Story

I built a message broker that speaks the Kafka wire protocol, so any Kafka client (librdkafka, kafka-python, kcat, etc.) works without code changes.

  The entire binary is 52KB. No JVM, no ZooKeeper, no third-party libraries —
  just C++20 with kqueue/epoll. Starts in <10ms, uses 0% CPU when idle.

  I built this because running Kafka locally for development is painful —
  gigabytes of RAM, slow startup, ZooKeeper/KRaft configuration. I just
  wanted something that accepts produce requests and gets out of the way.

  Technical details:
  - Single-threaded event loop (kqueue on macOS, epoll on Linux)
  - Memory-mapped log segments (1GB pre-allocated, sequential I/O)
  - Lock-free SPSC/MPSC ring buffers with cache-line alignment
  - Kafka protocol v0-v3 including flexible versions (ApiVersions, Metadata, Produce)
  - Auto-topic creation on first produce or metadata request

  The most interesting bug I hit: librdkafka sends ApiVersions v3, which uses
  Kafka's "flexible versions" encoding. But there's a special exception in the
  protocol — ApiVersions responses must NOT include header tagged_fields for
  backwards compatibility. One extra byte shifted every subsequent field,
  causing librdkafka to compute a ~34GB malloc that crashed immediately.

  Current limitations: no consumer groups, no replication, single-threaded,
  no auth. It's v0.1.0 — consume support is next.

  MIT licensed, runs on macOS (Apple Silicon + Intel) and Linux.