Comparison
EvictionvsOvercommit
Eviction
a healthy workload was terminated because the node it happened to be on ran out of disk, and it had done nothing wrong.
The platform terminating running workloads to relieve pressure on a node — memory, disk or PIDs — choosing victims by QoS class and by how far over their requests they are. It is a deliberate sacrifice rather than a failure, and it is why a workload can die for reasons entirely outside its own behaviour. Disk pressure from unbounded logs or image sprawl is the most common trigger and the easiest to prevent.
Full entry →Overcommit
the sum of every limit on the node is three times the node, and it works fine until the day everything is busy at once.
Allowing the total of workloads' limits to exceed a node's real capacity, on the assumption that they will not all peak together. It is the mechanism that makes shared infrastructure economical, and it is a bet rather than a configuration. The bet loses in exactly the situation you care about — a correlated traffic spike — which is when eviction and throttling arrive together and the platform looks like it caused the incident.
Full entry →