Skip to content

// installation

Troubleshooting

Applies to

Product: Server · Audience: Platform Operator

Common failure patterns and their diagnosis – symptom, likely cause, solution. This page consolidates the troubleshooting sections of the installation pages, the FAQ and the Helm reference. Start with the basic commands, then look up your symptom in the tables.

Basic commands

kubectl -n basebox get pods -o wide
kubectl -n basebox get events --sort-by='.lastTimestamp' | tail -30
kubectl -n basebox describe pod <pod>
kubectl -n basebox logs <pod> --tail=200          # or -l app.kubernetes.io/name=<service>
kubectl -n basebox logs <pod> --previous          # after restart
kubectl -n basebox get pvc,cluster,ingress,secret
kubectl top pods -n basebox; kubectl top nodes
kubectl describe node <node> | grep -A8 -E "Allocatable|Allocated resources"
nvidia-smi; nvidia-smi -L

Internal connectivity test from a debug pod:

kubectl run -it --rm debug --image=busybox --restart=Never -n basebox -- sh
# inside the pod:
wget -qO- http://aisrv:8888/health; wget -qO- http://inference:8000/health

Pods and cluster

Symptom Cause Solution
Pods Pending No GPU/MIG resource free, no storage class, control-plane taint kubectl describe pod (events); describe node Allocatable; get pvc; tolerations or taint (Kubernetes)
CrashLoopBackOff Configuration error, missing secret, database unreachable logs --previous; check secrets (get secrets); DB pods Running?
ImagePullBackOff Registry access of the node runtime Test the pull from the runtime, not just from the workstation; pull secret
inference/ragsrv-support: "Temporary error in name resolution" k3s with systemd-resolved Point coredns at an external DNS – FAQ
Database connection fails DB cluster not ready, wrong secret get cluster; logs <db-pod>; secret *-database
GPU Operator pods Pending, "untolerated taint" Single node Tolerations values (NVIDIA / GPU)
helm install "context deadline exceeded" Large images, slow download --timeout 15m/120m; recreate namespace; watch pods

GPU

Symptom Cause Solution
nvidia-smi not found / "couldn't communicate with driver" Driver missing or incomplete sudo apt-get purge nvidia-*, ubuntu-drivers autoinstall, reboot
Node advertises no nvidia.com/gpu GPU Operator not ready, driver conflict get pods -n gpu-operator; logs; node labels
MIG resources missing after reboot MIG layout not persistent, discovery not mixed Recreate MIG; check MIG strategy; restart device plugin
OOM in inference logs Model + KV cache > VRAM Lower context, quantise KV cache, quantise more strongly, TP – Scaling
GPU throttles (temperature) Cooling, power budget nvidia-smi dmon; manufacturer/FAST LTA approval
Data parallelism produces nonsense (GPT-OSS 120B) Known issue Use tensor parallelism

Decision trees: Server Preparation Guide → Troubleshooting.

Login and interface

Symptom Cause Solution
"Organisation not found" instead of the login page VITE_BB_OIDC_FORCE_REALM missing Set to primary
Only a loading spinner Frontend cannot reach backend/IdP Browser console (F12); VITE_BB_GRAPHQL_URL, VITE_BB_OIDC_DOMAIN
/auth/… 404 in the console IdP path wrong VITE_BB_OIDC_DOMAIN = https://<domain>/auth/realms/ with trailing slash; check .well-known/openid-configuration
Redirect loop after login Auth fails at an unexpected point; global.domain ≠ external hostname Clear browser data; console (INVALID_TOKEN); aisrv logs
aisrv: "iss does not match" AISRV_OIDC_IDP_URL not a base URL or ≠ VITE_BB_OIDC_DOMAIN Align; set AISRV_OIDC_ISSUER_URL only with care – a foreign issuer would admit foreign tokens
404 on the hostname DNS ≠ ingress host Compare DNS and kubectl get ingress
Certificate not issued ClusterIssuer, ACME reachability get certificate,certificaterequest,order,challenge
GraphQL returns HTML Ingress rules/host Check ingress paths
Support button missing/wrong VITE_BB_SUPPORT_BUTTON_LINK Set URL or leave empty to hide

Chat, inference, RAG

Symptom Cause Solution
"data transmission failed (SSE)" Connection browser↔server or AISRV↔inference/ragsrv; proxy buffers aisrv logs; disable proxy buffering, timeouts; inference health
Chat does not answer, error visible Inference offline (no fallback) Check inference pod/endpoint; wait for warm-up after start
"Model not found" AISRV_LLM_MODEL ≠ /v1/models Copy the identifier exactly
401 from the inference endpoint Keys differ AISRV_LLM_API_KEY and VLLM_API_KEY from the same secret
Long requests abort AISRV_LLM_CONTEXT_SIZE > runtime context Align the values
Train of thought in the answer text Reasoning parser missing Configure the parser in vLLM
Uploads stay "processing" ragsrv ↔ ragsrv-support (API_KEY), webhook Logs of both; WEBHOOK_STATE_URL = http://aisrv:8888/rag/v1/state
OCR/STT very slow CPU mode or shared GPU Dedicated GPU/MIG
Connector "Test connection" fails Base URL, egress, TLS CA Connector pod logs; firewall to the target system; import private CA

Storage and database

Symptom Cause Solution
PVC Pending Storage class missing/not default get storageclass; set in values
Uploads/downloads fail Volume full Enlarge PVC; clear caches
CNPG cluster not healthy Storage, replication describe cluster; DB logs; check backup
Migration fails on start Checksum, incompatible state Logs (grep migrat); restore from backup – Updates

If nothing helps

  1. Note symptom, time, recent changes.
  2. Collect aisrv logs and logs of the affected component, get events, browser console, manifest – without credentials.
  3. support@basebox.ai; for suspected security issues datenschutz@basebox.ai.

Further collections: FAQ · Helm chart overview → Troubleshooting · Using Helm charts → Troubleshooting