Backpressure

Level: 301 · deep dive

Stub — an outline, not a lesson. There is no runnable example behind this page yet, so nothing on it has been through the check that backs every other claim in this library. The bullets below are the questions the finished page has to answer.

One line: An unbounded queue does not remove overload, it moves it into memory, where it ends the process — so every queue in the server needs a bound and a decision for when it is full: make the sender wait, refuse the work, or drop it.

What it has to cover

• An unbounded channel under a producer faster than its consumer, with memory measured as it grows
• Bounded mpsc ↗ and the three policies: send().await waits (the slowness travels back to the producer), try_send refuses (the client hears "busy" now), and dropping is lossy (acceptable for metrics, never for writes)
• Choosing a policy per queue in minidb — the store actor's queue, a logging queue — and choosing the bound from a measurement rather than a round number
• A Semaphore ↗ to limit concurrent connections, with acquire_owned so the permit moves into the connection task and is released when the task ends
• Where the permit is acquired decides what the limit protects. Before accept, excess clients wait in the kernel's listen backlog and hold nothing of yours; after accept but inside the spawned task, every waiting client already holds a socket and a file descriptor

The trap it exists for

Acquiring the permit inside the spawned task. The limit then caps work, not connections — accepted sockets pile up without bound while they wait for a permit, and the file-descriptor limit is hit anyway.

What minidb gains

A bounded queue in front of the store and a cap on concurrent connections, each with a stated policy for when it is full.

See also

• Who owns the state — the actor whose queue this bounds
• Channels — sync_channel, std's bounded channel
• Backpressure ↗, bounded channel ↗ and semaphore ↗ — in the Concurrency library

Po polsku

Nieograniczona kolejka nie usuwa przeciążenia, tylko przenosi je do pamięci, gdzie kończy się ono śmiercią procesu. Ciśnienie zwrotne (backpressure) to decyzja, co zrobić, gdy ograniczona kolejka jest pełna: kazać nadawcy czekać, odmówić albo porzucić pracę. Przy ograniczaniu liczby połączeń semaforem liczy się miejsce pobrania zezwolenia — przed accept nadmiarowi klienci czekają w kolejce jądra, po accept każdy z nich już zajmuje gniazdo i deskryptor pliku.

Szukaj po polsku: ciśnienie zwrotne · ograniczona kolejka · tokio bounded channel backpressure · tokio semaphore limit connections