 19 日 01:48

借鉴下

X_Y_Z · 2025 年1 月 20 日 09:03

哪位大佬测试一下，分享研究一下

residentevil · 2025 年1 月 20 日 09:12

看了半天的社区v2文档，感觉跟社区80的差异不大

system · 2025 年1 月 27 日 11:36

此话题已在最后回复的 7 天后被自动关闭。不再允许新回复。