A cluster node joins a busy broker cluster and reports healthy — does the load on the existing nodes drop yet?

## Membership and capacity are two different events When a cluster node starts and joins a broker or streaming cluster, it announces itself: the cluster records that a machine is present, reachable and willing to hold data. That is all joining does. It does not hand the newcomer a single **unit of ownership** — a partition (or queue) that exactly one node serves at a time. Until some unit is assigned to it, the node stores no records, accepts no writes and answers no reads. This is the gap that surprises people. The machine is provisioned, the process is healthy, the member count went up by one — and every existing node is working exactly as hard as it was a minute ago. The member list grew; the capacity did not. ## Why the relief arrives late, and costs something on the way Getting work onto the newcomer takes two further steps. 1. **Declare** a new mapping from units of ownership to the nodes that should hold them — a **reassignment plan**. Some platforms expect an operator to produce it, some compute one themselves, and some hosted ones do it invisibly. 2. **Execute** it. Executing means the newcomer fetches the records of each unit assigned to it from whichever node holds them now, until its **copy** — one complete stored duplicate of that unit — is current enough to join **the caught-up set** and be allowed to serve. Step 1 takes a second. Step 2 takes the bytes involved divided by whatever bandwidth is available for copying them, and it reads those bytes off disks that are already serving production traffic. A node that will eventually hold terabytes is useless for hours. Worse, while the copying runs the cluster is doing *more* work than before, not less: the existing nodes serve their normal traffic plus the outbound copying. | Stage | Typical elapsed | What the new node serves | |---|---|---| | Process up, listed as a member | seconds | nothing | | Units of ownership assigned to it | seconds | nothing yet | | Its copies are being filled | minutes to hours | nothing yet | | Its copies reach the caught-up set | — | reads, and leadership may move to it | Only the last row is relief. ## The first exception: detached storage Some designs keep a unit's records on shared or remote storage rather than on a node's local disk. There, ownership of a unit can be handed to the newcomer without copying anything: the records already sit where the new owner can read them, so it begins serving almost at once. Anything the previous owner has not yet pushed to the shared store still has to be settled, and caches start cold, so it is not literally free — but it is a change of seconds rather than a data movement of hours. That is the real fork in this subject: **where records live on local disks, capacity follows data; where storage is detached, capacity follows ownership.** ## The second exception: letting new work land there Not all relief requires moving anything. Where units of ownership are created continuously — new queues, new streams, a new period's units — a platform can place the *new* ones on the newcomer and leave existing ones untouched. Relief then arrives in proportion to how much of the traffic is new, with no copying at all. It is slow, cheap and only ever partial, so it complements a reassignment rather than replacing one. Queue-shaped brokers where a queue lives on its owning node behave the same way: the newcomer takes the queues placed on it, never the ones already elsewhere. ## What to do with this as an operator - **Expand ahead of the need, not at the moment of it.** The lead time is the copying, and it is longest exactly when the cluster is busiest. - **Measure relief in units served and catch-up distance**, not in node count. Catch-up distance is how far behind its leader a copy still is; while it is large, that copy consumes capacity rather than providing it. - **Expect the bill before the benefit.** The machine is charged from the moment it boots and contributes from the moment its copies are current. - **Do not read an empty newcomer as a broken one.** A node holding nothing an hour after joining is normal if nothing was assigned to it; it is a fault only if a reassignment was declared and no records are arriving. - **Say which of the two shapes you are on** before promising a timeline. The same sentence — "we added a node" — means hours on one design and seconds on another. The compact model: joining is an announcement, assignment is a decision, copying is the work, and capacity appears at the end of the third.