Reviewed 2026-10-25
What is verified, and what is not
Levelrail labels every feature by how well it is tested. See what is verified. This page lists what has evidence behind it, what was checked on one host only, and what has no evidence yet. Each row names its source and its date.
Verified, with the limit stated
Each row has a test or a dated measurement behind it. None of them is a run on a fresh VPS with a real public domain, so read the caveat in the same row.
| Claim | Evidence and caveat | Source | Date |
|---|---|---|---|
| Control plane idle footprint | Idle RSS 200 MB (peak 253 MB) and 1.5 percent CPU, 273 goroutines, on a 2 vCPU, 8 GB Linux test box with all features at defaults and no apps under load. One live instance, one measurement. | BenchmarksControl plane idle footprint | 2026-10-19 |
| Deploy approvals | Labelled stable. A live end-to-end test on one machine shows a second user's approval is the only thing that cuts over. No real infrastructure run. | Feature statusDeploy approvals | Evidence 2026-09-26, e2e rechecked 2026-10-05 |
| Postgres point-in-time restore and backups | Labelled stable. A live end-to-end test restores to before a marker row using an S3 compatible bucket, plus live Docker dump and restore tests, all local. The real S3 and R2 upload test is skipped in CI. | Feature statusPITR and database backups | Evidence 2026-09-26, e2e rechecked 2026-10-05 |
| Post-deploy watch and automatic rollback on hard signals | Labelled stable for crashloop, out-of-memory and 5xx signals: live end-to-end tests on a real Docker engine and real ingress on one machine roll back to the previous pinned image. The latency signal is beta and nothing was observed under real production traffic. | Feature statusPost-deploy watch and auto rollback | Evidence 2026-09-26, e2e rechecked 2026-10-05 |
| Sleep when idle | Labelled stable. A scripted live run against a real Docker daemon and the real ingress passes 10 checks (idle stop, wake on request, hold mode, manual wake, fleet counts). Per app and off by default. | Feature statusSleep when idle | Evidence 2026-09-26, e2e rechecked 2026-10-05 |
| Pipelines, deploy freeze, supply chain gate, status page, feature flags, IAM | Each is labelled stable and has a live end-to-end test through the real API on one machine. The page lists no real-infrastructure run for any of them. | Feature statusSummary table | Evidence 2026-09-26, e2e rechecked 2026-10-05 |
| Generic webhook and SMTP email notifications | Labelled stable. A live end-to-end test against a local receiver and a local SMTP server covers delivery, retry of a transient failure and delivery history. | Feature statusNotification channels | Evidence 2026-09-26, e2e rechecked 2026-10-05 |
Verified on one host only
These ran on a single machine or in a local lab. Nothing here shows the same result on another host, another architecture or the public internet.
| Claim | Evidence and caveat | Source | Date |
|---|---|---|---|
| Template catalog start-up | 359 templates in the catalog, 100 smoke-tested, 93 start healthy. Measured on one arm64 Mac with a smoke harness that emulates compose, not on a VPS. The separate end-to-end fleet that deploys samples is 16 templates. The other templates pass structural checks only. | Template statusTemplate status | 2026-10-25 |
| Earlier top-50 template run | 48 of 50 passed `docker compose up --wait` with healthy containers on an arm64 host. One failure needs an operator decision on a Docker socket mount, one pull failed because the pinned image has no arm64 build. | Template statusTemplate status, last run table | 2026-10-19 |
| Multi-node join flow | Labelled beta. Enrollment, cordon, drain with real container relocation and node pinning were verified with two agent processes, each with its own Docker daemon, on one machine. The WireGuard mesh and cross-host transport are not part of that run. | Feature statusMulti-node and WireGuard | Evidence 2026-09-26 |
| Load balancer | Hidden behind a flag. A live end-to-end test with real traffic and drain passes on one machine. It has not been exercised across real hosts. | Feature statusLoad balancer | Evidence 2026-09-26 |
