Zhikai-VM opened a new pull request, #70790:
URL: https://github.com/apache/doris/pull/70790

   <!-- Recommended: develop the change with the develop-with-evidence skill 
and write this description with the doris-pr-description skill, both from 
https://github.com/apache/doris-skills (the doris-skills submodule of this 
repository). -->
   
   ### Background
   
   <!-- How the feature or module this PR touches works today, in two to four 
plain sentences, for a reviewer who doesn't know it. Link the user doc if there 
is one. Leave the problem and the fix to Problem Summary. -->
   
   The query with 'QUANTILE_STATE_UNION(TO_QUANTILE_STATE(col, 2048))' consume 
huge memory in our performance test environment. We have node with 32 core 128 
Gb memory, and we can not finish the query with memory not enough error.
   
   ### What problem does this PR solve?
   
   <!-- Issue Number: "close #N" if this PR resolves the issue, "#N" if it only 
relates to it, "None" if there is none. A Feature needs an issue. -->
   <!-- Related PR: "#N (relation)" for each PR a reviewer should know about, 
such as "#123 (introduced the bug)", separated by commas. Write "None" if no 
related PR is known. -->
   <!-- PR Type: Bugfix, Performance, Feature, Enhancement, Refactor, Test, 
Build/CI/Chore, Revert, Release/Dependency, or Other. -->
   
   Issue Number: None
   Related PR:
   PR Type: Performance
   
   Problem Summary:
   
   <!-- Two to five sentences: the problem or need and who is affected; the 
approach in one sentence; when the diff is larger than that suggests, what else 
it had to handle, or that most of it is tests; and what existing users see 
after an upgrade. -->
   
   TO_QUANTILE_STATE will store the original values before the value count 
reaches to QUANTILE_STATE_EXPLICIT_NUM (2048). Then it will be converted to 
t-digest. Similar behavior in t-digest, unless the count of centroids in 
t-digest reaches to QUANTILE_STATE_COMPRESSION_MIN(2048), it won't compress the 
centroids it stores. This behavior will cause:
   1. if the rows for each group do not reach 2048, the query would just store 
the raw values, which would consume huge memory
   2. we cannot leverage t-digest to accelerate the query if the it happened in 
1st point.
   
   ### Details
   
   <!--
   For a diff of more than about 300 changed lines, start with "- **Change 
map:**": how many lines are production code, tests, generated files, and docs, 
then the code in parts, each with the problem it handles and its size.
   Then answer each field of your PR type on its own line, as "- **Field:** 
answer", in this order:
     Bugfix: Frequency, Root cause, Why existing tests missed it, Fix, Other 
paths and branches
     Performance: Why it was slow, Optimization, Results, Applicability and cost
     Feature: Scope, Design, Usage, Default behavior, Compatibility
     Enhancement: Before and after, Configs and defaults, Impact on existing 
users
     Refactor: Moves vs. logic changes, Behavior change, Compatibility, 
Performance impact, Rollback
     Test, adding tests: Scenarios, What existing tests miss, Fails when broken
     Test, fixing a test: Failure, Cause, Stability evidence
     Build/CI/Chore: Affected, Developer and release impact, License and 
third-party, Rollback
     Revert: Why not fix forward, Original author notified, Follow-up
     Release/Dependency: Versions, Upstream changes, Compatibility, Rollback
     Other: Why Other
   "None" is a valid answer when it is true. Don't restate the diff.
   -->
   
   In our performance test environment, we have about 100,000 devices. Each 
device have about 50 apps. We could like to have some jobs to calculate the P95 
value of the App's cpu/memory usages and the P95 values of IO by grouping 
device and app. Here is an example:
   
   INSERT INTO agg_table(...)
   SELECT  (QUANTILE_UNION(TO_QUANTILE_STATE(cpu_usage_percentage, 2048))), ...
   FROM app_performance_timeseries
   WHERE ...
   GROUP BY org_id, device_id, app_name;
   
   Basically, each device/app would have cpu/memory/IO metrics in every 
minutes. If we doing the job for each day, then by device/app_name, we would 
have 1440 rows of data for the QUANTILE_UNION/TO_QUANTILE_STATE. It does not 
reach the QUANTILE_STATE_EXPLICIT_NUM. So the query would just store all the 
row data. That would consume about:
   
   100,000 * 50 * 4(bytes of cpu_usage_percentage) * 1440(daily count) = 
28,800,000,000(bytes) about 29GB
   
   We have 5 similar queries which could take about 29 * 5 = 145(GB) memory.
   
   I know that decreasing the QUANTILE_STATE_EXPLICIT_NUM / 
QUANTILE_STATE_COMPRESSION_MIN would get less accurate 
   result. But I think maybe we should let the users to decide accurate result 
or better performance.
   
   ### Testing
   
   <!-- The tests, suites, or commands that ran, what each shows, and where it 
ran (locally or in CI). For a fix, the test that fails before it and passes 
after. -->
   
   The changes are not touching the core logic of the query/computation. I used 
gdb to verify that the configuration works .
   
   ### Release note
   
   <!-- One paragraph on what users will notice, or None. -->
   
   ### Check List (For Author)
   
   <!-- Behavior changed: "No", or "Yes (see Details)" with a Behavior change 
field in Details. -->
   <!-- Does this need documentation: "No", or "Yes" with the 
apache/doris-website PR or the plan. -->
   <!-- Pre-submission status: "Ready"; "Partially verified: " and what is not 
verified; or "Not ready: " and why, for a Draft PR. -->
   
   - Behavior changed: NO
   - Does this need documentation: I am not sure about this, better to have.
   - Pre-submission status: Ready
   
   ### AI Disclosure
   
   <!-- Tool: the agent and model, or None. Generated: the parts it wrote 
(code, tests, this description), or None. Human review: what you reviewed of 
the generated parts, or "Not applicable" without AI tools. You may add "- 
Prompts:" to summarize how you directed the agent. -->
   
   - Tool: None
   - Generated: None
   - Human review: Not applicable
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to