wenzhenghu commented on PR #63332:
URL: https://github.com/apache/doris/pull/63332#issuecomment-4552756225

   > This is a great find — thanks for digging into the Jetty 12 regression. A 
couple of questions on the approach:
   > 
   > 1. Are the Jetty-level maxUnconsumedRequestContentReads = -1 and the 
application-level bounded drain redundant?
   > 2. Misbehaving clients sending the body early is fundamentally a 
client-side problem — they aren't implementing Expect: 100-continue correctly, 
and the right fix is to push the clients to comply. Having FE actively consume 
the request body by default essentially makes FE take on body traffic it was 
never supposed
   >    to handle; under high-concurrency Stream Load, this will noticeably 
increase FE's network and thread pressure.
   
   
   ### 1. Stability Comparison
   Based on current test results:
   - The Doris 3.0 baseline instance did not reproduce `BrokenPipeError` under 
the 30 sets of extreme parameters already covered.
   - Doris 3.1 (after this PR modification) shows clear risk of 
`BrokenPipeError` when `idle_time_ms = 1000`.
   - After setting `idle_time_ms = 2000`, the representative test cases for 
Doris 3.1 (after this PR modification) have recovered to a 10/10 success rate.
   
   Therefore, the stability conclusion can be summarized as:
   - Doris 3.0 behaves more conservatively, with no premature disconnections 
observed within the current test scope.
   - Doris 3.1 (after this PR modification) is more susceptible to link 
conditions under smaller idle timeouts.
   - The stability of Doris 3.1 (after this PR modification) can be 
significantly improved by appropriately increasing the idle timeout.
   
   ### 2. Performance Comparison
   Based on the latency of successful requests:
   - Doris 3.1 (after this PR modification) @ 172.16.0.90 @ 2000ms: ~40.515s
   - Doris 3.0: Most samples range from 36s to 42s, and under more aggressive 
parameters, latency can rise to 100s+ or even 200s+.
   
   Therefore, the performance conclusion is:
   - On low-latency paths, Doris 3.1 (after this PR modification) is 
significantly faster than Doris 3.0.
   - On high-latency paths, the successful request latency of Doris 3.1 (after 
this PR modification) is close to that of Doris 3.0, but its stability depends 
on a more reasonable idle timeout configuration.
   - The main characteristic of Doris 3.0 is not high performance, but rather 
higher compatibility achieved through longer connection hold times.


-- 
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