Skip to main content
Expanse works with whatever schedules your jobs today. You don’t change schedulers, and your researchers don’t change how they submit.

​
Schedulers

Talk to us about your setup and we’ll tell you straight how Expanse would connect to it.

​
If you already collect telemetry

Expanse can also start from the telemetry you already collect, read from Kafka, Prometheus and the like, with nothing running on the cluster itself. How far this takes you depends on what your pipeline records. Expanse uses the job events and metrics it carries. Monitoring pipelines rarely keep source code or the processes and commands inside a job, and they often sample GPUs less often than the collectors do, so predictions and diagnoses have less to work with. Two things are needed: every record tied to the job it came from, and a field mapping we agree with you. During the engagement we work through, with you, what your pipeline supports, and where collectors would add the most.

​
Hardware

NVIDIA GPUs, including MIG partitions, and CPU-only nodes. Collectors run on x86 and Arm. If you run other accelerators, such as AMD GPUs, talk to us.

​
Where Expanse itself runs

Your clusters can be in your own data centre, in a public cloud, or both. Expanse’s own services run in your network, on infrastructure you provide: your own servers or your cloud account. Kubernetes is the simplest route, and we ship Helm charts for it. If Kubernetes isn’t an option, the same services run on virtual machines and we agree the layout with you. If your network has no outbound path at all, Expanse runs fully air-gapped, with an offline licence and nothing leaving the network. Security covers what each option means.