Lazy build join output to improve performance of ALL join

[liuneng1994]

Changelog category (leave one):

• Performance Improvement

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Lazy build join output to improve performance of ALL join

Information about CI checks: https://clickhouse.com/docs/en/development/continuous-integration/

[liuneng1994]

I compare #54662 and #56996 using intel vtune. The main performance difference is front-end bound. From the code point of view, delaying output generation can improve the cache hit rate of L1.
@vdimir

[robot-ch-test-poll4]

This is an automated comment for commit 4af3395 with description of existing statuses. It's updated for the latest CI running

❌ Click here to open a full report in a separate page

Successful checks

Check name	Description	Status
A Sync	There's no description for the check yet, please add it to tests/ci/ci_config.py:CHECK_DESCRIPTIONS	✅ success
ClickBench	Runs [ClickBench](https://github.com/ClickHouse/ClickBench/) with instant-attach table	✅ success
ClickHouse build check	Builds ClickHouse in various configurations for use in further steps. You have to fix the builds that fail. Build logs often has enough information to fix the error, but you might have to reproduce the failure locally. The cmake options can be found in the build log, grepping for cmake. Use these options and follow the general build process	✅ success
Compatibility check	Checks that clickhouse binary runs on distributions with old libc versions. If it fails, ask a maintainer for help	✅ success
Docker keeper image	The check to build and optionally push the mentioned image to docker hub	✅ success
Docker server image	The check to build and optionally push the mentioned image to docker hub	✅ success
Docs check	Builds and tests the documentation	✅ success
Fast test	Normally this is the first check that is ran for a PR. It builds ClickHouse and runs most of stateless functional tests, omitting some. If it fails, further checks are not started until it is fixed. Look at the report to see which tests fail, then reproduce the failure locally as described here	✅ success
Flaky tests	Checks if new added or modified tests are flaky by running them repeatedly, in parallel, with more randomization. Functional tests are run 100 times with address sanitizer, and additional randomization of thread scheduling. Integrational tests are run up to 10 times. If at least once a new test has failed, or was too long, this check will be red. We don't allow flaky tests, read the doc	✅ success
Install packages	Checks that the built packages are installable in a clear environment	✅ success
Mergeable Check	Checks if all other necessary checks are successful	✅ success
Performance Comparison	Measure changes in query performance. The performance test report is described in detail here. In square brackets are the optional part/total tests	✅ success
SQLTest	There's no description for the check yet, please add it to tests/ci/ci_config.py:CHECK_DESCRIPTIONS	✅ success
SQLancer	Fuzzing tests that detect logical bugs with SQLancer tool	✅ success
Sqllogic	Run clickhouse on the sqllogic test set against sqlite and checks that all statements are passed	✅ success
Stateless tests	Runs stateless functional tests for ClickHouse binaries built in various configurations -- release, debug, with sanitizers, etc	✅ success
Stress test	Runs stateless functional tests concurrently from several clients to detect concurrency-related errors	✅ success
Style check	Runs a set of checks to keep the code style clean. If some of tests failed, see the related log from the report	✅ success
Unit tests	Runs the unit tests for different release types	✅ success
Upgrade check	Runs stress tests on server version from last release and then tries to upgrade it to the version from the PR. It checks if the new server can successfully startup without any errors, crashes or sanitizer asserts	✅ success

Check name	Description	Status
AST fuzzer	Runs randomly generated queries to catch program errors. The build type is optionally given in parenthesis. If it fails, ask a maintainer for help	❌ failure
CI running	A meta-check that indicates the running CI. Normally, it's in success or pending state. The failed status indicates some problems with the PR	⏳ pending
Integration tests	The integration tests report. In parenthesis the package type is given, and in square brackets are the optional part/total tests	❌ failure
Stateful tests	Runs stateful functional tests for ClickHouse binaries built in various configurations -- release, debug, with sanitizers, etc	❌ failure

[liuneng1994]

Aarch64


X86-64


A very strange phenomenon. In the previous version, the performance of ANY join on Aarch64 was reduced by 30%. In the current version of Aarch64, there is no regression, but the performance is reduced by 10% on x86.
I can't reproduce this regression on my machine (12900k), tried many ways. I added a new any join test case, and there were no performance problems after testing. So I suspect it's due to the specificity of the use case.
From an overall point of view, this optimization should have more advantages than disadvantages.
@vdimir

[baibaichen]

@alexey-milovidov @vdimir any comments?

[vdimir]

Sorry it slipped my mind.
In general looks reasonable. Maybe we can also introduce a way to support old behaviour, so if we found any cases with degradation we could fall back? But I'm not sure how to fall back, because introducing a setting also doesn't look practical as well.

[binmahone]

insert is a virtual function , it maybe heavy to call it repeatly in a loop. Will sth like insertSelective() helps perf here?

[vdimir]

Not sure if it's bottleneck or not, but it's worth trying, perhaps in another PR

[liuneng1994]

Sorry it slipped my mind. In general looks reasonable. Maybe we can also introduce a way to support old behaviour, so if we found any cases with degradation we could fall back? But I'm not sure how to fall back, because introducing a setting also doesn't look practical as well.

I have previously verified that as long as any join uses the original algorithm, the performance regression problem can be solved. I can add a subclass LazyAddedColumns to AddedColumns, and select which one to use in HashJoin::joinBlockImpl based on whether the join type is any join. Is this plan acceptable?

[vdimir]

I can add a subclass LazyAddedColumns to AddedColumns, and select which one to use in HashJoin::joinBlockImpl based on whether the join type is any join. Is this plan acceptable?

Yes, sounds good 👍

[liuneng1994]

Virtual function calls on AddedColumns can severely impact performance.
Using an insertSelective call in build_output instead of calling insert in a for loop can reduce a large number of virtual function calls, but the premise of doing so is to merge the original data in the hashmap into a large block. RowRef only saves the row number, not the block pointer

[liuneng1994]

[liuneng1994]

@vdimir Ready for review. The performance regression problem of any join has been solved.
Fortunately, I found that this version of the code also greatly improved some use cases in joins_in_memory.


Multiple CI results have shown performance improvement. The specific reasons for the improvement need to be analyzed. The core code hasn't changed much.

[liuneng1994]

@vdimir any comments?

[vdimir]

I assume is_join_get = false in all cases, because we call joinBlockImpl with Any for dictGet. But I'm not 100% sure, also if something changed we will get a wrong behavior, so we can keep this if.

[liuneng1994]

I don’t particularly understand the logic here. In order not to cause unnecessary bugs, I have retained this logic.

[vdimir]

Thank you for your effort!
I'll merge a PR once A Sync Pending is finished