NEW YORK — What happened
The 2026 State of Observability Report, compiled from audits across 420 enterprise Kubernetes environments, examined the operational reality of running continuous, eBPF-based profiling across fleet infrastructure.
Why it matters
Extended Berkeley Packet Filter (eBPF) technology allows engineers to run safe programs directly inside the operating system kernel. It revolutionized observability by eliminating the need to modify application source code or install heavy runtime agents. However, capturing granular call stacks from every CPU core across thousands of microservices has created a massive secondary problem: the cost of storing, compressing, and querying petabytes of profiling data.
Deep dive
The report confirmed that host-level CPU utilization for observability agents dropped from an average of 6.5% under traditional bytecode-instrumented agents to under 0.8% using eBPF probes. Yet, data egress and cloud storage costs associated with continuous telemetry rose by an average of 115% for teams collecting unfiltered stack traces. Organizations that did not implement dynamic sampling or edge aggregation at the Kubernetes node level found that their monthly observability invoices quickly rivaled their core compute spending.
Report check
Vendor marketing has historically framed continuous eBPF monitoring as virtually free and effortless. While the low-impact compute footprint inside the host kernel is rigorously verified by the audit, the marketing claim that data collection incurs negligible network and storage costs is demonstrably false. The report documents multiple instances of teams spending up to 28% of their cloud infrastructure budget strictly on managing observability data streams.
Open questions
Will standardized on-node telemetry filtering frameworks mature quickly enough to allow engineers to discard redundant call traces before they leave the host network interface?