From alert to root cause,
inside your own account.
SRE Orbit is self-hosted observability for platform and SRE teams on Kubernetes: metrics, logs, traces, dashboards and an AI SRE in one console. Your telemetry stays in your account and never passes through us.
In the design-partner pilot, the engineers who build it install it with you.
APM/Services/checkout-api
checkout-apiDegraded
Deployment · checkout · 12/12 pods · prod-use1
- Requests
- 6.47K/s
- p99 latency
- 517 ms↑ 33.6%
- Ready
- 12/12pods
- Last deploy
- 7f3ac102:08 UTC
p99 latency
| Series | First | Mean | Max | Last |
|---|---|---|---|---|
| p99 | 387 ms | 464 ms | 595 ms | 517 ms |
- 02:08Deploy 7f3ac1 on checkout-apirollout 12/12
- 02:08Spot drain on ip-100-64-42-193 replicas back in 94 s
Nocturne3 queries · read-only
Question
Why did checkout-api p99 jump at 02:08?
- recent_changescheckout
- get_eventscheckout
- promql_rangep99 · 24 h
Probable cause
Deploy 7f3ac1
p99 jumped with the rollout at 02:08 and settled 33.6% higher. The spot drain did not cause it: the 3 evicted replicas were back in 94 s and p99 stayed up. The HPA scaled after the spike.
Suggested fix
kubectl -n checkout rollout undo deploy/checkout-api
Nocturne only reads the cluster. Applying the fix stays with you.
A metrics engine of our own, measured in public.
Against Prometheus 3.1.0 and VictoriaMetrics v1.152.0, on the same machine with the same load of 20 thousand series. The figures came from the Go engine of the time; one script runs the same benchmark on today's Rust engine.
Method and results2.9×
less disk per sample than Prometheus
2.4×
less memory than Prometheus
3.3×
faster than Prometheus on query_range p90
1.7×
slower than VictoriaMetrics on that same query. The losses are on the scoreboard too.
Full scoreboard and the command behind it
| Measure | Orbit Stack | vs Prometheus 3.1.0 | vs VictoriaMetrics v1.152.0 |
|---|---|---|---|
| Disk per sample, sealed blocks · 20k series | 1.87 B | win2.9× less5.38 B | win1.2× less2.24 B |
| RSS with the head full · 20k series | 105 MiB | win2.4× less257 MiB | win2.5× less263 MiB |
| RSS with the head full · 200k series | 511 MiB | win1.7× less856 MiB | loss64% more312 MiB |
| Ingest accepted · 20k series | 555K/s | win1.19×467K/s | parity0.98×567K/s |
| Ingest accepted · 200k series | 582K/s | win1.35×429K/s | win1.12×520K/s |
| query_range p90 · 20k series | 33 ms | win3.3× faster108 ms | loss1.7× slower20 ms |
Disk per sample, sealed blocks · 20k series1.87 B
Prometheus 5.38 Bwin2.9× less
VictoriaMetrics 2.24 Bwin1.2× less
RSS with the head full · 20k series105 MiB
Prometheus 257 MiBwin2.4× less
VictoriaMetrics 263 MiBwin2.5× less
RSS with the head full · 200k series511 MiB
Prometheus 856 MiBwin1.7× less
VictoriaMetrics 312 MiBloss64% more
Ingest accepted · 20k series555K/s
Prometheus 467K/swin1.19×
VictoriaMetrics 567K/sparity0.98×
Ingest accepted · 200k series582K/s
Prometheus 429K/swin1.35×
VictoriaMetrics 520K/swin1.12×
query_range p90 · 20k series33 ms
Prometheus 108 mswin3.3× faster
VictoriaMetrics 20 msloss1.7× slower
Synthetic, scrape-shaped workload on one machine, 2026-09-22: 20,000 series for 4 h, 200,000 for 2 h. Noisier metrics compress worse for every engine.
PROM_IMAGE=prom/prometheus:v3.1.0 \ SERIES=20000 SPAN=4h BLOCK=2h \ services/rust/orbit-stack/bench/run.shThe same workload and Prometheus as the figures above, on the Rust engine. The benchmark page has the method and the results.
One console, from the alert to the span.
From the cluster down to the span, with every workload's cost in the same place and roles that decide who sees what.
