Scale to zero
How idle Silos databases spin down their compute, and what a cold start means for your first query.
Because a Silos database's durable state lives in object storage, the compute that serves it doesn't have to stay running. When a database goes idle, its runtime spins down — and you stop paying for compute until the next query. That's scale to zero.
How it works
The orchestrator watches each database's traffic. When a database stops receiving queries, it scales the runtime down to zero. The data is untouched and safe in object storage; only the compute goes away.
When a query arrives for a database that's scaled to zero, the orchestrator materializes it again — loads the latest snapshot, replays any newer WAL, and starts serving. This wake-up is the cold start.
Idle
No traffic. The runtime is scaled to zero and incurs no compute cost. Durable state stays in object storage.
A query arrives
The orchestrator selects a runtime and materializes the database from its snapshot plus newer WAL segments.
Serving
The database is current and serving queries on the real PostgreSQL engine, until it goes idle again.
Cold starts
The first query after a database has been idle pays the materialization cost — the cold start. Subsequent queries hit a warm runtime and don't.
Silos is engineered for fast cold starts, and low cold-start latency is an active design target — but it is not a benchmarked guarantee. Don't treat a specific cold-start time as an SLA. For latency-critical paths, design for the first-query cost or keep the database warm.
Designing around cold starts
- Hot paths: keep a frequently-used database warm by ensuring it sees regular traffic, so the first user request doesn't pay the wake-up cost.
- Spiky and idle workloads: lean into scale-to-zero — these are exactly the workloads that benefit most, since you pay nothing while idle.
- Background and batch jobs: a one-off cold start at the start of a job is usually negligible relative to the work itself.
What you pay for
Scale-to-zero means no compute charge while idle. You still pay for the durable state held in storage. Billing is usage-based across compute time, storage, and data transfer. See Pricing & billing.