6.7 KiB
Cloud plan enforcement
These limits apply to new Cloud offers (offer version 13). Existing purchased offers retain their saved limits. Self-hosted installations do not use these Cloud quota gates.
| Plan | Monthly price | CPU | RAM | Disk | Projects/apps | Running services | Build admission ceiling |
|---|---|---|---|---|---|---|---|
| Hobby | $12 | 1 vCPU | 4 GB | 50 GB | 3 | 4 | 1 |
| Starter | $24 | 2 vCPU | 8 GB | 128 GB | 10 | 15 | 4 |
| Pro | $47 | 4 vCPU | 16 GB | 250 GB | 25 | 40 | 8 |
| Scale | $119 | 8 vCPU | 32 GB | 600 GB | 50 | 100 | 20 |
Annual charges are $120, $240, $468 and $1,188 respectively, paid upfront. Yearly prices display these full annual amounts per year. Both intervals use the same capacity and application limits. Catalog changes do not rewrite existing paid subscription terms or their renewal prices.
Hobby includes email support. Starter is the featured plan. Projects/apps share one project allowance; services share the subscribed server's capacity. Container resource ceilings do not reserve that amount of memory for every service.
Traffic
Egress is excluded from plan limitations. This change introduces no bandwidth allowances, traffic caps, egress enforcement, or overage charges. Traffic analytics are independent of plan enforcement.
Deployment and upgrade behavior
Project creation and service admission enforce the saved offer on the selected managed server. Queued deployment service reservations count toward the service allowance; drafts do not. Confirmed stopped services release their running slot. Starting a stopped service reserves its slot before the runtime call.
Explicit build resources are validated rather than silently reduced. Capacity refusals allow customers to review the server or upgrade. Plan refusals identify the relevant server and allowance. A successful billing refresh enables retry; the deployment API still validates the request. Failed retries retain the modal and configured project.
Queue and execution ownership
Only explicitly submitted jobs enter the durable queue. Workspace admission is serialized across replicas. FIFO uses first submission time, not draft creation time. Global sweeps interleave workspaces and collect pages before dispatching. Permanent configuration errors terminalize only unclaimed jobs; temporary server errors, rate limits and resource pressure keep jobs queued. The UI shows queue position, the waiting reason, and an optional upgrade action. Maintenance that wins the race between reservation and activation returns only untouched work to the queue atomically; its submission order is preserved.
Migration 0182_cloud_build_activity.sql adds durable execution ownership and
resource reservations; 0183_cloud_build_exclusive_preparation.sql adds the
captured preparation requirement without rewriting an already-applied migration.
Different projects may build concurrently on a Cloud
server. Same-project builders remain serialized. Existing provisioning and
project mutation locks remain in place, and live or unresolved shared activities
exclude server lifecycle operations. Pre-deploy backups retain exclusive
preparation; jobs involving them wait for other activities to settle. The required
backup-policy set is captured at admission, including an explicitly empty set;
subsequent policy edits cannot change the preparation lock requirement mid-flight.
The plan values are ceilings. The scheduler also checks CPU reservations, measured free memory, and host headroom. Unknown measurements wait; explicit requests that cannot fit the purchased server fail with an actionable resource message. Reservations conservatively subtract outstanding memory even when current usage already includes part of a build. Available capacity can therefore permit fewer parallel builds than the plan ceiling. No scheduler can infer arbitrary source code's peak memory usage from its repository alone.
Automatic builds use the server's measured available capacity rather than a fixed 1-vCPU / 4-GB worker. Memory leaves 5% operating-system headroom (128–512 MB) and subtracts outstanding build reservations. Automatic CPU subtracts outstanding reservations from sampled idle capacity, bounded by unreserved server CPU. Running builders may be conservatively counted twice, as with memory. An explicit dimension is preserved; only missing or zero dimensions are automatic. Unknown samples or extreme CPU pressure wait instead of launching a worker. The plan concurrency allowance is a ceiling; automatic builds can consume the available budget and queue subsequent builds. Smaller explicit build limits permit concurrent builds when they fit. The minimum is 0.25 vCPU and 128 MB RAM. Automatic allocation is selected at admission and stays fixed for that worker; it is not resized during a build. An explicit request waits for its chosen resources rather than being silently reduced. Allocated resources are printed in the build log. Available capacity is not a guarantee that every application's compiler fits.
Single-app builds release their slot after source building and before deployment. Compose releases after its image-build batch. Image-only deployments do not reserve a build slot; retained-artifact operations reserve for a possible source fallback and release once reuse is confirmed. Slot release does not release host command ownership. Unconfirmed commands retain their reservation for recovery. Docker-owned BuildKit workers must also be confirmed absent before capacity is released; stopping an exec process alone does not prove its worker stopped.
Execution locks cover backups, build, deployment and acknowledgement. Detached workers cannot inherit the submitting request's lock ownership. Recovery must try-acquire that exact execution lock and cannot stop a live sibling's commands. An interrupted execution is marked failed only after ownership is obtained; its commands and owned artifacts are recovered before capacity is released. Jobs that were still waiting remain queued and resume automatically. Builds already interrupted during execution require redeployment rather than silently replaying release commands or migrations.
Queue dispatch and command recovery have separate periodic jobs so slow recovery does not prevent ordinary dispatch. No timeout is treated as proof that a remote command stopped. Linked installations retain their existing exclusive protocol; the shared activity protocol is restricted to Cloud-local managed servers.