Overview
Needs attention5 firing · 2 critical
- PromoEngineCrashLoopCritical · All 3 replicas in CrashLoopBackOff for more than 15mpromo-engine2h 28m
- PaymentsErrorBudgetBurnCritical · Error budget burning at 14.2×, 31% of the 30d budget spent todaypayments-api42m
- KubeletVersionSkewWarning · Node group system runs kubelet v1.30.9 against control plane 1.31aws-node1d 4h
- SearchIndexDiskFillingWarning · PVC at 84%, projected full in 3d 6h at the current write ratesearch-index5h 12m
| Cluster | Status | Nodes | CPU | Cost / mo |
|---|---|---|---|---|
| Production US East | Degraded | 42 | 15% | $18.4k |
| Production EU Central | Healthy | 24 | 8% | $9,860 |
| Staging US West | Healthy | 9 | 13% | $2,140 |
APM/Traces/POST /checkout
Duration517 msStatusOK12 spans · 5 services
Service and operation
checkout-webPOST /checkout
checkout-webHTTP POST checkout-api
checkout-apiPOST /v1/orders
checkout-apicart.load
cart-sessionGET /cart/{id}
checkout-apiorders.create
checkout-apipricing.apply
checkout-apiinventory.reserve
checkout-apiHTTP POST payments-api
payments-apiPOST /v1/charges
payments-apiledger.append
payments-ledgerINSERT ledger_entries
Span attributes · checkout-api
- service.name
- checkout-api
- service.version
- 2.14.3
- k8s.namespace.name
- checkout
- k8s.deployment.name
- checkout-api
- http.request.method
- POST
- http.route
- /v1/orders
Dashboards/Kubernetes overview
Nodes
42
us-east-1 · v1.31
Pods
204
38 namespaces
Monthly cost
$18.4k
prod-use1
Spot coverage
62%
target 80%
Requests per second by cluster
| Series | First | Mean | Max | Last |
|---|---|---|---|---|
| prod-use1 | 20.2K req/s | 18.9K req/s | 22K req/s | 20.1K req/s |
| prod-euc1 | 8.9K req/s | 9.3K req/s | 10.9K req/s | 8.3K req/s |
| stage-usw2 | 1.2K req/s | 1.2K req/s | 1.5K req/s | 950 req/s |
Cost by namespace · per month
Workloads not ready
| Workload | Ready | Status |
|---|---|---|
| checkout/promo-engine | 0/3 | CrashLoopBackOff |
| payments/payments-api | 9/10 | Degraded |
| search/query-router | 5/6 | Progressing |
| observability/node-exporter | 41/42 | Degraded |
| kube-system/aws-node | 41/42 | Degraded |
Costs/prod-use1
- Run rate
- $18.4k
- +4.4% vs last month
- Forecast
- $19.2k
- vs last month
- Spot coverage
- 62%
- target 80%
- Waste
- $645
- 22 over-requested workloads
Cost by namespace
- ingest$5,821
- search$3,960
- payments$2,874
- checkout$2,413
- observability$1,234
- kube-system$866
Where to cut7 open
- search-apiAutoscale$39/mo
- orbit-ingestRightsize$35/mo
- search-indexRightsize$27/mo
- checkout-webScale down$14/mo
It runs in your account. Nothing comes back to us.
One Helm release installs the console, the API, Nocturne and the engine on your Kubernetes, with the agents that feed the engine. There is no backend of ours to receive your data.
- Roles and audit trail
- 12 permissions checked on every route. Every login, change and denied request becomes an audit event: who, what, where and when.
- Sessions
- An HttpOnly cookie. The service token never reaches the browser.
- Network and containers
- Default-deny NetworkPolicies, and non-root containers on a read-only root filesystem with every capability dropped.
- Data residency
- Metrics, logs, traces, dashboards, users and the audit trail stay on volumes in the region you install in.
Roles, permissions and one audit event
| Permission | viewer | editor | admin |
|---|---|---|---|
| estate.readClusters, workloads, pods and events | Granted | Granted | Granted |
| telemetry.queryMetric panels and PromQL | Granted | Granted | Granted |
| apm.readServices, service map and traces | Granted | Granted | Granted |
| logs.readLive pod logs | Granted | Granted | Granted |
| dashboards.readOpen dashboards | Granted | Granted | Granted |
| dashboards.writeCreate, edit and delete dashboards | Not granted | Granted | Granted |
| nocturne.useAsk Nocturne AI | Not granted | Granted | Granted |
| workloads.writeScale workloads and change requests | Not granted | Granted | Granted |
| config.readView workload configuration and reveal secrets | Not granted | Not granted | Granted |
| users.manageUsers, roles and sessions | Not granted | Not granted | Granted |
| audit.readRead the audit trail | Not granted | Not granted | Granted |
| settings.manageConsole settings | Not granted | Not granted | Granted |
Admins define custom roles from the same set of permissions.
{
"ts": 1758470400000,
"actor": { "username": "admin", "role": "admin" },
"action": "user.update",
"target": { "type": "user", "name": "maria" },
"outcome": "success",
"details": {
"before": { "role": "viewer" },
"after": { "role": "editor" }
},
"ip": "198.51.100.17",
"requestId": "4bf92f3577b34da6"
}Questions a platform team asks first
What is SRE Orbit?
