Kafka's pieces are each simple — partitions, keys, consumer groups, offsets, lag — but together they tie people in knots. So I built an interactive playground where you produce and consume and watch it happen, no broker required.
▶ Live demo: https://dev48v.github.io/kafka-playground/
Source (single file, zero deps): https://github.com/dev48v/kafka-playground
The four ideas, made visible
1. Keys decide partitions. Produce a message with a key and it's hashed to a partition — and the same key always goes to the same partition:
key = "user-42" → hash("user-42") % partitionCount → Partition 0 (every time)
That single fact is why Kafka can only guarantee ordering within a partition, not across the topic. All of user-42's events stay ordered because they're all in one partition. Produce with no key and messages spread round-robin — faster fan-out, no per-key ordering.
2. A consumer group splits the partitions. Every partition is owned by exactly one consumer in the group. Add a consumer and the partitions rebalance across the group — that's how you scale out consumption.
3. More consumers than partitions = idle consumers. Give a 3-partition topic a 4th consumer and it just sits there. Partitions are the unit of parallelism, so your consumer count is capped by your partition count. Seeing the idle consumer light up grey makes this stick instantly.
4. Offsets and lag. Consuming advances the group's committed offset per partition. The gap between the newest message and the committed offset is lag. Produce faster than you consume and watch lag climb; flip on auto-consume and watch it drain.
Why a playground beats a diagram
You can read "partitions are the unit of parallelism" and still be surprised the first time a consumer goes idle. Dragging the consumer count past the partition count and seeing it happen teaches it in one move. Same with key stickiness — produce three messages with the same key, watch them stack in one partition, and the ordering guarantee stops being abstract.
It's one index.html. A deliberately simplified model (round-robin assignment vs Kafka's range/sticky/cooperative assignors; a small key hash vs murmur2) — but the behaviour you see matches real Kafka.
If this made Kafka click, a star helps others find it: https://github.com/dev48v/kafka-playground
