Mount Mayhem at Netflix: Scaling Containers on Modern CPUs

This is a nice writeup.

The discussion of CPU microarchitecture and how it affects the performance of heavily-lock-contended programs running on CPUs with many cores is interesting but maybe not really salient to solving the problem here. I feel that it's akin to comparing BMW vs Ferrari sports cars by how fast they can drive after all four tyres have been removed. These are all many-core CPUs so regardless of how gracefully they are at queue up for a single lock, if you make them all queue for a single lock then you still are only getting at best 100/(number of cores)% of the performance those CPUs should be able to deliver.

I do have some nagging doubts about what you might call technical profligacy here:

• is it really helpful for those containers to have 50 layers each? are all 50 layers contributing a large amount of space savings by allowing sharing of common files? could some of those perhaps be squashed together into a smaller number of layers?
• is it really helpful to be spinning up hundreds of containers in response to one user request? could you perhaps use some kind of futuristic 1970s technology such as "multiple executables" to have some software from multiple teams go into a single container, or even "libraries" in order to have some software from multiple teams go into a single process?

That said, Netflix do run a successful business that makes a lot of money by delivering a movie service that, as far as I've seen, has practically 100% uptime, so empirically none of these things are showstopper problems.