/sota-async-concurrency

name: sota-async-concurrency
description: >-
  State-of-the-art rules for writing and auditing asynchronous and concurrent
  code across runtimes (Python asyncio, JS/Node, Go, Rust, JVM). Use when
  building anything with async/await, threads, processes, event loops, task
  groups, channels, or queues — and when auditing existing code for race
  conditions, deadlocks, leaked tasks, blocked event loops, missing
  cancellation, or backpressure failures. Not for latency/throughput profiling
  and optimization — use sota-performance. Trigger keywords: async, await,
  concurrency, parallelism, threads, race condition, deadlock, event loop,
  channels, queues, semaphore, mutex, cancellation, timeout, backpressure,
  task group, goroutine, tokio, asyncio.

SOTA Async & Concurrency

Purpose

Concurrency bugs are the most expensive class of defect: they pass tests, ship, and then corrupt data or hang production under load. This skill encodes the 2026 state of the art for concurrent design and the bug catalog auditors need to spot defects **by reading code**, without reproducing them. Concepts are cross-language; per-runtime notes are inlined where semantics genuinely differ (GIL, goroutine scheduling, tokio executors, Node's single loop).

Two operating modes. Pick one explicitly before starting.

BUILD mode

When writing new concurrent code:

1. **Classify the workload first.** I/O-bound → async/event loop. CPU-bound → threads (if runtime has real parallelism) or processes. Mixed → async front-end + bounded worker pool. Read [`rules/01`](rules/01-models-and-structure.md) before choosing; on an event loop, keep blocking and CPU work off it ([`rules/04`](rules/04-event-loop-hygiene.md)). 2. **Structured concurrency is the default.** Every task lives inside a scope (TaskGroup / nursery / errgroup / JoinSet) that joins or cancels it. Spawning a task with no owner is a design smell requiring written justification. 3. **Bound everything.** Every queue, channel, connection pool, in-flight request set, and spawn loop gets an explicit capacity. Unbounded = OOM with a delay timer. 4. **Every await gets a timeout policy** — a number, or a documented reason why it inherits one from an enclosing scope ([`rules/05`](rules/05-cancellation-timeouts-shutdown.md)). 5. **Cancellation is a feature you build, not an exception you ignore.** Propagate context/AbortSignal/CancelledError; clean up in finally blocks; design shutdown order (stop intake → drain → deadline → force) ([`rules/05`](rules/05-cancellation-timeouts-shutdown.md)). 6. **Shared mutable state needs an owner.** Prefer message passing or single-owner tasks; if you must lock, define lock ordering and never hold a lock across an await. Races vs race conditions, atomicity and TOCTOU: [`rules/02`](rules/02-correctness.md); choosing the primitive: [`rules/03`](rules/03-primitives.md). Ambient per-request context (tenant, user) is reset when the request ends, or a reused worker serves the next request with it: [`rules/07`](rules/07-audit-bug-catalog.md) §3. 7. Re-read the audit checklists at the end of each rules file against your own diff before declaring done.

AUDIT mode

**Coverage:** every `rules/` file's Audit checklist applies in AUDIT mode. Mark one *not applicable* only with its reason in the report — never skip a file because the thing it checks for is absent, since that absence may be the finding.

When auditing existing code, you find races, deadlocks, and leaks by reading — grep is your debugger. Workflow:

1. **Map the concurrency topology.** What spawns tasks/threads? What shares state? What are the queues and their bounds? Draw the lock set and the channel graph mentally before judging any line. 2. **Sweep with targeted greps**, then read each hit in context:

• Fire-and-forget: `create_task(` / `ensure_future(` without a stored

handle; bare `go func(`; `.then(` with no `.catch(`; floating promises; `tokio::spawn` whose JoinHandle is dropped.

• Missing await: async calls whose return value is discarded.
• Blocking in async: `time.sleep`, `requests.`, `open(` , sync DB drivers,

`fs.readFileSync`, `bcrypt.hashSync`, `std::thread::sleep` inside async fns.

• Lock across await: `async with lock:` / `mutex.lock()` bodies containing

`await` / `.await`.

• Unbounded: `Queue()` with no maxsize, `make(chan T)` fed by fast

producers, `unbounded_channel`, spawn-in-loop with no semaphore.

• Check-then-act: `if x in d:` … `d[x]`, exists-then-create, read-modify-write

on shared counters without atomics/locks. 3. **For each suspect, prove the interleaving.** State the two (or more) execution orders and which one breaks. A finding without an interleaving is a style note, not a concurrency bug. 4. Read [`rules/07`](rules/07-audit-bug-catalog.md) for the full bug catalog with signatures.

Severity conventions

| Severity | Criteria | Examples | |---|---|---| | CRITICAL | Data corruption, deadlock, or unbounded resource growth reachable under normal load | Lost-update race on money/state; lock-ordering deadlock on hot path; unbounded queue fed by network input | | HIGH | Hang, leak, or wrong result under plausible (load/error/timeout) conditions | Fire-and-forget swallowing exceptions; no timeout on external call; lock held across await; blocking call on event loop | | MEDIUM | Degraded behavior, starvation, or fragility under contention | Writer starvation on RwLock; missing jitter on retries; thundering herd on cache expiry; spurious-wakeup-unsafe condvar wait | | LOW | Latent hazard or convention violation with no current trigger | Orphanable task that today happens to finish first; missing cancellation propagation in a path that is never cancelled yet |

Escalate one level when the affected state is money, auth, or durability.

Finding format

[SEVERITY] file:line — short title
Race window / failure mode: the exact inter