Node Join — Minimal Remap and Staged Handoff Diagram

Minimal remap on join

Only the joining node's ranges stream over the wire: ~9% of keys for 10 to 11, never a full rehash
Bulk copy first, then catch up on writes until lag is near zero, and only then flip ownership
In-flight requests during the flip are forwarded or safely retried — no stranded writes

Weights without rehash

A 2x node takes 2x virtual nodes (ring) or 2x logical shards (fixed count): shares follow capacity
Weights change only the new node's portion; every other key stays exactly where it was
Virtual nodes also smooth the ring: 100 per node keeps the worst load near 1.2x the mean

Movement versus routing

Data movement and routing updates are separate events: the map version flips after the data arrives
Clients cache the map with a short TTL; a stale cache reads from the old owner, which forwards
A leave mirrors a join: only the departed node's ranges are rehomed, replicas promote locally