Anon2Tokyo commented on issue #16414: URL: https://github.com/apache/dubbo/issues/16414#issuecomment-5226185529
STEP： The runtime reproduction is timing-sensitive because it depends on how HTTP/2 DATA frames are coalesced, but the scenario I used is: - Dubbo `3.3.6` - Triple provider without generated stub inheritance - client-streaming RPC - the client sends many small messages quickly, for example `2000` messages with `8` bytes each The important part is that the provider should expose a plain Java service interface instead of extending the generated Triple stub, so the request goes through the no-stub method-discovery path. Example run shape: ```powershell # build the repro app mvn.cmd -q -DskipTests package mvn.cmd -q dependency:build-classpath "-Dmdep.outputFile=target\classpath.txt" # start provider .\run-provider.ps1 -ConfigName application-provider -LogName provider-tri -ClasspathFile target\classpath.txt # run consumers repeatedly .\run-consumers.ps1 -ConfigName application-consumer -RunName native-tri -Iterations 30 -Chunks 2000 -ChunkBytes 8 -ClasspathFile target\classpath.txt ``` A more aggressive run is: ```powershell .\run-consumers.ps1 -ConfigName application-consumer -RunName native-tri-stress -Iterations 60 -Chunks 10000 -ChunkBytes 1 -ClasspathFile target\classpath.txt ``` Expected failure signatures are usually one of: ```text reserved bits not zero IllegalReferenceCountException LengthFieldStreamingDecoder Upload failed ``` The underlying frame layout that triggers the bug is: ```text DATA frame 1: [complete message 1][prefix of message 2] DATA frame 2: [remaining bytes of message 2] ``` With the old implementation, the prefix of message 2 can be buffered in the temporary `GrpcStreamingDecoder` used during method discovery, then lost when the stream switches to the actual decoder. This PR avoids that by reusing the stream's existing decoder during method discovery. One note: I do not think this is 100% deterministic as a black-box runtime repro, since it depends on local scheduling and HTTP/2 frame packing. The regression test in this PR directly covers the problematic decoder state transition. -- 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]
