Skip to content

SYSTEM Cited by 3 sources

OpenSearch

OpenSearch is the Apache-2.0-licensed fork of Elasticsearch + Kibana maintained by the OpenSearch community (originally forked by AWS after Elastic's 2021 license change to SSPL/Elastic License v2). It is a distributed search and analytics engine with a REST API, Lucene-backed indexing, and a Kibana-compatible dashboards component (OpenSearch Dashboards).

How Fly.io uses it

Fly.io runs OpenSearch as the permanent-retention audit trail for token operations from tkdb — "since virtually all the operations that happen on our platform are mediated by Macaroons, this audit trail is itself pretty powerful." (Source: sources/2025-03-27-flyio-operationalizing-macaroons.)

Distinct from Fly's metrics stack (Prometheus) and tracing stack (Honeycomb + OpenTelemetry) — OpenSearch here serves the forensic-audit use case specifically, where arbitrary ad-hoc queries over history are the primary requirement.

Seen in

  • sources/2025-03-27-flyio-operationalizing-macaroons — Fly.io's permanent-retention audit trail for token operations.
  • sources/2026-01-06-lyft-feature-store-architecture-optimization-and-evolution — OpenSearch as the embedding-feature store inside Lyft's dsfeatures. Embedding features require specialised kNN / vector-similarity indexing that DynamoDB doesn't natively provide, so they are split off to OpenSearch while scalar / metadata features stay in DynamoDB + ValKey. Second OpenSearch use-case shape on the wiki: the Fly.io entry above is audit-log / full-text search; Lyft is vector-similarity serving.
  • sources/2026-03-06-pinterest-unified-context-intent-embeddings-for-scalable-text-to-sql — OpenSearch as the substrate for Pinterest's internal Vector Database as a Service. Multiple teams across Pinterest (Analytics Agent table + query search, AI Table Documentation, assorted LLM features) run on one shared OpenSearch-backed platform with Hive as source-of- truth + Airflow for index creation / ingestion DAGs. Supports hybrid queries combining vector similarity with metadata filters ("Tier-1 tables semantically similar to user_actions containing impression data"). Third OpenSearch use-case shape on the wiki: multi-tenant internal vector-DB-as-a-service platform — Fly.io = audit log; Lyft = single-team vector feature store; Pinterest = company-wide vector platform.
Last updated · 542 distilled / 1,571 read