KAFKA-21140: Add randomized testbed for streams group task assignors

[gabriellefu]

• TaskAssignorTestbed — generates a random topology and group from
  a seed
  (two profiles: SMALL for logical corners, LARGE for production
  shape), runs the
  assignor until the assignment is a fixed point, then applies random
  rebalance
  events (process add/drop/restart/retag, member add/drop, standby
  count change)
  and converges again. A failure prints the seed and the full history;
  replay a
  single scenario with STREAMS_ASSIGNOR_FUZZ_SEED=<seed> or shift
  the whole run
  with STREAMS_ASSIGNOR_FUZZ_BASE_SEED=<seed> (system properties
  work from an IDE).
  • AssignmentInvariants — properties every assignor must satisfy,
    asserted on
    every assignment: output covers exactly the group members, every
    task is active
    on exactly one member, standbys only on stateful tasks, at most
    numStandbyReplicas standbys per task, no process holds two copies
    of a task.
  • AssignmentMetrics — grades an assignment without asserting:
    active/total
    spread per member, load spread per process, process/member
    stickiness, task
    moves, rack diversity per tag. The summary over a run is printed so
    two versions
    of an assignor can be compared.
  • StickyTaskAssignorFuzzTest — runs StickyTaskAssignor through
    the testbed
    and adds the checks its algorithm guarantees: no member above
    ceil(activeTasks / members), every stateful task gets exactly
    min(numStandbyReplicas, processes - 1) standbys, and rack-aware
    tags do not
    change the active assignment.

Reviewers: Lucas Brutschy lbrutschy@confluent.io

[Copilot]

Copilot review overview

🟡 Changes recommended

The convergence tests currently pass when the assignment never reaches a fixed point.

Review effort: Balanced
Findings: 1

Open (1)

• Fail convergence tests when assignment oscillation exceeds cutoff · New

What changed in this PR

Adds deterministic randomized testing for Streams task assignors across topology and group changes.

Changes:

• Generates seeded small and production-scale scenarios with rebalance events.
• Validates assignment invariants and records quality metrics.
• Exercises StickyTaskAssignor-specific guarantees.

File	Description
TaskAssignorTestbed.java	Generates scenarios, events, and convergence runs.
StickyTaskAssignorFuzzTest.java	Adds randomized StickyTaskAssignor tests.
AssignmentMetrics.java	Calculates assignment quality metrics.
AssignmentInvariants.java	Validates universal assignment constraints.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[lucasbru]

minor comment

[lucasbru]

warmupTasks is always Map.of() here, and MemberAssignment (the assignor's own output) never carries warmup tasks either, so a Scenario can never synthesize a member holding one. That means collectStandbyCandidates/isCurrentlyAssignedStandbyOrWarmupTask in StickyTaskAssignor (the warm-up stickiness logic from KAFKA-21001) never gets exercised by this fuzzer, even though the rest of the scenario generator is fairly thorough about modeling production-like changes.

[lucasbru]

Two things worth adding here, I think:

1. Timing. There's no reference implementation in this module to compare against, so I wouldn't gate CI on perf, that tends to be flaky, but it'd be useful to at least measure and print per-rebalance assignment time in the summary, so a future accidental complexity regression at least shows up in the output.

2. Right now this only runs as a JUnit test with the full SMALL/LARGE profiles, and the summary is print-only, nothing here is asserted except the assignor-specific checks. Could be worth splitting into two entry points: a main that runs the full profiles and prints the summary tables for manual comparison, and a separate fast test with a fixed seed and a small scenario count that asserts fixed bounds on a couple of the metrics, so a real quality regression actually fails the build instead of just printing a worse number in CI output that nobody reads.