Hammrly — Open-source Kubernetes control plane

The story

The problem Hammrly is built for

Shared Kubernetes clusters need more than raw Job YAML. Tenants need authenticated self-service, fair access to scarce GPUs, and one place to run interactive sessions and batch work—without turning every scientist into a cluster operator.

Read the Hammrly story →

Architecture

Submit once. Orchestrate everywhere.

Clients hit a JWT-secured gateway. Workloads travel over a fast in-memory database to an orchestrator that persists lifecycle state and optionally submits Kubernetes Jobs through Kueue.

Client UI Portal & API clients

Gateway Auth · validate · accept

In-memory database Quick, durable handoff

Orchestrator Persist · submit · watch

Kubernetes Jobs via Kueue

Write path

Gateway never writes Postgres. The orchestrator is the sole automated writer of submission lifecycle rows.

Read path

Query API serves job and session state from PostgreSQL so the UI can list, poll, and open ready sessions.

Screenshots

See the control plane in action

Launch interactive notebooks and desktops, track fair-shared capacity, and govern multi-tenant workloads from one place.

Interactive sessions

New session

NameKindResourcesStatus

analysis-notebook notebook 2 CPU · 4Gi ready

astro-desktop desktop 4 CPU · 8Gi · GPU running

campaign-batch-12 headless 8 CPU · 16Gi queued

carta-viewer carta 2 CPU · 8Gi ready

Web UI — browse tenant sessions and open ready workloads.

1
2
3
4

from hammrly import cluster
jobs = cluster.list_jobs(tenant="demo")
print(f"{len(jobs)} active")
jobs[0].status

12 active
'ready'

Interactive Jupyter sessions on governed cluster capacity.

astronomy-desktop — GPU session

Remote desktop workspace

GPU-aware desktop sessions for interactive analysis.

Get started

Build on Hammrly

Clone the repo, bring up the Compose stack, and submit your first session in minutes.

Get Started on GitHub