Under active development. Breaking changes expected. APIs, installers, and UI may shift between releases.
Master and workers
Today the Workstation runs as a single-node cluster: it is its own master. The scheduler, the worker seam, and the job-offload contract are in place. Multi-machine workers are being built.
TOPOLOGY
A star around the master
One Workstation is always the master. Compute-node workers register with it and offer their GPUs. The master schedules each job onto the worker best suited to run it.
JOB OFFLOAD
One contract for heavy work
Reconstruction and vision inference are too heavy for a drone companion and often for a single machine. The offload contract lets any plugin push that work to the cluster and read the result back.
SUBMIT
Enqueue a job
A plugin or the operator submits a reconstruction or vision job to the cluster. The master accepts it, records it, and places it on the queue.
ATTACH
Write the dataset
The capture data for the job is written to the cluster. Workers read it directly, so a job carries everything it needs to run without a round trip.
READ
Poll and fetch
The submitter reads job status and pulls the result when it is ready. The world model or the detections come back through the same contract.
SCHEDULING
How the master places work
The cluster is capability-aware. Each node reports its GPU and memory, and the master weighs jobs onto the machines that can finish them fastest.
- Master
- Always exactly one
- Failover
- Any capable node can take over
- Workers
- Register and offer GPUs
- Scheduling
- Capability-weighted (GPU, memory)
- Offload
- Reconstruction, vision inference
- Callers
- Any plugin, the operator
- Today
- Single-node, master-only
EXPLORE
Related
One console, many GPUs
The cluster lets a Workstation borrow the compute of every machine on the bench. Read about the worker profile.
See compute nodes