The control plane records node state far less often

The control plane now saves each node's peer-discovery details only when they change, instead of on every heartbeat, cutting its routine database writes about sevenfold. Delivery is unchanged; two rare recovery cases behave slightly differently.

Every edge node checks in with the control plane every 30 seconds. Until now, each check-in rewrote the node’s details in the database edge nodes use to find each other, whether or not anything had changed. The same was true of two pieces of usage-metering bookkeeping.

Now:

Peer-discovery details are written when a node’s address, region, status or location changes, and otherwise refreshed every 15 minutes.
Metering bookkeeping is saved once a minute or once every five minutes, instead of on every check-in.

Together that is about seven times fewer writes. Nothing about how content is served or how usage is counted changes during normal operation.

Two edge cases change, and both apply only if the control plane itself stops without shutting down cleanly:

A node that went offline can stay listed as a peer for up to an hour, rather than ten minutes. Requests already route around a peer that doesn’t answer.
Up to five minutes of one node’s traffic can be counted again after the restart. Usually this restores traffic the crash would otherwise have lost. It only double-counts if the crash lands within five minutes of the hourly usage save.

A clean shutdown, including every normal deploy, saves everything first and is unaffected.