Last week we fixed a bug in our pruning logic that was quietly costing us CPU and time to finality. The shape of the graph tells the story: things get worse when pruning starts, and go back to normal once the fix lands.
We're running a 512-validator devnet with ethlambda, as another step towards scaling to higher validator counts.
32 nodes, 4 subnets, one aggregator per subnet, 16 validators per node.