Distributed search — Azure
Service mapping
- Azure AI Search (formerly Cognitive Search) — managed search service: inverted-index full-text (BM25), facets, suggesters/autocomplete, indexers that pull from data sources (Blob, Cosmos DB, SQL), and optional vector/semantic ranking. The default managed search on Azure; it abstracts shards behind replicas (query throughput/HA) and partitions (index size/capacity).
- Self-managed OpenSearch/Elasticsearch on VMs/AKS — when you need full engine control or portability; then the generic recipe applies.
When to pick which
AI Search when you want a managed, batteries-included search with built-in indexers and suggesters and don't want to operate a cluster; self-managed OpenSearch when you need low-level shard/analyzer control or cross-cloud portability.
Limits / things that bite (verify against current docs)
- Capacity is set by replicas × partitions (search units); both are bounded per tier and the tier caps total index size and doc count — pick the tier from your corpus bytes, and re-tiering can mean a rebuild.
- Indexers (the pull pipeline) have batch-size and run-frequency limits; large or fast-changing corpora can lag — monitor indexer freshness.
- Per-tier caps on indexes, fields, and queries-per-second.
Pitfalls
- Under-provisioning partitions for a growing corpus, then hitting a tier change.
- Relying solely on built-in indexers for near-real-time freshness without a
push path / durable buffer in front (→
messaging-streaming). - Treating it as a primary store rather than a reindexable derived copy.
- Lock-in: skillsets, indexers, and semantic ranking configs are Azure-specific.