erract - Structured Error Handling for Rust

erract provides production-ready error handling that solves:

• Actionable errors: Categorized by what callers can do (retry, fail, etc.)
• Explicit retry semantics: No guessing from error messages
• Rich context: Automatic location capture with zero overhead
• Type safety: Enforce context at module boundaries

Philosophy

Two audiences, two needs:

• Machines: Need flat structures, clear error kinds, predictable codes
• Humans: Need rich context, call paths, business-level information

This library combines the best ideas from:

• Apache OpenDAL's error design
• The exn crate
• Years of production error handling experience

Quick Start

use erract::prelude::*;

fn process_user(id: u32) -> erract::Result<String> {
    // Use static constructors for zero-allocation common errors
    let user = lookup_user(id)
        .or_raise(|| Error::not_found())?
        .ok_or_else(|| {
            Error::permanent(ErrorKind::NotFound, format!("user {} not found", id))
                .with_context("user_id", id.to_string())
                .raise()
        })?;

    Ok(user)
}

fn lookup_user(id: u32) -> erract::Result<Option<String>> {
    if id == 0 {
        // Static message - zero allocation
        return Err(Error::not_found().raise());
    }
    Ok(Some(format!("User{}", id)))
}

Features

• High-performance Context Arena: Thread-local storage for dynamic context, making large-scale context addition up to 40% faster than alternatives.
• Zero runtime overhead: Uses #[track_caller] instead of expensive backtraces
• Domain-specific errors: HTTP, Database, and Storage error kinds
• Error trees: Support for multiple concurrent failures
• Type safety: Compiler enforces context at module boundaries
• Zero-copy static strings: Use Cow<'static, str> for static messages
• Optimized serialization: Fast JSON and machine-readable output

API Highlights

Static Constructors (Zero Allocation)

// Common errors with pre-defined static messages
let err = Error::not_found();           // "not found"
let err = Error::timeout();             // "operation timed out"
let err = Error::permission_denied();   // "permission denied"
let err = Error::validation_failed();   // "validation failed"
let err = Error::unexpected();          // "unexpected error"

Context Methods

// For string context (zero-copy for static strings)
error.with_context("key", "value")

// For non-string values (uses ToString)
error.with_context_value("count", 42)

Serialization

let error = Error::not_found()
    .with_context("resource", "user")
    .with_operation("fetch");

// Human-readable
println!("{}", error);  // "not found (operation: fetch) [resource: user]"

// Machine-readable
error.to_machine_string()  // "kind=not_found;status=permanent;message=not found;..."

// JSON (optimized, ~220ns)
error.to_json()  // {"kind":"not_found","status":"permanent",...}

Comparison with anyhow

erract and anyhow serve different use cases:

Performance Benchmarks

Run cargo bench to reproduce. Results on Linux x86_64:

Key Insights:

• erract excels at error creation, high-pressure context addition (via Arena), and structured queries
• anyhow excels at deep error propagation and human-only display formatting
• Memory: erract::Error is 112 bytes (structured), anyhow::Error is 8 bytes (pointer)
• Trade-off: erract uses more stack space and thread-local memory to achieve superior performance in context-heavy scenarios

When to Choose

Choose erract when:

• Building production services with retry logic
• Need machine-parseable errors for alerting/metrics
• Want explicit error semantics at compile time
• Multiple teams consume your error data

Choose anyhow when:

• Building CLI tools or quick prototypes
• Error handling is human-only (logs, not machines)
• Minimal boilerplate is priority
• Don't need structured error classification

See benches/README.md for detailed benchmark methodology.

License

This project is licensed under either of

• Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.