Details
-
Bug
-
Status: Closed
-
Critical
-
Resolution: Fixed
-
proton-0.12.0, proton-0.12.1
-
None
-
Any
Description
Actual Issue/scenario hit by Microsoft Azure EventHubs:
We have a pattern where customers sends messages in a burst to our Queue and stop sending and then wait for all of them to be received.
Because of this issue in Proton-j Amqp implementation - we can see many bytes were stuck in the SSL Decode Buffer and were only picked up when new transport frames arrive.
Given that this is a 1 line fix - it would be great if you folks can publish an incremental minor update to 0.12.X.
Here are my findings after debugging through this issue:
- When incoming bytes arrive on the SocketChannel – proton-j client gets signaled by nio & as a result it unwinds the transport stack – as a result all the TransportInput implementations performs its task on the Read Bytes and hands off to the Next Layer in the stack (transport to ssl, ssl to frameparser etc).
- While unwinding that stack, SimpleSSLTransportWrapper.unwrapInput reads(16k bytes) from _inputBuffer and the result - decoded bytes are written to _decodedInputBuffer – as an intermediate buffer.
- It then flushes bytes from intermediate buffer to the next layer & invokes an _underlyingInput.Process() – to signal it that it has bytes in its input buffer.
- If the underlyingInput (lets say FrameParser) buffer size is small – lets say 4k – then decodedInputBuffer will be left with 12k bytes & Over time this accrues.
The fix here is to flush decodedInputBuffer to the Next transport in the Network Stack & call _underlyingInput.Process() - until decodedInputBuffer is empty. Here’s the pull request:
https://github.com/apache/qpid-proton/pull/73
Attachments
Issue Links
- relates to
-
PROTON-1185 [proton-j] Send operations like SendLink.send() & RecvLink.flow() (aka writing to the transport stack) doesn't flush all bytes to the Network
- Closed