Re: parallel distinct union and aggregate support patch

Hi
Here is patch for parallel distinct union aggregate and grouping sets support using batch hash agg.
Please review.

how to use:
set enable_batch_hashagg = on

how to work:
like batch sort, but not sort each batch, just save hash value in each rows

unfinished work:
not support rescan yet. welcome to add. Actually I don't really understand how rescan works in parallel mode.

other:
patch 1 base on branch master(80f8eb79e24d9b7963eaf17ce846667e2c6b6e6f)
patch 1 and 2 see https://www.postgresql.org/message-id/2020101922424962544053%40sohu.com
patch 3:
extpand shared tuple store and add batch store module.
By the way, use atomic operations instead LWLock for shared tuple store get next read page.
patch 4:
using batch hash agg support parallels

发件人： bucoo(at)sohu(dot)com
发送时间： 2020-10-19 22:42
收件人： pgsql-hackers
主题： parallel distinct union and aggregate support patch
Hi hackers,
I write a path for soupport parallel distinct, union and aggregate using batch sort.
steps:
1. generate hash value for group clauses values, and using mod hash value save to batch
2. end of outer plan, wait all other workers finish write to batch
3. echo worker get a unique batch number, call tuplesort_performsort() function finish this batch sort
4. return row for this batch
5. if not end of all batchs, got step 3

BatchSort paln make sure same tuple(group clause) return in same range, so Unique(or GroupAggregate) plan can work.

path 2 for parallel aggregate, this is a simple use
but regress failed for partitionwise aggregation difference plan
from GatherMerge->Sort->Append->...
to Sort->Gahter->Append->...
I have no idea how to modify it.

Same idea I writed a batch shared tuple store for HashAgg in our PG version, I will send patch for PG14 when I finish it.

The following is a description in Chinese
英语不好，所以这里写点中文，希望上面写的不对的地方请大家帮忙纠正一下。
BatchSort的工作原理
1. 先按group clause计算出hash值,并按取模的值放入不同的批次
2. 当下层plan返回所有的行后,等待所有其它的工作进程结束
3. 每一个工作进程索取一个唯一的一个批次, 并调用tuplesort_performsort()函数完成最终排序
4. 返回本批次的所有行
5. 如果所有的批次没有读完，则返回第3步
BatchSort plan能保证相同的数据(按分给表达式)在同一个周期内返回，所以几个去重和分组相关的plan可以正常工作。
第2个补丁是支持并行分组的，只做一次分组，而不是并行进程做每一次分组后，主进程再进行二次分组。
这个补丁导致了regress测试中的partitionwise aggregation失败，原来的执行计划有所变更。
补丁只写了一个简单的使用BatchSort plan的方法，可能还需要添加其它用法。

用同样的思想我写了一个使用shared tuple store的HashAgg在我们的AntDB版本中(最新版本暂未开源），适配完PG14版本后我会发出来。
打个广告：欢迎关注我们亚信公司基于PG的分布式数据库产品AntDB，开源地址 https://github.com/ADBSQL/AntDB

bucoo(at)sohu(dot)com