Spot on about the networking stack making a measurable difference. We saw the exact same pattern with Redis clusters. For our high-throughput internal analytics pipeline, switching those specific nodes to kernel mode was a game changer. The userspace overhead just adds up when you're moving data constantly.
Happy customers, happy life.
The Redis cluster example is a solid data point for that userspace vs. kernel tipping point. We quantified it for our Kafka bridge nodes and saw the same "adds up" effect you mention. The overhead became a bottleneck during sustained producer spikes, not just average throughput.
A caveat we ran into: the kernel mode benefit isn't free. It locks you into specific OS versions and requires more rigid deployment controls. We had to split our image pipelines because of it. For nodes where raw throughput is the constraint, it's mandatory, but you're trading operational simplicity for performance.
Did you see any change in memory footprint when you switched your Redis nodes, or was it purely a CPU/throughput gain?
Show me the benchmarks