README
¶
Armada is a system for scheduling and running batch jobs (e.g., a compute job for training a machine learning model) over Kubernetes clusters. Armada is designed for high availability and to handle scheduling hundreds of jobs per second (with potentially millions of jobs queued) over thousands of nodes.
To achieve this, Armada, unlike previous Kubernetes batch schedulers (e.g., kube-batch), can schedule jobs over multiple Kubernetes clusters simultaneously and hence scale beyond the limitations of a single Kubernetes cluster (which we find to be about 1000 nodes). In addition, Kubernetes clusters can be connected and disconnected from Armada on the fly without disruption. Jobs are submitted to job queues, of which there may be many (for example, each user could have a separate queue), and Armada divides compute resources fairly between queues.
Armada is loosely based on the HTCondor batch scheduler and can be used as a replacement of HTCondor, provided all nodes are enrolled in Kubernetes clusters.
Armada is not
- A service scheduler (i.e., Armada jobs should have a finite lifetime)
- Designed for low-latency scheduling (expect on the order of 10 seconds from job submission)
- Designed to schedule jobs over underlying systems other than Kubernetes clusters
Documentation
For an overview of the architecture and design of Armada, and instructions for submitting jobs, see:
For instructions of how to setup and develop Armada, see:
For API reference, see:
We expect readers of the documentation to have a basic understanding of Docker and Kubernetes; see, e.g., the following links:
Directories
¶
| Path | Synopsis |
|---|---|
|
cmd
|
|
|
internal
|
|
|
pkg
|
|
|
api
Package api is a reverse proxy.
|
Package api is a reverse proxy. |
|
api/binoculars
Package binoculars is a reverse proxy.
|
Package binoculars is a reverse proxy. |
|
api/lookout
Package lookout is a reverse proxy.
|
Package lookout is a reverse proxy. |