chaoyoung commented on issue #16427:
URL: https://github.com/apache/dubbo/issues/16427#issuecomment-6757400325

   Additional controlled observations on released Dubbo 3.3.7, Netty 
4.2.19.Final and OpenJDK 25 may help separate the two failure modes in this 
report.
   
   **Whole-message progress:** with the actual Triple 8 MiB receive window and 
request(1) issued before DATA, a 9 MiB payload stalls after 8,388,608 DATA 
bytes with no stream WINDOW_UPDATE. The connection window is replenished 
independently. An early-credit control completes all 9,437,189 wire bytes. At 
the boundary, payload 8 MiB - 5 completes; payload 8 MiB - 4 stalls one byte 
short. A synthetic 65,535-byte stream window similarly stalls a 128 KiB message.
   
   **Reachability qualification:** the default Java provider's 8 MiB 
maxResponseBodySize includes the five-byte gRPC prefix. Its actual serializer 
accepts payload 8 MiB - 5 and rejects payload 8 MiB - 4. Therefore the 
large-message case is relevant to peers/configurations permitting messages 
larger than the receive window; it is not evidence that untouched default Java 
peers fail on an ordinary 128 KiB response. Actual Triple 128 KiB responses 
completed with either early or late application demand.
   
   **Window synchronization:** after a completed SETTINGS/ACK handshake on this 
exact Netty version, both advertised and enforced stream windows were 8 MiB. 
This fixture did not reproduce the advertised/local mismatch; it does not 
exclude a pre-ACK race or the original PHP/older-Netty behavior.
   
   The results use paired EmbeddedChannel codecs and an official provider 
serializer, not a PHP interoperability or socket-load test. The whole-message 
progress observations corroborate the scenario addressed by #16443; I have not 
applied that PR to this fixture and am not claiming its fix has been 
independently verified. It would be useful to evaluate message/window progress 
(#16443) and SETTINGS synchronization (#16431) separately. Unconsumed bytes 
before application demand are normal manual backpressure, not by themselves a 
leak.
   
   Investigation and this summary were AI-assisted.
   


-- 
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