Architecture

A Lunora deployment is a small set of Cloudflare primitives wired together by @lunora/runtime. Each piece is single-purpose and replaceable.

            ┌────────────────────────────────────────────────┐
            │                  Worker (Edge)                 │
            │    @lunora/runtime · createWorker({ ... })     │
            └────────────────────────┬───────────────────────┘
                                     │
        ┌────────────────┬───────────┼────────────┬───────────────┐
        ▼                ▼           ▼            ▼               ▼
   ShardDO ×N        SchedulerDO     D1          R2          Queues
   (SQLite)         (alarms)     (.global())  (files)     (background)

Tier 1 — Shard-local

ShardDO is the workhorse. Every non-global table lives inside it. Default sharding routes to a single __root__ DO; .shardBy(field) partitions the namespace so each value of field gets its own DO with its own SQLite, its own CPU budget and its own hibernation timer.

Reasoning rules:

• Real-time subscriptions stay in-process. A delta never crosses DO boundaries; the mutation that wrote the row is also the broadcast source.
• Single-region writes. A DO is pinned to a region near its creator, and writes are strongly consistent within that region.
• Account-unlimited. You can have millions of ShardDO instances per account.

Tier 2 — Global (D1)

.global() tables move to Cloudflare D1. Use this for identities, billing, cross-tenant audit logs, anything that needs to be queryable across shards. Reads can be served by regional replicas; writes route to the primary. @lunora/d1 threads the withSession(bookmark) API so your client gets read-your-writes without a sticky-session router.

Tier 3 — Blob + async

• R2 — @lunora/storage issues signed URLs; the browser uploads directly without proxying through the Worker.
• Queues — for fire-and-forget work; Lunora emits jobs from a mutation via ctx.queue.send(...). Idempotency is on you.
• KV — useful for edge-cached config and feature flags. Lunora does not ship a wrapper; use the binding directly.

Why this beats single-DO

Zeroback's single-DO model caps out at 10 GB SQLite and ~1 000 req/s. Lunora's opt-in sharding moves both ceilings to N × 10 GB and N × 1 000 req/s without changing your call sites. The cost is tooling: codegen has to know how to address shards, and the docs have to be honest about cross-shard queries being fan-out.

See also

• Sharding
• Limits