What’s on the floor · today

1. 1ChatGPT and API appsChatGPT and the API. Bugs that wrote twice are fixed, lazy writes smooth spikes, clients back off before retrying, and connections go through PgBouncer.
2. 2Rate-limit gateRate limits at four layers: application (this gate), connection pooler, proxy and query. Above a ceiling it sheds traffic, so a sudden spike cannot overwhelm the database.
3. 3CacheServes most reads. Cache locking and leasing: one request per missed key fetches from PostgreSQL; the rest wait for the refill.
4. 4PgBouncer behind a Kubernetes ServiceEach read replica has its own Kubernetes deployment of PgBouncer pods, all behind one Kubernetes Service, co-located with clients and replicas. Average connection time fell from 50 ms to 5 ms in OpenAI’s benchmarks.
5. 5PostgreSQL primaryStill one unsharded writer. It takes the writes and only the reads that are part of write transactions. The ORM can rate-limit or block an expensive query’s digest, and schema changes are lightweight only, with a 5-second timeout.
6. 6Hot standbyA hot standby, continuously synchronised, promoted quickly when the primary fails. Reads carry on from the replicas meanwhile.
7. 7Azure Cosmos DBAzure Cosmos DB takes shardable, write-heavy workloads, and every new table: PostgreSQL accepts no new ones. Write-heavy workloads that are harder to shard are still being moved.
8. 8Read replicasNearly 50 across regions, several per region with headroom, split into high- and low-priority instances; products and services are kept apart the same way. Six are drawn. WAL from the primary, or through a relay with cascading replication (in testing).
9. 9Robot trays and depotEach tray carries one query slip from the apps through the machines and back; one tray stands for many queries. Hover a tray to read its slip.