| SAML and OpenID Connect sign-in | Labelled beta. SAML ran against Keycloak 26.0.8 and SimpleSAMLphp 1.15, and generic OIDC against Keycloak 26.0.8, in a local lab. Not tested against Okta or Entra ID. | Feature statusSummary table and Identity and access | 2026-10-20 |
| MySQL, MariaDB and MongoDB backup and restore | Labelled beta. Live end-to-end tests dump, verify and restore through the real API from a local run. MySQL and MariaDB also have point-in-time restore. MongoDB has none. | Feature statusSummary table | Evidence 2026-09-26, e2e rechecked 2026-10-05 |
| Vendor notification channels | Labelled beta. 15 vendor kinds and browser push were tested against local fakes that check each vendor's documented payload and auth header. No real vendor account has accepted a message. | Feature statusNotification channels (all 18) | Evidence 2026-09-26, e2e rechecked 2026-10-05 |
| Upgrade restore points | Labelled beta. A harness upgrades the real v0.8.1 and v0.8.2 binaries with seeded data and checks the restore point, an automatic restore after a failed verification and a refused downgrade, on one machine. | Feature statusSummary table | Evidence 2026-09-26 |
| Load driver ceiling | A load driver and a fixture app on an Apple M4 Max over loopback held 5000 requests per second with 0 errors. That is the ceiling of the driver and the fixture, not a figure for a Levelrail deployment. | BenchmarksLoad driver against the fixture app | 2026-10-18 |
Not yet verified
The evidence for these is missing, thin or pending. Absence of a row elsewhere on this page does not mean it works.
| Claim | Evidence and caveat | Source | Date |
|---|---|---|---|
| A stable release | Not yet labelled 1.0. The changelog lists 0.8.4 (2026-10-24) as its newest dated release and v0.8.6 as unreleased. APIs, the app spec format and the on-disk layout can change between releases. | ChangelogTop of the changelog | 2026-10-25 |
| Runs on real infrastructure | The feature status page says almost no feature has a run on a fresh VPS with a real public domain and real vendor endpoints. Multi-node enrollment is the one documented exception, and it was local. | Feature statusEvidence tiers | Evidence 2026-09-26 |
| WireGuard mesh and cross-host remote transport | Still unverified. Multi-node stays beta. | Feature statusMulti-node and WireGuard | not yet |
| Throughput, latency and failover through the ingress on a Linux server | Not yet measured. The benchmarks page marks anything outside a dated, hardware-labeled row as not measured and claims no limit for it. | BenchmarksTop of the page | not yet |
| Real S3 or R2 upload, real email providers, real traffic export destinations | No live account has been exercised. The real S3 round trip test is env-gated and skipped in CI. | Feature statusPITR and database backups, Email delivery, Traffic export | not yet |
| Spend caps, ingress abuse defense, automatic cleanup | Labelled beta, backed by unit-level tests. No end-to-end test, no run against real traffic or a real disk filling up. | Feature statusSummary table | not yet |
| Resilience under a real process kill on a real host | Proven in tests with a stalled control plane and a sealed snapshot. A real process kill under load on a real host is listed as not proven. | Feature statusResilience check | not yet |
| GPU models, Cloudflare tunnel, in-app AI chat | Hidden behind flags. Tested with fakes only, or needing hardware or an account no test touches. | Feature statusSummary table | not yet |
How to read this
Stable, beta and flag-gated labels come from the feature status page, where stable means unit tests, a doc page and either a live end-to-end test or a documented live verification. It does not mean tested on production infrastructure. Run Levelrail beside your current setup first, and read the benchmarks and the template status before you size or trust anything.
Questions
Is Levelrail stable?
Levelrail labels every feature by how well it is tested. See what is verified. A single node is the well-tested path. The rows above say what each claim rests on.
Where do these numbers come from?
Each row names the page in the project documentation that holds the test or measurement, and a date. Anything without a dated, hardware-labeled row is listed as not yet verified.
How often is this page updated?
It was reviewed on 2026-10-25. When a new run changes a row, the row changes and its date moves.