Proposal: opt-in server request admission and response budgets

Aperta
  • Qualcosa non va?
#319 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 2 giorni

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
5/5
Tempo stimato
Più di una settimana
Idoneità per principianti
12/100
Tipo di issue
Funzionalità
Chiarezza
Da chiarire
Stato di attività
Attiva
Stack tecnologico
swift
Ambito
backend

Direzione di ricerca

This is a design proposal, not a bug or a scoped change. The payload names only the proposed Server.Configuration and StdioTransport.Configuration shapes, so start by reading the existing Server and stdio transport configuration types to see how limits are currently handled. Done means maintainers agree on the admission, budget and shutdown contracts before any patch. Do not start an implementation from this issue alone.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

enhancement

After verifying #266's stdio serialization and #298's early-cancellation changes,
there is a separate resource-control need for hosts that expose deterministic tools
with bounded input and output contracts. A bounded input frame alone does not bound
SDK-created request tasks, queued calls, or an aggregate JSON-RPC batch response.

Would an opt-in configuration surface along these lines fit the SDK?

Server.Configuration(
    strict: true,
    maxConcurrentRequests: 4,
    maxQueuedRequests: 8,
    maxBatchItems: 8,
    maxResponseBytes: 65_535,
    maxPendingControlResponses: 64)

// Separate proposed transport configuration, naming still to be agreed:
StdioTransport.Configuration(
    maxInboundFrameBytes: 16_384,
    maxOutboundFrameBytes: 65_536,
    maxBufferedMessages: 1,
    maxPendingWrites: 4)

The numbers are host policy examples. Existing callers would retain unlimited
execution/batch/response defaults. This is a distinct proposal from #266/#298 and
would benefit from agreement on the contracts before an upstream patch:

  • Reserve request ownership and admission before any suspending context lookup.
    Queue request values without allocating handler tasks. Return explicit overload
    errors when capacity is full. A separately bounded control-reply queue keeps
    reception available for cancellation while those errors await output. Count
    cancelled active work until it actually exits; retain correlation IDs through
    delayed batch delivery, separately from execution slots.
  • Allow cancellation notifications through while admitted work executes. Remove
    cancelled queued calls before their handlers start. Scope correlation/admission
    by the appropriate Server/session boundary; preserve HTTP transport routing
    rather than assuming a bare wire ID is globally unique.
  • Bound both per-item responses and the encoded aggregate batch envelope. Preserve
    IDs in explicit overflow errors; close the connection when even an error cannot
    fit. Never truncate a success payload. Count bytes before transport framing.
  • Retain completed batch results separately under a bounded accumulator, so a batch
    larger than the execution limit can finish. The final item holds a delivery slot
    while sending the aggregate. Define how configuration changes affect work already
    admitted; immutable per-connection limits would make the contract simpler.
  • Define host-driven shutdown completion and EOF cleanup. Waiting is cooperative;
    process deadlines remain a host responsibility. Avoid a handler awaiting a
    shutdown operation that itself waits for that handler.

The transport configuration should reject an oversized frame before accumulating
beyond its bound, count the delimiter consistently, bound pending/buffered messages,
and preserve FIFO writes through partial I/O. Backpressure must avoid silently
dropping a cancellation/control frame. A partial write failure needs an explicit
connection-failure policy before another frame may be written. These transport
APIs are proposed separately from Server admission; stock stdio presently remains
an unbounded input producer. Neither layer bounds arbitrary handler allocations
or notification-handler work. No networking/domain policy
belongs in these APIs.

A local development prototype based on release 0.12.1 and tests establish feasibility: the full native
macOS SDK suite passes 573 tests, and 28 focused native ARM64 Linux tests cover
4 active / 8 queued admission, overload, queued/running cancellation, retained
capacity during response delivery, batch8 through slots4, cancellation during a
batch, aggregate/individual budgets, too-small error budgets, negotiated batch
eligibility, EOF, and cooperative cleanup. This prototype is unpublished and is
not an upstream API commitment. No new dependency or broad protocol rewrite is
required for the stdio pilot. Full Linux test-target qualification remains blocked
by unrelated baseline URLProtocol helper diagnostics, separately documented.

Lingua principale
Swift
Stelle
1.6k
Fork
270
Merge medio
2g 7h
PR unite (30g)
15

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di modelcontextprotocol/swift-sdk

Tutte le issue di modelcontextprotocol/swift-sdk

Issue simili

Altre issue su Swift

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.