I’m planning a self-hosted Weaviate instance for a client whose data has to stay in Canada. PIPEDA, BC public body rules and Quebec Law 25 all push us toward keeping the vectors, source documents and backups on Canadian servers, so we’re running Weaviate ourselves on a Linux VPS instead of a managed service in a US region. I’d like a sanity check from people who self-host it.
1. Sizing. For about 2 million objects with 768-dimension vectors, is 8 vCPU / 32 GB RAM a sensible starting point?
2. Memory. How do you estimate HNSW memory use, and is vector compression (PQ or BQ) worth turning on from day one?
3. Disk. Local NVMe vs network block storage for the data directory: how much does it matter?
4. Deployment. Docker Compose on a single node, or go straight to Kubernetes with the Helm chart?
5. Vectorizer. Run the embedding model on the same VPS, or keep it on a separate box?
6. Backups. Do you use the backup module to an S3-compatible bucket, or filesystem snapshots?
7. Security. API key auth behind an Nginx reverse proxy with TLS: anything else you’d add?
What would you change in this setup, and what surprised you the first time you self-hosted Weaviate?