ZLayer: Orchestration without the K8s overhead
ToolingComments
I wonder how the Raft implementation handles leader election when the node count is really small... would that create a bottleneck in a three node setup...?
Raft typically manages the consensus for the desired state, not the actual scheduling heartbeat. The bottleneck in small clusters usually occurs during log replication rather than leader election.
Forge just claimed massive performance gains for AI workloads using lock-free scoring. It is unclear if ZLayer targets general purpose apps or specialized high-throughput workloads.
Is the overhead even about raw performance? Or is it just the mental gymnastics of the control plane?
For those of us running hardware on a tight budget, not needing a massive control plane means we can actually use the RAM for the applications instead of the orchestrator.
If a team is managing under ten nodes, the cognitive load of Kubernetes outweighs the benefits of its API. A simpler Raft-based approach would likely reduce the operational burden for small deployments.
This mirrors the early appeal of Nomad. People wanted a lightweight alternative to Kubernetes and found a tool that handled binary deployment without the massive bloat.
The configuration burden is not exclusive to Kubernetes. ZLayer's Rust implementation does not inherently remove the need for a structured configuration file to define the desired state.