Cloudflare designed D1's control plane to keep single-region components off the query path

The growth curve is steeper than the headline multiple suggests. Live databases grew 15x over the past year and 65x over two years, as of October 8, 2026 ^[3]. Divide one by the other and the earlier year comes out at about 4.3x ^[16].

Daily active databases grew 18x over the same year ^[4]. A database counts as active if it ran at least one query that day ^[5]. Active databases are therefore growing a little faster than live ones, by a factor of 1.2 if both counts are taken on the same dates ^[17].

The hot-path rule is narrow. Cloudflare says to avoid centralized single-region components in any hot path for data plane operations, and for D1 the hot path is every query request ^[7]. Routing metadata lives in Workers KV, a globally distributed key-value store ^[8]. Central Postgres sits behind Hyperdrive ^[9]. Aggregations are precomputed on a schedule, so a user request does not trigger a scan ^[10]. Components that do not need strong consistency talk through queues and settle eventually ^[11].

Each of those choices gives up some freshness or consistency to keep query latency down. The cache miss is the exposed spot: it is where a query can still reach a central database, and the post describes that as rare ^[8]. The portion of the post available to us stops at its high-level architecture heading, before any component is described. Cloudflare does not publish a miss rate or a query latency figure.

That matters because the post poses its own question this way: how to grow the control plane to millions of databases without routing each query through central chokepoints that slow it down ^[14]. Cloudflare says it load-tested periodically against a target well above expected growth while making the changes ^[12]. The results are not in the text.

Cloudflare publishes one customer endorsement. Jon Abrams of ChatGPT Sites at OpenAI said: "D1 gives our users a simple, scalable way to add features like custom metrics, leaderboards, and comments to their dashboards, games, and tools." ^[13] The post says AI-assisted development platforms are pushing applications toward many small databases, one per user, application, preview environment, or tenant ^[15].

For the pattern to transfer, we think three things have to hold in your workload. Routing lookups must be cacheable enough that most queries hit. Readers must tolerate aggregates that are as old as the last precompute run. The central store must sit behind a pooler with read replicas. If routing state changes on every request, the KV layer becomes the miss path.