W&B Self-Hosted
AdvancedMl ProjectSelf-hosted Weights & Biases on your own Kubernetes cluster, with MySQL and your own object storage.
Published 27 September 2026
About W&B Self-Hosted
W&B Self-Hosted runs the Weights & Biases platform on infrastructure you control, through the W&B Kubernetes Operator: the official Helm chart brings up the full server, with MySQL as the customer-managed metadata database, Redis behind it, and object storage of your own for experiment artifacts. This is the same platform as W&B Cloud, on your cluster.
The product is the tracking and collaboration layer for ML work: experiment dashboards, run comparison, dataset and artifact versioning, sweeps, evaluation, and W&B Weave for tracing LLM applications. Training code instruments the same wandb SDK against your server that it would against the cloud, so nothing changes in the code when everything changes in where the data lives.
It's a licensed, heavy platform. Company use of self-managed W&B runs on an Enterprise license, with a trial license to start, and you're responsible for provisioning, maintenance, and upgrades. The reference architecture asks for a multi-node x86 Kubernetes cluster with 16 GB of memory per node for a test setup and 64 GB for production, plus a MySQL instance sized to match. Traffic enters through the cluster's ingress controller, and W&B requires a properly signed TLS certificate; self-signed ones aren't supported. MLflow Self-Hosted is the light single-server alternative.
The reward is total data ownership. Experiment runs, datasets, model artifacts, and LLM traces never leave your infrastructure, and SSO, audit, and residency requirements are met by construction rather than by a vendor's regional deployment.
Key Features
- ✓W&B Kubernetes Operator deployment through the official Helm chart
- ✓MySQL as the customer-managed metadata database
- ✓Redis and your own object storage behind the server
- ✓The full W&B feature set: dashboards, artifacts, sweeps, evaluation, and Weave tracing
- ✓The same wandb SDK instrumentation as W&B Cloud, pointed at your server
- ✓SSO, audit logs, and data residency on your own infrastructure
When to Use W&B Self-Hosted
- →ML teams in regulated industries that can't send experiment data to a vendor cloud
- →Organizations training on sensitive datasets that must stay on-premises or air-gapped
- →Enterprises standardizing experiment tracking and LLM tracing across many teams
- →Companies with an existing Kubernetes platform team to operate the deployment
- →Teams moving from W&B Cloud to their own infrastructure for data-residency reasons
Pros
- The full platform, experiment tracking through LLM tracing, at fixed infrastructure cost
- Every run, dataset, and model artifact stays on infrastructure you own
- The same SDK and UI as W&B Cloud, so teams relearn nothing
Cons
- The heaviest ML stack in this family: a multi-node Kubernetes cluster and MySQL to operate
- Company use needs an Enterprise license; the free self-hosted server is personal-only
- MySQL is required, not swappable: the platform does not run on PostgreSQL
Hosting Options for W&B Self-Hosted
Deploy W&B Self-Hosted on Hetzner
The budget path for a self-managed cluster: 64 GB x86 dedicated servers, the production node size, start around €50/mo, and 16 GB cloud servers for a test cluster around €20/mo each. Hetzner has no managed Kubernetes, so you run the cluster, ingress, certificates, and a MySQL node yourself; ARM servers are out, since W&B supports x86 only.
Deploy W&B Self-Hosted on Amazon Web Services
A documented target: the reference architecture uses EKS (about $73/mo per cluster before nodes) with RDS Aurora for MySQL and S3 for object storage, and W&B publishes Terraform modules for the whole stack. Three 16 GB worker nodes put a test setup in the hundreds of dollars a month.
Deploy W&B Self-Hosted on Google Cloud Platform
The same documented path on GKE, with Cloud SQL for MySQL and GCS behind it, covered by W&B's Terraform modules. GKE waives the management fee for one zonal cluster, so a test setup pays for its nodes, database, and storage.
Deploy W&B Self-Hosted on Microsoft Azure
AKS with Azure Database for MySQL and Blob Storage, the third documented cloud path for the operator deployment. AKS's free tier has no control-plane fee, and cluster access can use the organization's Microsoft Entra ID identities.
These are highlighted picks. To see all the tools, check the Hosting & Cloud category.
Frequently Asked Questions about W&B Self-Hosted
Why self-host W&B instead of using W&B Cloud?
Only for data control. W&B Cloud is free for personal use and Pro starts at $60/mo, managed and updated continuously. Self-managed W&B exists for strict on-premises, air-gapped, or regulatory requirements: company use runs on an Enterprise license, and you run the whole platform, from the Kubernetes cluster to MySQL backups. The SDK and UI are the same either way; the difference is where the data lives and who carries the operations.
