Operate
Status
An honest, as-of-now snapshot of what's implemented versus what's specified but not
yet built. architecture.md, configuration.md, security.md, and api.md
document the contract: the addenda's design, which is binding regardless of what's
on disk on any given day. This page documents the implementation, as rechecked on
2026-09-24 against services/, deploy/ and apps/web, through Addendum v0.5 (identity,
audit, dashboards and APM) and Addendum v0.7 (contact leads). Treat it as a snapshot, not a
guarantee, and prefer it over guessing from the source tree yourself.
What shipped after 2026-09-24 is not in the tables below: single sign-on and the second factor
(identity.md), stored logs (logs.md), ingestion control
(ingestion.md), Orbit SQL (sql.md), the AWS bill in FinOps
(finops-aws.md) and the S3 archive (archive.md). Where
this page and one of those disagree, that page is right.
This page tracks the project's status. There is no hosted uptime page, because there is
no multi-tenant SRE Orbit service to monitor. See architecture.md for why: nothing
about this product runs anywhere except your own account.
What works, piece by piece
| Piece | Status |
|---|---|
| Users, sessions and roles | Implemented and tested. Local accounts in orbit-api, signed and revocable 12 h sessions in an HttpOnly cookie, three built-in roles and custom ones over twelve permissions, checked on every route; the first admin from the Secret, from the demo accounts or from a one-time setup code; the login, account and admin pages in the Deck. cargo test in services/rust/orbit-api covers the rules, including a test that every route asks for a permission. See identity.md. |
| Audit trail | Implemented. Every sign-in, change and refusal as a fsynced JSONL event, with a histogram, facets, a search language and CSV export in the Deck and over the API; retention by day. Not tamper-evident by itself. See audit.md. |
| Dashboards | Implemented. Five built-in dashboards, an editor with six panel types, rows, thresholds and PromQL autocomplete, query, custom and interval variables, the JSON model, import from Grafana (converted in the browser), and versioned saves that refuse to overwrite a newer one. See dashboards.md. |
| Traces and APM | Implemented. An OTLP/HTTP receiver in orbit-stack with a write-ahead log, an in-memory head and checksummed segments; span metrics and service graph series under Tempo's names; services, service, map, search and trace pages in the Deck; the demo shop in compose and behind demo.shop.enabled in the chart. See apm.md. |
| Alerts | Implemented, derived. Six rules over the cluster's objects (there is no Alertmanager in this build), each with a runbook section in alerts.md; acknowledgement and silences kept by orbit-api and audited. |
| Contact leads | Implemented (Addendum v0.7). Stored under ORBIT_DATA_DIR/leads before the site's form is answered, listed at /deck/admin/leads, optionally POSTed to ORBIT_CONTACT_WEBHOOK_URL. |
| Nocturne AI | Implemented. Streams through web and orbit-api; its cluster tools read orbit-api with the asking user's session, so it reads what that person may read. With no provider configured it runs the mock provider. |
orbit-api live source | Implemented for Kubernetes; costs are an estimate. Clusters, nodes, workloads, events and PersistentVolumes through the Kubernetes API (kube-rs); usage from metrics.k8s.io or the engine (ORBIT_USAGE_SOURCE); FinOps from an embedded us-east-1 on-demand price list, labelled "source": "list-price". AWS Cost Explorer is not wired. |
orbit-stack, the metrics engine | Implemented and measured. Remote-write in, PromQL out, the figures in benchmark.md. The WAL, spill and compaction of the latest round are not yet measured by a report under bench/runs/. |
| Trace benchmark against Grafana Tempo | Run, round 2, re-run by an independent critic. The earlier trace figures are withdrawn. Ahead on reading a trace by id, on ingest and on memory, and on the services summary and the service map at 50,000 traces; search ahead only for values common in recent data and slower than Tempo for a rare or absent value at 3.4 M spans; the summary and the map slower on the first read after a flush. The figures, with the loss table, are in benchmark.md and CLAIMS.md only. |
web, its proxy and its own auth | Implemented. The /api/* proxy (GET, POST, PUT, PATCH, DELETE) attaches the bearer and the session server-side, refuses cross-site mutations, and gates workload writes on the operator token; HTTP Basic in front of every route but /healthz, which web now answers itself. pnpm check:proxy and pnpm check:auth prove both. |
| Helm chart | Implemented. Everything above, plus components.orbit-api.persistence and singleWriter, the identity keys in the Secret, demo.shop.*, and WEB_TRUST_PROXY derived from the ingress. |
| Compose | Implemented. Five services with orbit-shop, the named volumes orbit-api-data and orbit-stack-data, demo data and demo accounts by default. |
| Plain manifests | Implemented, with two gaps. orbit-api's data directory ships on an emptyDir (a claim is a documented hand edit, install/kubectl.md), and there is no demo shop on this path. |
ORBIT_AUTH_MODE=oidc | Not implemented. Refuses to start with a named error rather than ship half a verifier. Single sign-on for people is separate: an admin configures Google or an OpenID Connect provider at Admin > Integrations (identity.md). |
Known ceilings
Each of these is a ponytail: comment in the source naming its upgrade path, or a rule in
CONTRACT.md:
orbit-apiruns one replica. Identity, dashboards and the audit trail live in a single-writer file store; a rollout or a node loss is a short gap in the console. The upgrade path is Postgres.- The audit trail is scanned, not indexed. A query reads the day files of its window.
- No SCIM, no per-cluster or per-dashboard scoping.
- Traces are OTLP/HTTP only. No gRPC listener, no OTLP metrics or logs, no sampling, single node, and span links and scopes are not stored.
- Alerts are derived. Six rules, always firing, with no routing and no history.
- FinOps is a list-price estimate. No invoice, no month-over-month, no forecast.
- The engine is single node. No replication, no object storage, remote-write 1.0 only.
- Compose records
web's address for every visitor in the audit trail, because nothing in front ofwebnames the client (audit.md).
What this means if you're installing today
- Sign in first. Every Deck page needs a user. Put the first admin in the Secret
(
ORBIT_ADMIN_USER,ORBIT_ADMIN_PASSWORD_VERIFIER), or read the one-time setup code fromorbit-api's log; compose and demo data createadmin,editorandviewerwith the install's demo password, printed once in that log. Seeidentity.md. - Give
orbit-apia volume.values-prod.yamlandvalues-eks.yamldo; the chart's defaults and the plain manifests use anemptyDir, which loses every account and audit line with the pod. Back the volume up (operations.md). - Leave
ORBIT_TRUSTED_PROXIESempty.webis trusted through its bearer, and the client address comes from the ingress (WEB_TRUST_PROXY). Behind a load balancer or a CDN, the edge has to pass the client on, or the audit trail records the load balancer (audit.md). - Send traces over OTLP/HTTP to
orbit-stackon8082with its bearer, and name your Collector's namespace incomponents.orbit-stack.networkPolicy.allowFromNamespaces(apm.md). - Production is a live cluster or nothing:
values-prod.yamland the plain manifests setORBIT_DEMO_DATA=falseand bind the read-only ClusterRole;ORBIT_AUTH_MODE=oidcstill refuses to start.
Where to go next
- What changed and why:
changelog.md. - The target you're installing toward, regardless of today's gaps:
configuration.md,architecture.md. - The root
README.mdcarries the same honesty at repo-root scale, for a reader who hasn't openeddocs/yet.