360 Secure SkillHub
360 安全技能中心

m13-domain-error

该 Skill 以设计指南形式帮助区分用户、开发者和运维可见的领域错误,讨论重试、回退、熔断、超时与错误上下文,并给出 Rust 错误层级和重试模式示例,适合在错误处理方案设计阶段参考。

actionbook
热度 0.00
最近更新 2026-10-19 02:01:24
分类:开发依赖:纯提示词应用场景:概念解释 / 代码生成 / 代码审查

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add actionbook/rust-skills/m13-domain-error
Metadata
name
m13-domain-error
description
Use when designing domain error handling. Keywords: domain error, error categorization, recovery strategy, retry, fallback, domain error hierarchy, user-facing vs internal errors, error code design, circuit breaker, graceful degradation, resilience, error context, backoff, retry with backoff, error recovery, transient vs permanent error, 领域错误, 错误分类, 恢复策略, 重试, 熔断器, 优雅降级
user-invocable
false

Domain Error Strategy

Layer 2: Design Choices

Core Question

Who needs to handle this error, and how should they recover?

Before designing error types:

  • Is this user-facing or internal?
  • Is recovery possible?
  • What context is needed for debugging?

Error Categorization

Error Type Audience Recovery Example
User-facing End users Guide action InvalidEmail, NotFound
Internal Developers Debug info DatabaseError, ParseError
System Ops/SRE Monitor/alert ConnectionTimeout, RateLimited
Transient Automation Retry NetworkError, ServiceUnavailable
Permanent Human Investigate ConfigInvalid, DataCorrupted

Thinking Prompt

Before designing error types:

  1. Who sees this error?

    • End user → friendly message, actionable
    • Developer → detailed, debuggable
    • Ops → structured, alertable
  2. Can we recover?

    • Transient → retry with backoff
    • Degradable → fallback value
    • Permanent → fail fast, alert
  3. What context is needed?

    • Call chain → anyhow::Context
    • Request ID → structured logging
    • Input data → error payload

Trace Up ↑

To domain constraints (Layer 3):

"How should I handle payment failures?"
    ↑ Ask: What are the business rules for retries?
    ↑ Check: domain-fintech (transaction requirements)
    ↑ Check: SLA (availability requirements)
Question Trace To Ask
Retry policy domain-* What's acceptable latency for retry?
User experience domain-* What message should users see?
Compliance domain-* What must be logged for audit?

Trace Down ↓

To implementation (Layer 1):

"Need typed errors"
    ↓ m06-error-handling: thiserror for library
    ↓ m04-zero-cost: Error enum design

"Need error context"
    ↓ m06-error-handling: anyhow::Context
    ↓ Logging: tracing with fields

"Need retry logic"
    ↓ m07-concurrency: async retry patterns
    ↓ Crates: tokio-retry, backoff

Quick Reference

Recovery Pattern When Implementation
Retry Transient failures exponential backoff
Fallback Degraded mode cached/default value
Circuit Breaker Cascading failures failsafe-rs
Timeout Slow operations tokio::time::timeout
Bulkhead Isolation separate thread pools

Error Hierarchy

#[derive(thiserror::Error, Debug)]
pub enum AppError {
    // User-facing
    #[error("Invalid input: {0}")]
    Validation(String),

    // Transient (retryable)
    #[error("Service temporarily unavailable")]
    ServiceUnavailable(#[source] reqwest::Error),

    // Internal (log details, show generic)
    #[error("Internal error")]
    Internal(#[source] anyhow::Error),
}

impl AppError {
    pub fn is_retryable(&self) -> bool {
        matches!(self, Self::ServiceUnavailable(_))
    }
}

Retry Pattern

use tokio_retry::{Retry, strategy::ExponentialBackoff};

async fn with_retry<F, T, E>(f: F) -> Result<T, E>
where
    F: Fn() -> impl Future<Output = Result<T, E>>,
    E: std::fmt::Debug,
{
    let strategy = ExponentialBackoff::from_millis(100)
        .max_delay(Duration::from_secs(10))
        .take(5);

    Retry::spawn(strategy, || f()).await
}

Common Mistakes

Mistake Why Wrong Better
Same error for all No actionability Categorize by audience
Retry everything Wasted resources Only transient errors
Infinite retry DoS self Max attempts + backoff
Expose internal errors Security risk User-friendly messages
No context Hard to debug .context() everywhere

Anti-Patterns

Anti-Pattern Why Bad Better
String errors No structure thiserror types
panic! for recoverable Bad UX Result with context
Ignore errors Silent failures Log or propagate
Box everywhere Lost type info thiserror
Error in happy path Performance Early validation

Related Skills

When See
Error handling basics m06-error-handling
Retry implementation m07-concurrency
Domain modeling m09-domain
User-facing APIs domain-*