Hands-on Labs

English · 中文版

Install the project once from the repository root:

uv sync

Both labs use temporary directories and the public Direct API, so they leave no broker data behind.

1. Leader failure with acks=1

uv run python -m minikafka.labs.leader_failure

Expected landmarks:

acknowledged offset: 0
consumer-visible end (HW): 0
records after failover: 0

Watch the distinction between producer acknowledgement and committed visibility. The leader accepts offset 0, but the follower has not fetched it. Promoting that follower removes the unreplicated tail. This is the same durability risk represented by Kafka's acks=1, not a claim that MiniKafka implements Kafka's election protocol.

2. Consumer-group rebalance

uv run python -m minikafka.labs.rebalance

Expected landmarks:

member 1: orders-0, orders-1, orders-2, orders-3
member 2: orders-1, orders-3
overlap after refresh: False

Watch the generation transition. The first consumer initially owns four partitions; after the second joins, its old local assignment remains until refresh_assignment(). The refreshed assignments no longer overlap. When the second member leaves, the first owns all partitions again. MiniKafka performs this reassignment synchronously and has no Kafka-style revoke barrier.

Continue with executable evidence

The labs narrate two trade-offs; the test suite covers the full mechanism set. Use the behavior matrix to choose a focused test, or run:

uv run pytest -q