backend-engineering

Overview

Backend services are the load-bearing walls of software systems. They must be reliable under partial failure, observable in production, and secure against malicious input. This skill applies the engineering discipline required for production-grade services.

When to Use

• Before implementing a new service or significant backend feature
• When a service has reliability, performance, or security issues
• As part of the /review workflow for backend PRs

Process

Step 1: Define the contract

Document: endpoints/methods, request/response schemas, error codes, authentication requirements, rate limits. No implementation before the contract is defined.

Step 2: Design for failure

Every outbound call must have: timeout, retry with backoff, circuit breaker, fallback. Services that don't handle dependency failures will fail when their dependencies fail.

Step 3: Input validation at the boundary

Validate all incoming data before processing: types, required fields, length limits, format constraints. Reject invalid input — don't try to fix it.

Step 4: Idempotency for mutations

All write operations should be idempotent where possible. Use idempotency keys for payment and other sensitive operations.

Step 5: Database access

• No N+1 queries
• Transactions for operations that must be atomic
• Read replicas for read-heavy workloads
• Connection pooling configured

Step 6: Logging and observability

Implement structured logging, metrics (four golden signals), and distributed tracing. See observability-and-monitoring.

Step 7: Security

• Authenticate every request
• Authorize every resource access
• Validate all inputs
• No secrets in logs or responses See security-and-hardening.

Step 8: Graceful shutdown

Handle: SIGTERM, drain in-flight requests, close database connections, flush logs. Services that don't handle graceful shutdown cause errors during deployments.

Verification Requirements

• Contract documented before implementation
• All outbound calls have timeout and retry
• Input validation at the service boundary
• No N+1 queries
• Authentication and authorization on every request/resource
• Graceful shutdown implemented
• Four golden signals instrumented