Uber engineered native gRPC endpoints directly into OpenSearch to eliminate inefficient REST/JSON translation layers within their architecture. Their automated pipeline for syncing JSON APIs with Protobuf schemas, their internal integration strategy, and the resulting performance gains in production environments for high-throughput ingestion and vector search workloads.

- Uber implemented gRPC as an OpenSearch module
- To prevent divergence between REST and gRPC, Uber built a three-stage automated pipeline
- Removing the JSON-to-Protobuf translation layer reduced p99 index write latency by 60% for Uber’s M3 metrics system and decreased batch indexing job runtimes by 20-35%.
- Large vector searches, which serialize poorly in JSON, saw a 53% reduction in p50 latency and a 43% reduction in p95 latency.
- Combining gRPC with binary document formats like SMILE proved highly effective, executing 30% faster than REST JSON and 45% faster than gRPC JSON.

https://www.uber.com/us/en/blog/high-performance-grpc/