Glosarium
Contents

Architecture & System Design › Cloud Design Patterns

Geodes

Serving any request from any region, with data replicated everywhere.

Backend Engineer · Staff

Also known as: geodes, azure geodes, geo-distributed nodes

Geodes are dense multi-region units: identical full-stack deployments in many regions, each serving nearby users with local data ownership, replicating asynchronously — active everywhere, partitioned by geography, converging where global views need it. The pattern behind planetary-scale services that stay fast locally and available globally.

geode EU: serves EU users + owns EU data ⇄ async replication ⇄ geode US: mirrors
global queries: fan out, merge (eventual); local: single-digit ms

Geodes accept the fundamental trade explicitly: local writes converge globally (conflict design required), global reads merge (freshness bounded), and partitions split cleanly along geode lines (each side fully functional). Quorum and coordination stay intra-geode (fast); inter-geode is asynchronous by design.

The classic mistakes:

  • Synchronous inter-geode coordination. Cross-ocean consensus on hot paths adds hundreds of milliseconds per operation. Asynchronous between geodes; consensus within.
  • Conflict naivety. Concurrent cross-geode writes to shared entities need resolution designed upfront (partitioned ownership preferred; CRDTs/LWW with eyes open otherwise).
  • Residency violations. Replicating regulated data into geodes whose jurisdictions forbid it breaks law, not just latency. Data placement follows legal boundaries first.
  • Geode-count excess. Dozens of geodes multiply the operational surface (every deploy, migration and incident × N) past reason. As few as latency and law allow.
  • Global secondary indexes. Cross-geode queries needing global ordering/indexing reintroduce the coordination geodes avoid. Scope queries per geode; merge at read.
  • Failover fantasy. “Traffic shifts to the next geode” requires its capacity, data (lagged!) and compliance posture to serve those users. Verify, don’t assume.
  • Unowned global state. Counters, sequences and singletons spanning geodes become coordination bottlenecks or correctness hazards. Partition or approximate globally-shared state.

When to build them: planetary scale with local latency needs and partitionable data. Intra-geode consensus, inter-geode async, conflicts designed, laws obeyed — geography as architecture.