Kafka Consumer Assignment Visualizer

RangeAssignor works per topic, and that concentrates the low partitions

Range sorts the members, then for each topic separately hands out contiguous blocks starting from the first member. Subscribe to five topics of three partitions each with three members and member 0 gets partition 0 of all five. When partition 0 is hot, which on a keyed topic with a skewed key space it often is, that skew lands on the same consumer every time. It is also the default for the plain consumer, which is why this shape is so common.

3 members, topics orders=3 payments=3

range        m0: orders-0 payments-0
             m1: orders-1 payments-1
             m2: orders-2 payments-2

# fine here. With uneven partition counts the
# remainders pile onto the early members instead of
# averaging out across topics.

RoundRobinAssignor spreads better and moves almost everything

It lays every partition of every subscribed topic into one list and deals them round robin. The balance across multiple topics is better than range. The cost is that any membership change shifts the deal by one, so nearly every partition moves to a different consumer, which for a stateful consumer means rebuilding nearly every cache.

Stickiness is invisible in a first assignment

With no previous assignment to preserve, the sticky assignors aim purely for balance and produce the same shape as round robin. Their advantage shows up on the second assignment, where they keep members on the partitions they already had. This tool computes assignments from scratch, so it shows you the balance and not the stickiness, and it says so rather than implying otherwise. Balance is only half of why CooperativeStickyAssignor is the right default; the other half is that it revokes only the partitions that move.

Eager against cooperative is a different axis

Range and round robin both use the eager protocol: every partition is revoked from every member before reassignment, including the ones coming straight back. CooperativeStickyAssignor revokes only what moves and pays two rounds for it. The rebalance duration estimator on this site quantifies the difference; this page shows what each one assigns.

More consumers than partitions is permanent waste

Partitions are the ceiling on consumer parallelism. An extra consumer is assigned nothing, and it still joins every rebalance and still counts toward the join barrier, so it makes rebalances slower while consuming nothing. Partitions can be added and never removed, so the count is worth deciding up front.