DevZero vs Cluster Autoscaler

Builds a usage profile per workload from collected CPU, memory and runtime data and uses it to drive rightsizing recommendations.

The node agent detects JVM, Node.js, .NET, Go, Rust, GraalVM native image, Python, Ruby, Deno and Bun in every container with no workload changes. Low-privilege installs turn it off.

Collects pod-to-pod network flows with netfilter by default. An eBPF collector mode is available as an opt-in setting.

Tracks OOM kills, crash loops, CPU throttling and container startup lifecycle events for each workload.

Analyzes each node group and the workloads on it, then suggests node policy changes to requirements, capacity type, zones, limits and disruption settings.

Previews underused nodes whose pods fit elsewhere in their node group and drains them on demand or by policy. Also flags duplicate and overlapping node policies.

Splits node prices into CPU, memory and GPU unit costs and attributes them to each workload.

Prices network traffic by type, including cross-AZ, cross-region, internet egress and NAT paths, estimated at cloud list rates.

Calculates persistent volume claim storage cost for each workload.

Reads AWS, GCP and Azure billing exports to show commitment coverage and utilization and to price nodes at commitment-adjusted cost. It does not buy or manage commitments.

Sizes CPU and memory requests per container and adjusts memory limits with them. Applies changes in place on Kubernetes 1.33+, with a restart-based fallback when in place is unavailable.

Detects existing autoscalers automatically. An actuating VPA stands DevZero down. With an HPA it skips replica scaling, keeps vertical sizing, and offers an optional clamp against triggering scale-out.

Installs Karpenter with spot-to-spot consolidation turned on for AWS, Azure, GCP and OCI, with a per-cluster toggle in Karpenter settings.

Mirrors existing Karpenter NodePools as read-only cluster-managed policies, and a migration wizard moves them under DevZero management.

Checkpoints a running container with CRIU and restores it on the replacement pod with its in-memory state. Opt-in per policy, on labelled nodes in the same cluster.

Live migration checks workload kind, image coherence and node readiness first, and skips the workload without changes if a check fails. Live migration is opt-in per policy and part of Pro.

Alerts on application restarts, out-of-memory kills and CPU throttling, delivered by email or by webhook to Slack, Microsoft Teams, Discord or your own endpoint.

Records inbound request rate, errors and latency per container from socket probes in the network monitor. Off by default; enable it with the containerRedMetrics Helm value.

Horizontal scaling runs through KEDA ScaledObjects rendered from the workload rule. KEDA is an opt-in install, off by default, and DevZero creates no native HPA.

Raises the CPU request of detected Java containers to a startup floor, 75m by default, and can tune JVM heap flags when the heap rule is on. No other runtime has its own rule.

The sizing engine adapts percentile targets and headroom to each workload's observed CPU and memory behavior. No rule triggers scaling from a stored profile.

Raises memory right after an OOM kill, on by default, with configurable multipliers and cooldowns. A CPU throttling response exists in the operator but ships off.

Node policies select instances by category, family, size and capacity type, so one policy spans many shapes.

Savings estimates flag the share of billed node hours under a reservation or savings plan. It does not buy commitments or steer nodes onto them.

Supports fixed-size NodePools through spec.replicas on AWS, Azure and GCP once you turn on the Static capacity setting. OCI is not supported.

Node policy suggestions read workload profiles, such as spot suitability and startup time, to propose spot capacity and grace periods. You apply them; no rule scales nodes on profiles.