chore(query): turn on new agg hashtable

[Freejww]

I hereby agree to the terms of the CLA available at: https://docs.databend.com/dev/policies/cla/

Summary

In the previous PR, we implemented a new aggregation hash table. Now, it supports both singleton and cluster environments, and we have also added support for spill. This PR includes performance tests and try to enable the new aggregation hash table.

• Fixes #[Link the issue here]

Tests

• Unit Test
• Logic Test
• Benchmark Test
• No Test - Explain why

Type of change

• Bug Fix (non-breaking change which fixes an issue)
• New Feature (non-breaking change which adds functionality)
• Breaking Change (fix or feature that could cause existing functionality not to work as expected)
• Documentation Update
• Refactoring
• Performance Improvement
• Other (please describe):

[github-actions]

Docker Image for PR

• tag: pr-15155-a3d050a

note: this image tag is only available for internal use,
please check the internal doc for more details.

[Freejww]

Performance test on ClickBench/hits

Note: We use hits as a benchmark because there are many group by tests.

Local machine configuration: 32GB memory, 12-core CPU.

1. The new aggregation hash table has completely outperformed the old method.
2. In some high-cardinality or string grouping , the new method outperforms the old one.
3. More details will be introduced in a later blog.

Query	Local singleton	main	pr	improve		Local cluster two nodes	main	pr	improve		Local cluster three nodes	main	pr	improve		Cloud small	main	pr	improve		Cloud medium	main	pr	improve
Q1		0.006s	0.006s	0%			0.008s	0.008s	0%			0.012s	0.012s	0%			0.033s	0.032s	3%			0.025s	0.024s	4%
Q2		0.059s	0.059s	0%			0.079s	0.078s	1%			0.084s	0.083s	1%			0.168s	0.106s	37%			0.109s	0.079s	28%
Q3		0.17s	0.161s	5%			0.175s	0.182s	-4%			0.189s	0.186s	2%			0.139s	0.117s	16%			0.172s	0.115s	33%
Q4		0.191s	0.185s	3%			0.199s	0.208s	-5%			0.21s	0.206s	2%			0.338s	0.283s	16%			0.345s	0.318s	8%
Q5		0.544s	0.542s	0%			0.885s	0.811s	8%			1.075s	1.049s	2%			0.423s	0.419s	1%			0.571s	0.513s	10%
Q6		0.932s	0.858s	8%			1.381s	1.3s	6%			1.714s	1.303s	24%			0.666s	0.592s	11%			0.75s	0.631s	16%
Q7		0.049s	0.049s	0%			0.07s	0.07s	0%			0.067s	0.067s	0%			0.092s	0.092s	0%			0.099s	0.099s	0%
Q8		0.061s	0.061s	0%			0.081s	0.088s	-9%			0.097s	0.102s	-5%			0.089s	0.082s	8%			0.111s	0.086s	23%
Q9		0.862s	0.872s	-1%			1.11s	1.102s	1%			1.336s	1.207s	10%			0.604s	0.673s	-11%			0.782s	0.707s	10%
Q10		1.023s	1.044s	-1%			1.323s	1.274s	4%			1.618s	1.376s	15%			0.73s	0.802s	-10%			0.938s	0.717s	24%
Q11		0.415s	0.413s	0%			0.488s	0.48s	2%			0.557s	0.504s	10%			0.424s	0.398s	6%			0.472s	0.435s	8%
Q12		0.465s	0.461s	1%			0.54s	0.528s	2%			0.613s	0.572s	7%			0.402s	0.38s	5%			0.441s	0.405s	8%
Q13		1.081s	0.881s	19%			1.652s	1.416s	14%			1.995s	1.417s	29%			0.722s	0.533s	26%			0.838s	0.686s	18%
Q14		1.837s	1.424s	22%			2.514s	2.041s	19%			2.909s	2.122s	27%			1.158s	0.864s	25%			1.022s	0.934s	9%
Q15		1.279s	0.983s	23%			1.876s	1.536s	18%			2.27s	1.563s	31%			0.817s	0.603s	26%			0.885s	0.711s	20%
Q16		0.927s	0.696s	25%			1.501s	1.162s	23%			1.737s	1.393s	20%			0.563s	0.461s	18%			0.752s	0.699s	7%
Q17		3.03s	1.62s	47%			4.154s	2.631s	37%			4.361s	2.934s	33%			1.714s	1.049s	39%			1.435s	1.047s	27%
Q18		1.663s	0.969s	42%			1.757s	1.2s	32%			1.6s	1.313s	18%			1.084s	0.777s	28%			0.542s	0.519s	4%
Q19		6.223s	2.737s	56%			8.269s	4.699s	43%			9.316s	5.06s	46%			3.326s	1.872s	44%			2.515s	1.642s	35%
Q20		0.007s	0.007s	0%			0.04s	0.04s	0%			0.016s	0.016s	0%			0.047s	0.047s	0%			0.06s	0.06s	0%
Q21		2.497s	2.444s	2%			2.614s	2.566s	2%			2.682s	2.669s	0%			1.515s	1.512s	0%			1.109s	0.813s	27%
Q22		2.819s	2.795s	1%			2.958s	2.941s	1%			3.082s	3.082s	0%			1.821s	1.82s	0%			1.597s	0.998s	38%
Q23		5.848s	5.808s	1%			6.087s	6.007s	1%			6.488s	6.269s	3%			4.036s	3.95s	2%			2.915s	2.161s	26%
Q24		3.412s	3.398	1%			3.223s	3.204s	1%			3.441s	3.392s	1%			2.551s	2.566s	-1%			2.693s	2.524s	6%
Q25		0.718s	0.721s	-4%			0.738s	0.734s	1%			0.783s	0.783s	0%			0.755s	0.726s	4%			0.79s	0.447s	43%
Q26		0.53s	0.536s	-1%			0.515s	0.511s	1%			0.564s	0.547s	3%			0.388s	0.38s	2%			0.388s	0.336s	13%
Q27		0.769s	0.774s	-1%			0.775s	0.768s	1%			0.825s	0.804s	3%			0.748s	0.743s	1%			0.8s	0.462s	42%
Q28		3.177s	3.15s	1%			3.187s	3.186s	0%			3.263s	3.256s	0%			1.623s	1.652s	-2%			1.18s3	0.949s	20%
Q29		4.359s	4.394s	-1%			4.730s	4.701s	1%			5.426s	4.714s	13%			3.074s	3.033s	1%			2.02s6	1.833s	10%
Q30		0.137s	0.137s	0%			0.14s	0.14s	0%			0.147s	0.147s	0%			0.169s	0.169s	0%			0.168s	0.168s	0%
Q31		1.097s	0.993s	9%			1.519s	1.413s	7%			1.823s	1.386s	24%			0.843s	0.74s	12%			0.882s	0.692s	22%
Q32		1.72s	1.296s	25%			2.259s	1.946s	14%			2.586s	1.897s	27%			1.331s	1.099s	17%			1.349s	1.181s	12%
Q33		9.633s	3.772s	61%			13.7s	7.863s	43%			19.029s	8.305s	56%			4.258s	2.322s	45%			3.62s	2.352s	35%
Q34		5.485s	5.125s	7%			8.278s	7.39s	11%			8.982s	8.419s	6%			2.96s	2.412s	19%			2.936s	2.698s	8%
Q35		5.45s	5.106s	6%			8.361s	7.348s	12%			8.968s	8.478s	5%			2.963s	2.483s	16%			2.996s	2.682s	10%
Q36		0.738s	0.513s	30%			1.104s	0.908s	18%			1.406s	1.07s	24%			0.491s	0.41s	16%			0.721s	0.704s	2%
Q37		0.147s	0.138s	6%			0.244s	0.17s	30%			0.313s	0.192s	39%			0.167s	0.115s	31%			0.242s	0.184s	24%
Q38		0.136s	0.136s	0%			0.159s	0.159s	0%			0.167s	0.181s	-8%			0.124s	0.121s	2%			0.131s	0.131s	0%
Q39		0.103s	0.107s	-4%			0.12s	0.121s	-1%			0.124s	0.126s	-1%			0.105s	0.104s	1%			0.118s	0.117s	1%
Q40		0.281s	0.254s	10%			0.432s	0.33s	24%			0.493s	0.33s	33%			0.236s	0.203s	14%			0.326s	0.224s	31%
Q41		0.04s	0.039s	3%			0.08s	0.062s	23%			0.068s	0.066s	3%			0.086s	0.083s	3%			0.097s	0.075s	23%
Q42		0.045s	0.037s	18%			0.058s	0.059s	-2%			0.063s	0.064s	-1%			0.08s	0.073s	9%			0.083s	0.069s	17%
Q43		0.032s	0.032s	0%			0.051s	0.054s	-6%			0.06s	0.052s	17%			0.075s	0.072s	4%			0.079s	0.068s	14%