How much infrastructure does W&B need?
The reference architecture sizes per node: 2 cores, 16 GB of memory, and 100 GB of disk for a test or dev setup, and 8 cores with 64 GB for production, on a multi-node x86 Kubernetes cluster. MySQL 8.4 gets its own node of the same class (500 GB of disk in production), Redis 7 runs beside it, and object storage holds the artifacts. It's a genuine platform, not a single-container deployment.
Is there a free way to self-host W&B?
Only for personal use. W&B offers a free self-hosted server that starts in a single Docker container, but its terms don't allow corporate use. A company deployment, the Kubernetes platform this stack builds, runs on an Enterprise license; W&B provides a free trial license through its deployment portal to evaluate it first. W&B Server and W&B Self-Managed are two names for the same product.
What needs backing up?
Two things carry state: the MySQL database, which holds projects, runs, and metadata, and the object-storage bucket, which holds the artifacts. The cluster itself is disposable: redeploy it from the operator manifests or Terraform modules, point it at the same database and bucket, and the platform is back.
How is W&B exposed to users and training jobs?
Through the cluster's ingress controller, which routes the app, API, and console paths to their services; NGINX, Traefik, Istio, or a cloud load balancer's controller all work. W&B requires a valid signed TLS certificate and doesn't support self-signed certificates or custom CAs, so use a public certificate, for example from Let's Encrypt through cert-manager. Training jobs then log to that HTTPS hostname exactly as they would to W&B Cloud.
Stacks Related to W&B Self-Hosted
Airflow Self-Hosted
Infrastructure
Self-hosted Apache Airflow: the standard data-pipeline scheduler on your infrastructure.
GitLab Self-Hosted
Infrastructure
Self-hosted GitLab with Docker: your code, your CI, your infrastructure.
Sentry Self-Hosted
Infrastructure
Self-hosted Sentry error tracking on your own infrastructure, with PostgreSQL, Redis, ClickHouse, and Kafka under the hood.
Ghost Self-Hosted
Infrastructure
Self-hosted Ghost: publishing, memberships, and newsletters with your content and members on your own server.
Scores
Popularity3/5
Well known across ML engineering, but the self-managed deployment is an enterprise path and rare below that scale.
Learning Curve5/5
A multi-node Kubernetes cluster, a MySQL database, and an operator to run: the steepest stack in the ML family here.
Flexibility3/5
The platform is opinionated: MySQL is required, and the cluster is the deployment frame. Inside that frame, the full W&B feature set runs.
Performance4/5
Built and sized for team and organization scale, with autoscaling recommendations in the reference architecture.
Portability3/5
The operator and Terraform modules redeploy the platform anywhere, and MySQL dumps carry the metadata. The artifacts move through your own object storage either way.
Tools in the W&B Self-Hosted Stack
DevOps & CI/CD
Data & ML Libraries
W&B Self-Hosted Pricing
Company use of self-managed W&B runs on an Enterprise license at custom pricing, with a free trial license to evaluate it; the only free self-hosted option is a personal server that doesn't allow corporate use. On top sits the infrastructure: a multi-node Kubernetes cluster at 16 GB of memory per node for a test setup and 64 GB per node for production, plus MySQL and object storage. W&B Cloud is free for personal use and Pro starts at $60/mo.
16 GB of memory per worker node for a test setup and 64 GB per node in production, per the official reference architecture; x86 only.
Self-managed on the cluster or a managed cloud database; sized to match the cluster tier.
Your own S3, GCS, or Blob bucket for experiment artifacts.
Company self-managed deployments run on an Enterprise license, with a free trial license to start; the free self-hosted server is for personal use only.
W&B Self-Hosted System Requirements
source- Worker nodes
- 2 cores / 16 GB / 100 GB per node (test or dev); 8 cores / 64 GB / 100 GB in production
- MySQL node
- 2 cores / 16 GB / 100 GB (test or dev); 8 cores / 64 GB / 500 GB in production
- Database
- MySQL 8.4.x (PostgreSQL is not supported)
- Cache
- Redis v7.x
- Architecture
- x86 only (ARM unsupported)
- Platform
- Kubernetes (v1.34 or newer) with Helm v3 and an ingress controller
Numbers are per node from W&B's reference architecture (Models-only sizing); this is a multi-node Kubernetes platform rather than a single server, and company deployments run on an Enterprise license.