Chapter 25 — Async gRPC Server, Bounded Admission
A grpc::CompletionQueue is the one concurrency model underneath all of gRPC C++'s APIs, and a server built directly on it needs its own explicit admission control.
The full source for this example lives in
examples/concurrency/25-concurrent-grpc/— clone the repo,cdin, and./demo.sh.
Chapter reference: Doc 25 — Concurrent gRPC Services
What it shows
Every call is implemented directly against grpc::ServerCompletionQueue as the chapter’s CallData state machine — CREATE -> PROCESS -> (WRITE <-> DELAY)* -> FINISH — re-arming itself at each phase so the server never stops accepting new calls. A single process-wide atomic counter enforces a hard admission cap via a reserve-then-rollback pattern: a call that would push the count over the cap finishes immediately with RESOURCE_EXHAUSTED, the gRPC equivalent of an HTTP 503. Concurrent streaming clients, released from the same starting line, prove that exactly cap calls succeed at the cap and that overflow calls are rejected with no hang and no third status; a final check confirms the server’s per-queue drain loop lets it exit 0 within 10 seconds of SIGTERM.
Running it
cd examples/concurrency/25-concurrent-grpc
./demo.sh # build + run server+client inside the container + verify.lua
./demo.sh --clean # remove the image