python-concurrency-performance（Python concurrency 性能）

Python Concurrency and Performance

Overview

Correct concurrency starts with matching the model to the workload, not the developer's preference. This skill encodes defaults for model selection, cancellation/deadline behavior, and lifecycle safety—prioritizing explicit control over implicit magic.

Treat these recommendations as preferred defaults. When project constraints demand deviation, call out tradeoffs and compensating controls.

When to Use

• Selecting between asyncio, threading, multiprocessing, or concurrent.futures
• Propagating deadlines or cancellation through async call chains
• Bounding fan-out, backpressure, or semaphore-guarded concurrency
• Diagnosing race conditions, deadlocks, or priority inversion
• Profiling throughput bottlenecks before and after optimization
• Verifying no task or thread leaks on shutdown or lifecycle transitions

When NOT to Use

• Pure CPU-bound numeric work better served by NumPy/C extensions
• Single-threaded scripting with no concurrent I/O
• Distributed systems coordination (use a workflow/orchestration skill instead)

Quick Reference

• Choose the concurrency model by workload profile (I/O-bound → asyncio/threads; CPU-bound → multiprocessing).
• Keep cancellation and cleanup explicit—never rely on garbage collection to close resources.
• Bound fan-out and backpressure with semaphores or queue limits; unbounded spawning invites OOM.
• Measure before optimizing; re-measure after every change to confirm the win.
• Verify no task/thread leaks on any lifecycle-sensitive change (startup, shutdown, reconnect).

Common Mistakes

• Defaulting to threads for I/O-bound work — asyncio avoids thread-safety bugs entirely for network I/O; threads add synchronization overhead for no gain.
• Ignoring cancellation propagation — a cancelled parent that doesn't cancel children leaks tasks and holds connections open.
• Unbounded gather / submit calls — spawning thousands of tasks without a semaphore or bounded executor starves the event loop or exhausts OS threads.
• Optimizing without profiling — guessing at bottlenecks leads to complex code that solves the wrong problem; always profile first.
• Missing shutdown verification — tests that don't assert clean shutdown mask slow resource leaks that surface only in production under load.

Scope Note

• Treat these recommendations as preferred defaults for common cases, not universal rules.
• If a default conflicts with project constraints or worsens the outcome, suggest a better-fit alternative and explain why it is better for this case.
• When deviating, call out tradeoffs and compensating controls (tests, observability, migration, rollback).

Invocation Notice

• Inform the user when this skill is being invoked by name: python-concurrency-performance.

References

• references/concurrency-models.md
• references/deadlines-cancellation-lifecycle.md
• references/leak-detection.md