Caideyipi opened a new pull request, #18853:
URL: https://github.com/apache/iotdb/pull/18853

   ## Description
   
   A DataNode can keep rejecting new Thrift connections after a large-frame 
burst or leader failover because long-lived connections retain oversized RPC 
buffers. Requests at or below the initial buffer capacity never enter the 
normal shrink path, including the shrink requested by `flush()`. Once the 
shared pool is full, allocating even the initial 1 KB buffer fails repeatedly 
across RPC services.
   
   This change allows small frames to participate in gradual shrinking, with 
the initial capacity as the lower bound. When allocation fails, the memory 
manager reclaims oversized buffers whose frames have been fully consumed or 
sent and retries immediately. Candidates belong to the same memory control and 
are held through a weak registry. Nonblocking buffer locks prevent concurrent 
allocations from waiting on each other's buffers; unread frames and unflushed 
writes remain protected. The same lifecycle covers compression buffers.
   
   Failed transport opens also close their allocated buffers. This prevents 
reconnect attempts from leaking quota when a client constructor fails before 
the caller can obtain the transport. Allocation failures now include the pool's 
total, used, and available bytes, with matching English and Chinese diagnostics.
   
   ## Validation
   
   - `mvn test -pl iotdb-client/service-rpc`: **107 tests passed**, including a 
local Thrift server accepting a new connection while an old connection retains 
large-frame buffers.
   - Regression coverage for small-frame shrinking, pressure recovery, 
unread/unflushed payload integrity, Snappy scratch buffers, concurrent 
allocation, memory-control ownership, failed reconnect cleanup, and pool 
diagnostics.
   - Full-reactor English and Chinese `test-compile` with tests skipped.
   - Formatting, Checkstyle, locale key/format parity, and inline diagnostic 
checks.
   
   The original three-DataNode failover workload has not been rerun with this 
patch.
   
   This PR has:
   
   - [x] been self-reviewed, including concurrent allocation and reclamation.
   - [x] updated buffer policy documentation and comments explaining frame 
ownership and lock ordering.
   - [x] added regression tests for the new paths and preserved rejection when 
only active buffers occupy the pool.
   


-- 
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]

Reply via email to