There Is No "Now" in Distributed Systems

This piece draws a direct line from Einstein's special relativity — the idea that there is no universal 'now' — to distributed systems engineering. NTP-synchronized clocks are typically only within 1-100 milliseconds of each other, which invalidates the common assumption that a smaller timestamp means an earlier event. This flawed assumption quietly breaks distributed locks, event-sourcing replay, log correlation, and cache invalidation in production systems.

The fix is Lamport's 1978 logical clocks and their extension, vector clocks, which track causality rather than wall-clock time. Systems like DynamoDB, Riak, and CRDTs rely on this approach to resolve conflicting writes. The CAP theorem is reframed as a physical constraint: during a network partition, you must choose between consistency and availability — there's no third option, just as there's no faster-than-light shortcut in physics.

For cases where real wall-clock time is unavoidable, like financial transactions or audit logs, the article points to Google's TrueTime API in Spanner, which returns a bounded uncertainty interval instead of a single timestamp. The practical takeaway: use logical clocks for causal ordering, match consistency guarantees to the specific data's needs, and always treat timestamps as ranges rather than exact points.