proposal of fixing vttablet container OOMKilled because consolidator

jwangace commented 5 days ago

Description

This fix use config ConsolidatorMaxQueryWait to limit the max number of Wait() for a consolidated query, more same query will still getConn() where is protected by connection pools.

With configureable ConsolidatorMaxQueryWait (default to 0 means unlimit, user can adjust this value accordingly), this will prevent when the Result of first query been returned, same Result would be sent concurrently to large number of clients, cause memory abruptly went high.

Related Issue(s)

https://github.com/vitessio/vitess/issues/17243

Checklist

[ ] "Backport to:" labels have been added if this change should be back-ported to release branches
[ ] If this change is to be back-ported to previous releases, a justification is included in the PR description
[ ] Tests were added or are not required
[ ] Did the new or modified tests pass consistently locally and on CI?
[ ] Documentation was added or is not required

Deployment Notes

vitess-bot[bot] commented 5 days ago

Review Checklist

Hello reviewers! :wave: Please follow this checklist when reviewing this Pull Request.

General

[ ] Ensure that the Pull Request has a descriptive title.
[ ] Ensure there is a link to an issue (except for internal cleanup and flaky test fixes), new features should have an RFC that documents use cases and test cases.

Tests

[ ] Bug fixes should have at least one unit or end-to-end test, enhancement and new features should have a sufficient number of tests.

Documentation

[ ] Apply the release notes (needs details) label if users need to know about this change.
[ ] New features should be documented.
[ ] There should be some code comments as to why things are implemented the way they are.
[ ] There should be a comment at the top of each new or modified test to explain what the test does.

New flags

[ ] Is this flag really necessary?
[ ] Flag names must be clear and intuitive, use dashes (-), and have a clear help text.

If a workflow is added or modified:

[ ] Each item in Jobs should be named in order to mark it as required.
[ ] If the workflow needs to be marked as required, the maintainer team must be notified.

Backward compatibility

[ ] Protobuf changes should be wire-compatible.
[ ] Changes to _vt tables and RPCs need to be backward compatible.
[ ] RPC changes should be compatible with vitess-operator
[ ] If a flag is removed, then it should also be removed from vitess-operator and arewefastyet, if used there.
[ ] vtctl command output order should be stable and awk-able.

harshit-gangal commented 1 day ago

Website docs would also need an update https://vitess.io/docs/22.0/user-guides/configuration-advanced/query-consolidation/

jwangace commented 1 day ago

This would also require addition to summary notes for v22 over https://github.com/vitessio/vitess/blob/main/changelog/22.0/22.0.0/summary.md

You can see v21 summary notes to see the template to add new configuration to release notes

Updated.

jwangace commented 1 day ago

Website docs would also need an update https://vitess.io/docs/22.0/user-guides/configuration-advanced/query-consolidation/

Hi @harshit-gangal This is not in this repo, is it?

codecov[bot] commented 1 day ago

Codecov Report

Attention: Patch coverage is 87.50000% with 2 lines in your changes missing coverage. Please review.

Project coverage is 67.38%. Comparing base (0b51839) to head (473d6a4).

Files with missing lines	Patch %	Lines
go/sync2/fake_consolidator.go	0.00%	2 Missing :warning:

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #17244 +/- ## ======================================= Coverage 67.37% 67.38% ======================================= Files 1573 1573 Lines 253113 253124 +11 ======================================= + Hits 170538 170564 +26 + Misses 82575 82560 -15 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

🚨 Try these New Features:

Flaky Tests Detection - Detect and resolve failed and flaky tests
JS Bundle Analysis - Avoid shipping oversized bundles

GrahamCampbell commented 15 hours ago

Will actually have the desired effect of reducing memory, or will this just cause lots of unconsolidated queries, using even more memory?

GrahamCampbell commented 15 hours ago

Might a better fix be to chunk the responses to the waiting sessions, replying to them in chunks of n instead of all at once?

jwangace commented 15 hours ago

Will actually have the desired effect of reducing memory, or will this just cause lots of unconsolidated queries, using even more memory?

This is a very good question, let me start with a simple scenario:

it is a k8s deployment
vttablet set memory limit to 200MB
a result from 1 query is 20MB
first batch of queries we have 10 waits
second batch of queries we have 20 waits
3nd batch of queries we 30 waits
now what about if we have 10,000 waits, wound it cause OOM?
if not, what about 20,000 waits? until wait reaches where VTTablet can not.

The Results, regardless it's just a pointer, meaning no dumplicated content of copies, however there are costs in sending back Results for all the waitted clients, which could be a number when a VTTablet can not handle because of memory limits.

One high memory resulted shoud be the network buffer, now the content of Result will be duplicated copies in each sending-back-socket, with large contents be conjested, it's very hard to control the memory used by tcp/ip buffers, etc.

With this scenario we can quickly realize something isn't correct, assuming VTTablet can handle unlimited waits.

unconsolidated queries will be throttled by connection pools, in this solution.

jwangace commented 15 hours ago

Might a better fix be to chunk the responses to the waiting sessions, replying to them in chunks of n instead of all at once?

Hoestly, this was another solution in my mind, but it has it's shortcomings, perticularly when the wait is huge, where it might take too many minutes until the last chunk of wait get the result, but usually there are already settings in the middle could cut it already and the application, most likely don't want to wait too many minutes, if the application retries, this could cause exponential backups, which will worse the suituation, for this reason it's better to fail quick OR to throttle queries.

This solution, is simpler. When reaches to max waits, the following queries are fall to get connections, which will quickly use up all connection pools, which essentially let VTTablet to throttle the query, it's a trade off of being slower rather than constantly been OOMKilled.

vitessio / vitess