Scaling

Default: Container DuckDB over files in R2. WASM only when the working set is tiny.

Options

	Small (WASM)	Normal (Container)	Huge (R2 SQL)
Engine	DuckDB in a Worker	DuckDB in a Container	Cloudflare scan SQL
Data	Same R2 files	Same R2 files	Same R2 files
Fits	~48 MB Parquet / 64 MB RAM	Multi-GB interactive	Warehouse-scale scans
Idle	scales to zero	sleeps (cache clears)	pay per scan

Move compute, not files

Switch when touched Parquet exceeds ~48 MB, queries spill past 64 MB, or you need full extensions / Iceberg.

     org_acme
        │
        ├── worker (WASM)     ──┐
        │                       ├──▶  r2://…/org_acme/**   (unchanged)
        └── container (native)──┘

Placement is metadata in D1. No lake copy.

R2 SQL

One-off monster scans: "engine":"r2sql". Needs R2_SQL_TOKEN. Dashboards stay on DuckDB.

After Container sleep

Local cache is gone. Next query re-reads from R2. See Limits.

Don’t

• Put analytical tables in D1
• Copy Parquet when changing placement
• Run an always-on warehouse for weekly queries
• Treat WASM and Containers as the same size box