What happened
The Kubernetes project has published benchmarks of node swap backed by local NVMe SSDs. Swap support on nodes reached General Availability in v1.34. The authors tested it on CI kernel builds, headless Chrome sandboxes and isolated Python sandboxes, and report density gains of up to 3×, often with little or no latency cost.
Why it matters
Memory is usually the first limit a cluster hits. Many workloads touch a lot of RAM at start-up, then sit idle while that memory stays resident. You pay for it, and it caps how many pods fit on a node.
The post gives some concrete examples. A Linux kernel build ran with a 300 MB limit instead of 600 MB, and finished in 374s against 433s for the baseline. gVisor-sandboxed Python pods went from 80 to 240 per node. Kata-based Chrome pods went from 40 to 50, a smaller gain.
There is a clear warning too. Squeezing the kernel build down to 200 MB pushed active memory into swap and execution time rose by over 40%. Swap is insurance for bursts, not a substitute for the RAM your workload actively uses.
For a small team, this could mean fewer nodes for CI runners or sandboxed jobs. It could also mean unpredictable slowdowns if you size limits too aggressively. These numbers come from one set of tests, so your results will differ.
What to do this week
- Check your Kubernetes version. Node swap support is GA from v1.34, so older clusters need an upgrade plan first.
- Confirm your nodes use cgroup v2, which the swap support relies on for separate swap accounting.
- Find your most memory-bound workloads, such as CI runners, and compare requested memory with what they actually use while idle.
- If you test swap, use a separate node pool with fast local NVMe disks. Avoid slow disks.
- Load-test before changing limits. Watch latency and I/O wait, not only pod count, and keep a rollback path.
Need a hand?
Deciding whether swap suits your workloads, and how to size limits safely, is the kind of question we help with in our Kubernetes consulting. We can help plan a small, reversible trial on one node pool. The first call is free and is with an engineer.
Source: Scaling Kubernetes Workloads with Node Swap
Related service: Kubernetes Consulting