Every query is metered

When data access goes through a semantic layer, you get something that’s usually hard: a single, honest record of usage.

The summary endpoint

Every query the engine runs records its bytes scanned and row count against the principal that asked. GET /v1/tenants/{id}/usage/summary?from=… gives you the total over any window:

{
  "queries": 1284,
  "bytes_scanned": 41203847112
}

That’s it. No per-warehouse export, no reconciling three billing consoles.

Why it’s reliable here

The number is trustworthy because there’s no side channel. A member can’t drop to raw SQL to avoid the meter, and an agent on MCP is constrained to the same three tools as everyone else. If a byte was scanned on your warehouse through datahashi, it’s in the summary.

What we do with it

Today: a limiter and a cache in front of the engine, both keyed off the same usage signal. The cache means a repeated question doesn’t re-scan; the limiter means one runaway notebook doesn’t take the warehouse down for the tenant.

Pre-aggregation is the next step, and it’s deliberately not built yet — the reasoning is in the repo’s known-issues.md.