Layer 4 · self-hosting reality check
What it actually takes to self-host MLflow
The docs say 1 GB. In practice you want 4 GB server. Here is the honest version — real requirements, real monthly cost, what you will be maintaining, and the one thing that catches people out.
Usually reached from Weights & Biases alternatives, where MLflow is one of the picks.
| RAM — documented minimum | 1 GB |
|---|---|
| RAM — what it really needs | 4 GB server |
| CPU | 2 vCPU |
| Disk | Artifacts dominate — model checkpoints are large and nobody deletes them |
| Monthly cost | $24/mo server plus object storage, against Weights & Biases at roughly $50 per user per month |
| Setup time | 2 hours for a team-grade install |
| How you install it | `mlflow server` with PostgreSQL as backend store and S3-compatible object storage for artifacts |
| Ongoing maintenance | Moderate. Artifact storage growth is the recurring cost. |
| Where it stops scaling | Large teams and millions of runs with a proper backend. The file store is the only real ceiling. |
The thing that catches people out
The default file-based store has no concurrency control — two runs writing simultaneously can corrupt it, and the UI crawls past a few thousand runs. It is fine for one person and wrong for a team. Deploy with PostgreSQL and object storage on day one; retrofitting after you have history is far more annoying than doing it first.
When not to self-host MLflow
Your team lives in W&B Reports. Turning a training run into something a non-ML stakeholder reads is W&B's genuine advantage and MLflow has no equivalent.
Every guide here carries this section. A site that only ever tells you to self-host is selling something — the useful answer is sometimes no.
Other Layer 4 self-hosting guides
- Self-hosting Ollama8 GB VRAM for a 7B model at usable speed; 24 GB for 30B-class
- Self-hosting vLLM24 GB VRAM minimum for useful production serving
- Self-hosting QdrantVectors × dimensions × 4 bytes, in RAM, plus overhead — 1M × 768d is roughly 3 GB
- Self-hosting pgvector8 GB — the HNSW index wants to be resident
- Self-hosting LlamaIndex4 GB for the app; your vector store is the real cost
- Self-hosting faster-whisper5 GB VRAM for large-v3 in float16; 2 GB with int8
Common questions
- How much RAM does MLflow actually need?
- 4 GB server in practice. The documented minimum is 1 GB, which is the figure at which the process starts rather than the figure at which it works under real use. 2 vCPU alongside it.
- What does self-hosting MLflow cost per month?
- $24/mo server plus object storage, against Weights & Biases at roughly $50 per user per month This is commodity VPS pricing and excludes your time, which is the larger cost for most people — budget for moderate. Artifact storage growth is the recurring cost.
- How long does it take to set up MLflow?
- 2 hours for a team-grade install, via `mlflow server` with PostgreSQL as backend store and S3-compatible object storage for artifacts.
- When should I NOT self-host MLflow?
- Your team lives in W&B Reports. Turning a training run into something a non-ML stakeholder reads is W&B's genuine advantage and MLflow has no equivalent.
- What is the most common mistake when self-hosting MLflow?
- The default file-based store has no concurrency control — two runs writing simultaneously can corrupt it, and the UI crawls past a few thousand runs. It is fine for one person and wrong for a team. Deploy with PostgreSQL and object storage on day one; retrofitting after you have history is far more annoying than doing it first.