Error handling in large distributed systems is a global architectural decision, not a local code-level choice. Whether to crash or continue depends on three key factors: failure correlation patterns, whether higher layers can handle failures, and if meaningful continuation is possible. The right approach varies by system architecture—traditional web services handle low error rates through instance replacement, while serverless and Erlang-style systems can tolerate higher crash rates. Blast radius reduction techniques like cell-based architectures and shuffle sharding provide resilience when error handling decisions prove wrong.
Sort: