Datadog migrates its live Stream Router from FoundationDB to PostgreSQL with Claude and Cursor; 45-minute operations now take ~1 second
Stream Router, the Datadog control plane that tells metrics producers and queriers which Kafka cluster, topic and partitions to use, was hitting FoundationDB transaction size limits. Engineers redesigned the relational schema by hand, then used Claude and Cursor for method-by-method refactoring: Claude first wrote markdown documents on the intent of each function in the old KV controller, then for each method the team stubbed a PostgreSQL implementation, ran the end-to-end suite, and fed failing tests, old code, schema and docs to the model until tests passed. A blue/green deployment with a validator comparing responses every 30 seconds ran for weeks before cutover. The project took about 3 months from design to production, with operations once estimated at 45 minutes now taking about 1 second and database costs down 90%.
The numbers (as reported)
| Design to production | ~3 months (Dec 2025 – Feb 2026); proof of concept in 4 weeks |
|---|---|
| Slowest operations | Estimated 45 minutes → ~1 second |
| Routing dataset size | 40x smaller in PostgreSQL |
| Latency | Hundreds of ms → a few ms |
| Database cost | Down 90% |
| Production incidents during deployment | None |
Facts
- Company / subject
- Datadog
- Tools
- Claude, Cursor
- Codebase
- brownfield
- Work type
- Migration
- Stack
- PostgreSQL, DuckDB, FoundationDB, RocksDB, Kafka, gRPC
- Published
- 2026-06-23
- Last checked
- 2026-11-11
What the source claims Practitioner (own blog/paper)
Claimed by the builder, writing about their own work. Not checked independently.
- Operations estimated at 45 minutes now complete in about 1 second 45 minutes → ~1 second
- Database costs fell 90% after retiring FoundationDB -90% database cost
- Deployment to production proceeded without incident
“Deployment to production proceeded without incident.”
- The team says it would not have finished in time without AI tools
“We would not have completed the implementation in the time frame we had without AI tools.”
Caveats
Self-reported by the Datadog team that did the work. The AI tools were not autonomous: humans designed the schema, reviewed every generated skeleton, and wrote the optimised SQL (batching, UNNEST, CTEs), which the models then copied. Most of the performance and cost gains come from the new relational design, not from the AI tools; the post credits AI with making the implementation possible in the time frame. No count of methods, PRs or lines is given, and token spend is described only as significant.
More stories by tool: CursorClaude
Related topics: MigrationsBrownfield codebases