Hypertenant analytics · should be easy

rubberduck

Hypertenant OLAP on Cloudflare.

Scale-to-zero DuckDB over R2 Parquet. Per-org ducklings, one wrangler deploy, graduate compute without moving the lake.

$curl -X POST …/query -d '{"sql":"SELECT count(*) FROM lineitem"}'

session · org_tpch

httpPOST /v1/orgs/org_tpch/query

eventsession.started org="org_tpch"

eventduckling.placement tier=worker

eventstorage.read r2://org_tpch/lineitem/*.parquet

sqlSELECT count(*) FROM lineitem

→ 600572 rows · wall 174ms

eventsession.completed

Three services. One deploy. SQL over Parquet.

No ceremony. Need more than a Worker later? Scaling.

Workers + R2 + D1. Two Workers because WASM exceeds the 10 MiB script cap — platform constraint, not aesthetic split.

Scale to zero between queries. ~$5/mo Workers Paid floor when idle — CPU-ms and egress still meter at volume (see limits).

DuckDB beats SQLite/D1 by 25–98× on groupbys at 1M rows in-process. Cold client latency is higher; see benchmarks.

→

Query org_tpch or org_clickbench now. Bearer demo-token.

Scale-to-zero DuckDB over R2 Parquet. Per-org ducklings, one wrangler deploy, graduate compute without moving the lake.

Workers + R2 + D1. Two Workers because WASM exceeds the 10 MiB script cap — platform constraint, not aesthetic split.

Scale to zero between queries. ~$5/mo Workers Paid floor when idle — CPU-ms and egress still meter at volume (see limits).

DuckDB beats SQLite/D1 by 25–98× on groupbys at 1M rows in-process. Cold client latency is higher; see benchmarks.