> ## Documentation Index
> Fetch the complete documentation index at: https://docs.levelrail.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Known limits

> Current Levelrail limits by area: release maturity, multi-node, features not yet verified on real infrastructure, upgrades, templates and beta flags.

Last updated 2026-11-08 (release/v0.9.5 branch).

Current limits, each with the page it comes from. Labels and evidence change with every release, so check [feature status](https://docs.levelrail.com/feature-status.md) for the latest.

## Release maturity

* There is no stable release yet. Every release so far is a beta. Source: [roadmap](https://docs.levelrail.com/roadmap.md).
* Single node is the stable scope. Multi-node, stacks and flagged features are beta or experimental. Source: [feature status](https://docs.levelrail.com/feature-status.md).

## Multi-node

* The join flow is verified across two real Docker daemons, not across a real WAN. Source: [roadmap](https://docs.levelrail.com/roadmap.md), [multi-node](https://docs.levelrail.com/multi-node.md).
* The WireGuard mesh and cross-host remote transport are unverified outside fakes and local Docker daemons. Source: [roadmap](https://docs.levelrail.com/roadmap.md).
* Multi-cloud node onboarding has no run against a real cloud metadata endpoint. Source: [feature status](https://docs.levelrail.com/feature-status.md).
* Capacity cleanup covers the control plane node only. Remote nodes monitor and block. Source: [feature status](https://docs.levelrail.com/feature-status.md).

## Not verified on real infrastructure

* A stable label means tests plus a live end-to-end run on one machine. It does not mean a run on a fresh production host. Source: [feature status](https://docs.levelrail.com/feature-status.md).

* The build log's builder line (Dockerfile or Railpack, plus the buildpack file warning) is unit tested through its helper only. No real BuildKit build has run with it, because the build client has no seam to run without BuildKit. Source: [importing apps](https://docs.levelrail.com/importing-apps.md), [feature status](https://docs.levelrail.com/feature-status.md).

* Public ACME has one recorded run on a live VPS. Renewal close to expiry, DNS-01 wildcards and a deliberate failure against the production CA are not yet exercised live. Source: [roadmap](https://docs.levelrail.com/roadmap.md), [ACME runbook](https://docs.levelrail.com/acme-verification-runbook.md).

* Idle footprint was measured in a Docker Desktop container and, on 2026-11-08, on an arm64 Ubuntu VM (build 243a8ee00, 12 containers running): 184 MiB resident and about 1.3% of one core over five idle minutes. Other hosts and sizes are still open. Source: [roadmap](https://docs.levelrail.com/roadmap.md).

* Load and failover numbers have no published results yet. Source: [roadmap](https://docs.levelrail.com/roadmap.md), [benchmarks](https://docs.levelrail.com/benchmarks.md).

* The resilience checks are proven in tests. A real process kill under load on a real host and the container restart policy on a real reboot are not proven. Source: [feature status](https://docs.levelrail.com/feature-status.md).

* The 15 vendor notification kinds have never had a real vendor accept a send. Source: [feature status](https://docs.levelrail.com/feature-status.md).

* SAML and SCIM were run against Keycloak and SimpleSAMLphp in a local lab, not against Okta or Entra ID. Source: [feature status](https://docs.levelrail.com/feature-status.md).

* Proven on a real host (arm64 Ubuntu VM, 2026-11-08, build 243a8ee00): `apps run`, `apps scale` and `apps restart --service`; runtime options in `docker inspect` and a single container recreate on edit; the compose Postgres backup schedule, typed-confirm restore with a safety dump, `databases export` and `databases clone`, plain dump `databases import` and the Redis import guard across databases; the per-domain request body cap (413, enforced when the application reads the body); the Valkey TLS toggle with a `rediss` connection; managed databases on a private per-database network, moved live on upgrade without downtime; one-off run access rules (a deploy-only token is refused, root is allowed, a read-only token gets the output withheld); and an SSH deploy key manual deploy and git poller with trust-on-first-use host key pinning.

* Still no real-host run for: the basic auth gate on static previews (401 without credentials), and SSH deploy key host key pinning from a pasted key.

* The 1Password Connect provider is tested against a local fake API only. The Add provider save from the dashboard is not confirmed in a browser run. Source: [vault secret references](https://docs.levelrail.com/vault-secret-references.md).

* The Infisical machine identity login is tested against a local stand-in server only. Source: [vault secret references](https://docs.levelrail.com/vault-secret-references.md).

* The KeyDB health probe is unverified on the image, and the Dragonfly probe has no run against a real image. Source: [managed databases](https://docs.levelrail.com/managed-databases.md).

* The preview migration step is tested with a fake runner, not against a live Postgres database. Source: [preview environments](https://docs.levelrail.com/preview-environments.md).

## App controls

* The one-off command, per-service replica and restart controls, and the Logs service selector are unit and component tested with mocked fetch. None has run against a live server, and none has a browser check at phone or desktop width. The Logs selector's navigate-on-change path has no test. Source: [feature status](https://docs.levelrail.com/feature-status.md).

## Upgrades

* Upgrade restore points are beta. Systemd and agent reconnection after an upgrade, and a live host run, are not yet proven. Source: [feature status](https://docs.levelrail.com/feature-status.md), [upgrade safely](https://docs.levelrail.com/upgrade-safely.md).
* Restoring across a schema change discards data written since the pre-upgrade snapshot. The attempt records how many minutes that was. Source: [upgrade safely](https://docs.levelrail.com/upgrade-safely.md).
* A release older than your database schema is blocked and not offered. Source: [upgrade safely](https://docs.levelrail.com/upgrade-safely.md).
* After the first verified release signature, an upgrade to a release without a signature bundle is refused. Source: [upgrade safely](https://docs.levelrail.com/upgrade-safely.md).
* Upgrade and downgrade testing across versions with real data is still part of the work toward a stable release. The populated-database SQLite test uses synthetic rows, and the domains table is not seeded yet. Source: [roadmap](https://docs.levelrail.com/roadmap.md).

