APCu is a very performant form of in memory cache, almost as fast as using local variables and thus a good candidate for high load applications.

Despite that it's usage is rare because it resides in a single computing node. Most high load applications use an underlying infrastructure of autoscaling VMs or K8s, which means that there are multiple computing nodes with no cache coherency. This presentation explores these issues, possible strategies for establishing internode communications (hint: they don't reliably work), infrastructure side solutions and how one can use the APCu despite this shortcoming. We will present in broad terms how Altenar GP incorporated APCu in our system that receives 3.5K requests per second, the failures we had along the way and the performance benefits we saw with our final implementation. We also clarify critical shortcomings of the APCu documentation that may have severe consequences.