How Uber Builds Fault-Tolerant OpenSearch Clusters Across Zones

Uber shares its recipe for keeping OpenSearch running when an AZ goes down. No magic AI - just topology, shard placement, and regular drilling.

The fact

InfoQ publishes (July 17, 2026) an experience report from Uber on building OpenSearch clusters resilient to zone failures (AZ). The article details placement patterns, failover, and testing - and specifies that Uber stacks a home-made "isolation-group" system on the container orchestration platform Odin above the OpenSearch primitives.

Our take

Two points to remember. Zone-aware placement is engineering work, not a checkbox. OpenSearch does not magically redistribute shards after an AZ failure - the topology must be thought out in advance (allocation awareness, forced awareness, replicas per zone). Uber details the concrete hooks, with the contribution of its own isolation-group system on Odin - which is missing in 90% of the documentation circulating.

The discipline of drills makes the difference. Predicting failure is easy; repeating it in pre-production, less so. This is what the harness/platform wave (see #1211 at QCon AI Boston) brings to AI: continuous drills, applied to infrastructure research for ten years, become the norm for agents in production.

To watch

The Uber approach (OpenSearch + isolation-groups + Odin) is transposable to other distributed data stores (Cassandra, Kafka). For teams building the RAG/vector layer above OpenSearch: zone resilience is played upstream of embedding - the AI layer inherits the floor that its base sets.