Embedded SQLite etcd data corruption risk in k3s-server for local integration testing

Answer to the main question

For a short‑lived, single‑node k3s cluster that runs locally without cloud credentials, the default SQLite‑backed etcd remains the simplest and most convenient choice. It requires no external service, keeps the bootstrap script minimal, and is sufficient for workloads that generate <1 GB of etcd data and perform <10 writes per second. If your integration test involves more nodes, higher write rates, or a data size approaching the 10 GB limit, switch to an external HA etcd cluster to avoid corruption and downtime.

Operational signals that SQLite‑etcd is nearing its limits

• Frequent etcd: write failed or etcd_server_error entries in /var/log/k3s.log.
• High API‑server latency or timeouts when querying resources.
• Disk usage of /var/lib/rancher/k3s/server/db approaching 8–10 GB.
• Repeated disk full or i/o timeout messages in etcd logs.

Practical verification steps

# Check etcd health
kubectl get etcd --raw /health

# Inspect etcd data size
du -sh /var/lib/rancher/k3s/server/db

# Look for write errors in k3s logs
grep -i "write failed" /var/log/k3s.log

Node‑token hygiene for short‑lived test clusters

• Generate a unique --token when starting the server: k3s server --token $(openssl rand -hex 16).
• Do not reuse the default admin token in shared or reused workspaces.
• After each test run, delete the token secret: kubectl delete secret node-token-$(kubectl get secret -o jsonpath='{.items[0].metadata.name}') (or simply wipe the data directory).
• If you need to keep the cluster alive between runs, rotate the token with k3s server --token <new> and update any join commands.

When to consider moving to external etcd

If you observe any of the operational signals above, or if your test suite creates many nodes or performs heavy write workloads, it is safer to provision a small external etcd cluster (3‑node HA) and point k3s to it with --datastore-endpoint and --datastore-cafile parameters. This eliminates the SQLite size and concurrency limits and gives you a more realistic HA environment for integration testing.

Missing diagnostic detail

To fine‑tune the recommendation, could you tell me the current size of the /var/lib/rancher/k3s/server/db directory?