Products / For platform admins
TAIP Admin
AvailableThe whole AI cluster on one screen. kubectl optional.
TAIP Admin is a web-based administration console purpose-built for AI infrastructure. One Go binary serves the API and the SPA. It auto-detects metrics-server, Kueue, KServe, Training Operator, cert-manager, Gateway API, DRA, and VPA — each integration lights up when its API appears and disappears cleanly when it doesn't — and connects to Prometheus, Alertmanager, and Grafana with one URL each, degrading gracefully when they're absent. Resource accounting works in three tiers: requests and capacity from the K8s API alone, live usage with metrics-server, 30-day history with Prometheus. Storage is a console of its own — Ceph clusters and Longhorn discovered as backends, every PV joined back to the backend serving it, capacity projected forward, and claims nothing mounts flagged as reclaimable. The platform's own applications get a health view with restart, scale, and rollback. It is engineered, not just assembled: listing 617 secrets takes 0.5 seconds instead of 62 via metadata-only reads — and secret values never reach the browser.
Specification
- Version
- v1.6.21 — generally available
- Footprint
- Single Go binary · one Helm release · amd64 + arm64
- Auto-detected
- metrics-server · Kueue · KServe · Training Operator · cert-manager · Gateway API · DRA · VPA
- Connects to
- Prometheus · Alertmanager · Grafana — one URL each, optional
- GPU telemetry
- Per-accelerator — NVIDIA DCGM · Ascend NPU: utilization, memory, temp, power
- Storage
- Ceph (several clusters) and Longhorn as backends · PV → backend join · capacity projection · unused-claim detection · per-tenant GiB-hours
- Applications
- Health across platform apps · restart, scale, rollback · alerts attributed to the app that owns the pod
- Identity
- Users, groups, invitations, sessions, service accounts · who can actually launch which app
- Roles
- OIDC · admin / viewer split · secrets never sent to browser
- Languages
- English · 简体中文
Proof, not promises
See it in one block.
No proprietary SDKs, no rewrites — TAIP Admin meets your tools where they already are.
$ helm install taip-admin taip/taip-admin
detected metrics-server ✓ kueue ✓ kserve ✓ cert-manager ✓
detected gateway-api ✓ dra ✓ vpa ✓ training-operator ✓
connected prometheus ✓ alertmanager ✓ grafana ✓ # one URL each
# remove a component and its pages vanish cleanly
# 617 secrets listed in 0.5s (metadata-only) · values never leave the cluster▌One binary, one Helm release. The console grows and shrinks with your stack — across Kueue versions, without rebuilds.
Capabilities
What TAIP Admin gives you
GPU and AI workloads, first-class
Extended resources, DRA device browsing, and per-accelerator telemetry for NVIDIA DCGM and Ascend NPU — utilization, memory, temperature, power. A cluster GPU heatmap shows idle-vs-active capacity with owner attribution, alongside topology and MIG. KServe InferenceServices and ServingRuntimes, plus Kueue queue management via API discovery — one binary across versions.
Storage you can actually account for
Ceph clusters and Longhorn are discovered as storage backends — declared by a StorageBackend CRD that supplies fleet identity, console link, and metrics job — and every PV is joined back to the backend serving it. Health, capacity, OSD state, pools, and filesystems come from Prometheus; capacity is projected forward rather than reported as a single number; claims that nothing mounts are flagged so a reclaim is a decision, not an archaeology project. Expand a PVC, change a reclaim policy, set the default class from the UI — and read per-tenant storage as GiB-hours, chargeable like GPU time.
The platform's own apps, watched
A health view over the TAIP applications themselves — traffic signal per app, alerts attributed to the pod that raised them, and restart, scale, or rollback without leaving the page. The War Room shows them as tiles beside the node grid, so an on-call responder sees the cluster and the products on it at once.
Three-tier resource accounting
Requests, limits, and capacity from the K8s API alone; live CPU and memory when metrics-server is present; 1h–30d history when Prometheus is configured. The same UI scales with your stack — and stays fast: 617 secrets in 0.5s, not 62s.
Alerts, silences, and Grafana deep-links
Severity-coded alert tables with one-click silence creation pre-filled from the alert's matchers. Live alert badge in the sidebar. Open-in-Grafana buttons that carry cluster, node, namespace, and pod context with them. Configure outbound receivers — email/SMTP or a CloudSentry webhook — from the console.
War Room for incidents
A full-screen NOC dashboard with auto-refresh, live event SSE feed, a node grid carrying GPU state with hover detail, application tiles, and scheduling and resource panels — built for wall displays and on-call shifts. Node cordon and drain with real-time eviction progress.
Audit trail, idle reclaim, and an app catalog
A structured audit log of every mutating admin action, with optional persistent history you can query in-app. Idle-resource detection flags idle GPUs, inactive notebooks, and stale jobs to reclaim. An OCI/Helm app catalog browses charts from your registry and installs them through a guided wizard — and a bilingual user guide ships inside the binary.
Identity, topology, and queue analytics
Manage platform users and groups, invitations, live sessions, MFA devices, and service accounts from the console — and see, per user, which TAIP applications they can actually launch, honouring the identity provider's own application gate rather than guessing. Visualize cluster hierarchy and intra-node NUMA topology. Read Kueue queue analytics — wait time, depth, fairness, and preemptions — and curate the portal app catalog and broadcast announcements.
How it works
Install it, then operate the cluster.
- Step 01
Point it at a cluster
One Go binary, one Helm release, one optional CRD. OIDC for SSO. The whole console is a single process.
- Step 02
Integrations light up automatically
Kueue, KServe, DRA, VPA, cert-manager, Gateway API — auto-detected from the cluster. Prometheus, Alertmanager, and Grafana connect with one URL each.
- Step 03
Operate and respond
Severity-coded alerts, one-click silences, War Room dashboard, live event SSE, node cordon and drain — without a kubectl tab open.
Who it's for
Built for these teams
- Platform engineers running shared AI clusters
- On-call responders investigating incidents
- Auditors and read-only viewers (admin/viewer roles built in)