SRE Orbit is self-hosted observability for Kubernetes: one console for Prometheus-compatible metrics, stored container logs, OpenTelemetry traces with a service map, custom dashboards and Nocturne AI, an assistant that finds root cause through read-only tools. One Helm release runs web, orbit-api, nocturne and the orbit-stack engine in your own cloud account, with the agents that feed the engine, and roles and an audit trail built in.
Is SRE Orbit an alternative to Datadog?
For Kubernetes metrics, logs, distributed tracing, dashboards, alerts and AI-assisted root cause, yes, and it runs in your account instead of a vendor's cloud. It covers less ground than Datadog: it has no real user monitoring, synthetic tests or security products. Choose it when keeping telemetry in your own account matters more than that breadth.
Does SRE Orbit support OpenTelemetry traces?
Yes, over OTLP/HTTP only. The orbit-stack engine accepts traces as protobuf or JSON, optionally gzipped, with the same bearer token as metrics remote-write, and it has no gRPC receiver: point an OpenTelemetry Collector's otlphttp exporter, or any SDK's OTLP/HTTP exporter, at it. From the spans it derives request, error and latency series under the names Tempo's metrics-generator uses, draws the service map and keeps traces for the retention you configure.
Does data leave my cluster?
Not to us. SRE Orbit has no hosted backend: metrics, logs, traces, dashboards, users and the audit trail stay on volumes in your account. The one outbound call is Nocturne AI's request to the model provider you configure. Point it at a model you host, or keep the default mock provider, and nothing leaves.
How do roles and the audit trail work?
Every user has a role. The built-in viewer, editor and admin roles map to a closed set of permissions such as apm.read, dashboards.write and users.manage, and admins can define custom roles from the same set. orbit-api records each login, logout, user or role change, dashboard change, workload write and permission denial as an audit event with the actor, target, outcome, IP address and request id, and the console filters them by actor, action and outcome.
Does SRE Orbit support SSO?
Yes. People sign in through Google Workspace or one OpenID Connect provider, such as Okta, Microsoft Entra ID, Keycloak or Auth0. An admin configures each provider under Admin > Integrations, and the login page shows its button only while it can sign someone in. Local accounts stay for the emergency admin, and SSO is part of the Free plan.
How is SRE Orbit priced?
Per vCPU of the nodes it monitors, across every cluster, as a 30-day average. Free covers up to 250 vCPU with unlimited clusters and users. Enterprise (in your clusters) and Enterprise Cloud (run by us) add Nocturne AI, the MySQL statement views (SQL text, live sessions, EXPLAIN) and archive rehydration, through a key orbit-api checks offline. Passing a limit never stops ingest or deletes data. The pricing page has the plans and what happens past each limit.
Is SRE Orbit ready for production?
It is early. The components are tested and hardened and the Helm chart is the production path, but SRE Orbit has no production customers yet. We are looking for design partners who run it in their own accounts with the engineers who build it, and the contact form is how you start.
Can I install SRE Orbit myself today?
Not from a public registry yet. The images and the Helm chart come with a design-partner pilot: book a demo, and the engineers who build SRE Orbit install it in your cluster with you. The install guide shows every step, so your team can review it before the pilot.
Which AI model does Nocturne AI use?
The one you configure. Nocturne AI talks to Anthropic, to any OpenAI-compatible endpoint, including a model you host, or to a built-in mock that needs no key. Two tiers sit behind one interface: reasoning for root cause and FinOps plans, fast for short answers. The router picks the tier for each request, so a browser cannot choose to spend the expensive one.
How was the benchmark measured, and what does it not show?
One script pushes one deterministic, scrape-shaped remote-write workload into Prometheus 3.1.0, VictoriaMetrics v1.152.0 and Orbit Stack on the same machine, one after the other: 20,000 series over four hours and 200,000 over two, 2 h blocks. Orbit Stack wins disk (2.9x less per sample than Prometheus, 1.2x less than VictoriaMetrics) and RSS at 20,000 series. It loses query latency to VictoriaMetrics (1.5x slower at p50, 1.7x at p90), uses 64% more RSS than it at 200,000 series, and ties it on ingest at 20,000. The workload is synthetic, and noisier metrics compress worse for every engine. The method and every figure are on the benchmark page.
Which Kubernetes versions and distributions are supported?
Kubernetes 1.30 or later: the chart declares it, because the production values use the pod lifecycle sleep action that went stable in 1.30. The chart uses core APIs only, so any CNCF-certified distribution works; the product is developed against EKS and a local k3d cluster. The Compose path needs no Kubernetes.
See SRE Orbit on your own clusters
Tell us what you run. A person replies within one business day to book a demo of the console, and because SRE Orbit is early, the engineers who build it install it with you.
- Installed with you, next to the tools you run today.
- A copy of your remote-write, so you compare disk, memory and query latency on your own data.
- Direct work with the engineers who build it.