[github-actions]

Docker Image for PR

• tag: pr-15155-845b9c7

note: this image tag is only available for internal use,
please check the internal doc for more details.

[sundy-li]

There are explain tests that need to be fixed.

[sundy-li]

We use random and fixed SQL queries to cover the correctness tests.

Test Codes in https://github.com/sundy-li/mmbend/tree/master/examples

[BohuTANG]

Great 👍

Wizard [SELECTS] tests passed, with results compared to Snowflake. SQL selects script here: https://github.com/datafuselabs/wizard/blob/main/checksb/sql/selects/check.sql

How to run the tests:

python3 checksb.py --database selects --warehouse COMPUTE_WH --case selects

[github-actions]

ClickBench Report

• hits: https://benchmark.databend.rs/clickbench/pr/15155/8521796985/hits.html
• internal: https://benchmark.databend.rs/clickbench/pr/15155/8521796985/internal.html
• load: https://benchmark.databend.rs/clickbench/pr/15155/8521796985/load.html
• tpch: https://benchmark.databend.rs/clickbench/pr/15155/8521796985/tpch.html

[compasses]

May I know the test carried on memory limited ?

Performance test on ClickBench/hits

[sundy-li]

May I know the test carried on memory limited ?

No major difference if blocks are spilled, because the bottleneck is io issue.

[compasses]

got thanks
So I think this kind of implementation may not efficient than general-purpose hash table like Clickhouse did.

[sundy-li]

So I think this kind of implementation may not efficient than general-purpose hash table like Clickhouse did.

I disagree with that. The older one is the general-purpose hashtable you indicate.

[compasses]

Oh, OK, I mean without IO effect. Since most of agg workloads are memory-bound. And the way of implementation above seems need indirection access? Or maybe I missed something, but memory bandwidth is important for high cardinality keys.

[Freejww]

Oh, OK, I mean without IO effect. Since most of agg workloads are memory-bound. And the way of implementation above seems need indirection access? Or maybe I missed something, but memory bandwidth is important for high cardinality keys.

Indirect access is that we only compare the high 16 bits of the hash. If they match, then we proceed to compare the actual data. General-purpose hashtable compares the entire hash value before proceeding to compare the actual data.

[compasses]

OK got you thanks. Since the general-purpose hashtable will compare hash values if they saved the hash, if not just use it for probing, but any way the memory layout is more cache-friendly.