Concurrency

Layer 1: Language Mechanics

Core Question

Is this CPU-bound or I/O-bound, and what's the sharing model?

Before choosing concurrency primitives:

• What's the workload type?
• What data needs to be shared?
• What's the thread safety requirement?

---

Error → Design Question

Error	Don't Just Say	Ask Instead
E0277 Send	"Add Send bound"	Should this type cross threads?
E0277 Sync	"Wrap in Mutex"	Is shared access really needed?
Future not Send	"Use spawn_local"	Is async the right choice?
Deadlock	"Reorder locks"	Is the locking design correct?

---

Thinking Prompt

Before adding concurrency:

1. What's the workload?

   • CPU-bound → threads (std::thread, rayon)
   • I/O-bound → async (tokio, async-std)
   • Mixed → hybrid approach

2. What's the sharing model?

   • No sharing → message passing (channels)
   • Immutable sharing → Arc<T>
   • Mutable sharing → Arc<Mutex<T>> or Arc<RwLock<T>>

3. What are the Send/Sync requirements?

   • Cross-thread ownership → Send
   • Cross-thread references → Sync
   • Single-thread async → spawn_local

---

Trace Up ↑ (MANDATORY)

CRITICAL: Don't just fix the error. Trace UP to find domain constraints.

Domain Detection Table

Context Keywords	Load Domain Skill	Key Constraint
Web API, HTTP, axum, actix, handler	domain-web	Handlers run on any thread
交易, 支付, trading, payment	domain-fintech	Audit + thread safety
gRPC, kubernetes, microservice	domain-cloud-native	Distributed tracing
CLI, terminal, clap	domain-cli	Usually single-thread OK

Example: Web API + Rc Error

"Rc cannot be sent between threads" in Web API context
    ↑ DETECT: "Web API" → Load domain-web
    ↑ FIND: domain-web says "Shared state must be thread-safe"
    ↑ FIND: domain-web says "Rc in state" is Common Mistake
    ↓ DESIGN: Use Arc<T> with State extractor
    ↓ IMPL: axum::extract::State<Arc<AppConfig>>

Generic Trace

"Send not satisfied for my type"
    ↑ Ask: What domain is this? Load domain-* skill
    ↑ Ask: Does this type need to cross thread boundaries?
    ↑ Check: m09-domain (is the data model correct?)

Situation	Trace To	Question
Send/Sync in Web	domain-web	What's the state management pattern?
Send/Sync in CLI	domain-cli	Is multi-thread really needed?
Mutex vs channels	m09-domain	Shared state or message passing?
Async vs threads	m10-performance	What's the workload profile?

---

Trace Down ↓

From design to implementation:

"Need parallelism for CPU work"
    ↓ Use: std::thread or rayon

"Need concurrency for I/O"
    ↓ Use: async/await with tokio

"Need to share immutable data across threads"
    ↓ Use: Arc<T>

"Need to share mutable data across threads"
    ↓ Use: Arc<Mutex<T>> or Arc<RwLock<T>>
    ↓ Or: channels for message passing

"Need simple atomic operations"
    ↓ Use: AtomicBool, AtomicUsize, etc.

---

Send/Sync Markers

Marker	Meaning	Example
Send	Can transfer ownership between threads	Most types
Sync	Can share references between threads	Arc<T>
!Send	Must stay on one thread	Rc<T>
!Sync	No shared refs across threads	RefCell<T>

Quick Reference

Pattern	Thread-Safe	Blocking	Use When
std::thread	Yes	Yes	CPU-bound parallelism
async/await	Yes	No	I/O-bound concurrency
Mutex<T>	Yes	Yes	Shared mutable state
RwLock<T>	Yes	Yes	Read-heavy shared state
mpsc::channel	Yes	Optional	Message passing
Arc<Mutex<T>>	Yes	Yes	Shared mutable across threads

Decision Flowchart

What type of work?
├─ CPU-bound → std::thread or rayon
├─ I/O-bound → async/await
└─ Mixed → hybrid (spawn_blocking)

Need to share data?
├─ No → message passing (channels)
├─ Immutable → Arc<T>
└─ Mutable →
   ├─ Read-heavy → Arc<RwLock<T>>
   └─ Write-heavy → Arc<Mutex<T>>
   └─ Simple counter → AtomicUsize

Async context?
├─ Type is Send → tokio::spawn
├─ Type is !Send → spawn_local
└─ Blocking code → spawn_blocking

---

Common Errors

Error	Cause	Fix
E0277 Send not satisfied	Non-Send in async	Use Arc or spawn_local
E0277 Sync not satisfied	Non-Sync shared	Wrap with Mutex
Deadlock	Lock ordering	Consistent lock order
future is not Send	Non-Send across await	Drop before await
MutexGuard across await	Guard held during suspend	Scope guard properly

---

Anti-Patterns

Anti-Pattern	Why Bad	Better
Arc<Mutex<T>> everywhere	Contention, complexity	Message passing
thread::sleep in async	Blocks executor	tokio::time::sleep
Holding locks across await	Blocks other tasks	Scope locks tightly
Ignoring deadlock risk	Hard to debug	Lock ordering, try_lock

---

Async-Specific Patterns

Avoid MutexGuard Across Await

// Bad: guard held across await
let guard = mutex.lock().await;
do_async().await;  // guard still held!

// Good: scope the lock
{
    let guard = mutex.lock().await;
    // use guard
}  // guard dropped
do_async().await;

Non-Send Types in Async

// Rc is !Send, can't cross await in spawned task
// Option 1: use Arc instead
// Option 2: use spawn_local (single-thread runtime)
// Option 3: ensure Rc is dropped before .await

---

Related Skills

When	See
Smart pointer choice	m02-resource
Interior mutability	m03-mutability
Performance tuning	m10-performance
Domain concurrency needs	domain-*