## Templates and databases

* The last local catalog smoke (2026-10-25, plus eight hand-booted on 2026-10-27) passed 101 of 108 templates that need no operator configuration. The rest of the catalog is untested. Source: [feature status](https://docs.levelrail.com/feature-status.md), [templates status](https://docs.levelrail.com/templates-status.md).

* Point-in-time restore covers Postgres, MySQL and MariaDB only. MongoDB has no oplog here. TLS is built for Postgres, Redis, Valkey, KeyDB, MySQL, MariaDB and MongoDB (on at creation for Postgres, Redis and, from the TLS release on, MySQL, MariaDB and MongoDB; older MySQL, MariaDB and MongoDB databases, Redis and every Valkey and KeyDB database can be switched with `PUT /api/v1/databases/{name}/tls`, and Valkey and KeyDB stay plaintext until switched). A local Docker TLS handshake passed for KeyDB. Dragonfly is not enabled yet: its server refuses to start with TLS and no password, so the toggle rejects it. ClickHouse, Meilisearch, Qdrant, Typesense and Memcached have no TLS, and no TLS path except the Valkey toggle with a `rediss` connection (arm64 Ubuntu VM, 2026-11-08) was run on a real host. Source: [roadmap](https://docs.levelrail.com/roadmap.md), [feature status](https://docs.levelrail.com/feature-status.md).

* Database backups for compose services (Postgres, MySQL, MariaDB) can be restored through the API, `apps backups restore` and a Restore button in the dashboard card (root only). The schedules, run history, dashboard card and restore are built and unit tested, and the Postgres schedule and typed-confirm restore ran on a real host (arm64 Ubuntu VM, 2026-11-08). MySQL and MariaDB compose services and the dashboard card have not. Source: [compose apps](https://docs.levelrail.com/compose-apps.md), [feature status](https://docs.levelrail.com/feature-status.md).

* `databases export` and `databases clone` are unit tested with fakes and ran against a managed Postgres database on a real host (arm64 Ubuntu VM, 2026-11-08). Other engines have not. Source: [managing databases](https://docs.levelrail.com/managing-databases.md), [feature status](https://docs.levelrail.com/feature-status.md).

* `databases import` restores into an empty database, or over existing data only with `--force`. The emptiness check counts tables, so a Postgres database with tables but no rows is refused. The check and the restore are not atomic. A Redis restore stops and restarts the container, and backup schedules are not paused during an import. A plain dump import into Postgres and a Redis import with keys in a non-default database ran on a real host (arm64 Ubuntu VM, 2026-11-08); other engines have not. Source: [managing databases](https://docs.levelrail.com/managing-databases.md), [feature status](https://docs.levelrail.com/feature-status.md).

## Behind flags or beta

* Seventeen features sit behind the platform gate and are off by default. Ten are experimental: AI models on GPU nodes, the load balancer, platform as code, Cloudflare Tunnel, access roles, global environments, AI control, plugins, mail DNS and remote MCP. Seven are beta: AI assistant chat, resource advisor, CLI console, mobile API, git push deploy, canvas, and SAML and SCIM. Turn one on in Settings, Beta features, with `levelrail-cli settings beta-features set <feature> on`, or list it in `APP_EXPERIMENTAL`. Features wired at server start show "restart to apply" until the control plane restarts. Add `--restart` to the CLI command to restart it at once (root or a supervised install). Source: [experimental features](https://docs.levelrail.com/experimental-features.md), [feature status](https://docs.levelrail.com/feature-status.md).
* Saved `abuse.*` runtime settings reach the ingress limiter on the next policy save and proxy refresh; unit tested only, not benchmarked on real traffic. Source: [ingress abuse defense](https://docs.levelrail.com/abuse-defense.md), [feature status](https://docs.levelrail.com/feature-status.md).
* `git remote add --apply` and `rotate --apply` keep the push token out of `.git/config` by default and read it from `LEVELRAIL_GIT_TOKEN` at push time. With `--store-token-in-config` the token is stored unencrypted in the repository config, so anyone who can read that repository can push deploys to the app. The flags are unit tested with fakes and have not run against a live control plane. Source: [git push deploy](https://docs.levelrail.com/git-push-deploy.md).
* Git push, webhook and manual git builds honor `dockerfilePath` from `levelrail-definition.json` when no Dockerfile path is stored and the base directory has no `Dockerfile`. On a push, `imageName` and `dockerfileLines` are refused with a message; webhook and manual builds ignore the file in that case and fall back to the default build. Source: [git push deploy](https://docs.levelrail.com/git-push-deploy.md).
* The `pause` spend cap action serves a maintenance response and does not stop containers. Source: [roadmap](https://docs.levelrail.com/roadmap.md).
* `migrate coolify` is built and unit tested against fixtures only. No run against a real Coolify instance is recorded, so field names such as the command health check, network and swap fields are unverified. Backup schedules and S3 destinations are listed as manual steps, not created. Source: [Coolify exit gaps](https://docs.levelrail.com/coolify-exit-gap-report.md), [feature status](https://docs.levelrail.com/feature-status.md).
* Database create and Prometheus read bodies are capped (1 MiB and 8 MiB, the latter after snappy decoding) and answer 413 beyond that. Source: [performance](https://docs.levelrail.com/performance.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.