All high-performance database kernels do full kernel bypass e.g. they control their storage hardware, CPU affinity/context-switching, memory, network, etc explicitly. For all practical purposes Linux turns into little more than a device driver. This enables integer factor gains in performance via various optimizations. Ironically, it also makes the code simpler because behavior is explicit.

Linux is specifically designed to support this type of usage. The necessary syscalls were added decades ago, originally to support databases. Kubernetes intercepts these syscalls because they break its abstractions; while they appear to function like the underlying kernel syscall, the resultant behavior is not the same and generally unsuitable for these types of database architectures. The practical effect is degraded and unpredictable performance because it violates invariants that core optimizations rely on.

This has been kicked around by Kubernetes people for years, including within my own orgs because we use a lot of Kubernetes. No one has every been able to make this type of software achieve comparable performance, even when we've used a lot of hack-y workarounds. Kubernetes was not designed to allow software to interact with the Linux kernel in this way. Consequently, this type of software is deployed on VMs or bare metal in practice, even if everything else is on Kubernetes.

Sounds like parent misunderstood how PDs work in k8s or maybe is referring to THPs - still doesn’t make a ton of sense