Shared vs isolated compute
Shared compute and isolated compute trade flexibility and per-second cost against isolation and predictable performance. This page lays out the practical differences so you can pick per environment.
At a glance#
| Question | Shared | Isolated (Dedicated) |
|---|---|---|
| Dashboard option | Tenancy → Shared | Tenancy → Isolated on an environment (Dedicated on a deploy) + Machine type |
| What you pay for | The CPU and memory each running replica requests, per second | Each leased node, per second, regardless of utilisation |
| Resource ceiling per replica | 4 vCPU / 8 GiB | The machine type's capacity |
| Resource floor per replica | 0.125 vCPU / 256 MiB | 0.125 vCPU / 256 MiB |
| Memory per vCPU | 1–4 GiB per vCPU | No ratio rule |
| Node sizes | Not applicable | 2 or 4 vCPU today. The current list is under Machine type. |
| Other tenants on the machine | Yes, isolated by container boundaries | No |
| Node metrics on the Nodes page | No | Yes |
| Stopping workloads | Billing for a stopped workload stops immediately | Billing stops when nothing remains on the node |
Choose shared compute when#
- Traffic is light, bursty or unpredictable, and you would rather pay only for what runs.
- You want small replicas, well under a single vCPU.
- The environment is staging, preview or internal, where noisy neighbours are an acceptable trade-off.
- You don't want to think about machine sizes at all.
Choose isolated compute when#
- You need predictable performance with no other tenants on the machine.
- You run enough replicas to keep a node busy, so a fixed node price costs less than per-request billing.
- The workload is stateful, such as a database, or compliance requires single-tenant machines.
- You want node-level metrics for capacity planning.
How the cost shape differs#
On shared compute, cost follows your requests. Doubling replicas doubles the cost, and removing a replica removes its cost. On isolated compute, cost follows nodes. Adding a replica that still fits on the current node costs nothing extra, and one that doesn't fit leads to another node being leased. For current rates, use the pricing calculator, which covers both models.
Mixing models#
Tenancy is set per environment, and a deploy can override it. A common setup is shared compute for staging and the marketing site, and dedicated nodes for the production API and database. You can change an environment's tenancy later from Projects→Edit.