Kafka

pyrula.kafka is a Kafka client backed by Rust. The API tracks confluent-kafka, so existing knowledge carries over. Two things differ on purpose: every networked call takes an explicit timeout, and there’s a native record path for throughput.

from pyrula import kafka

On a 12-partition, 100k-record benchmark it runs ahead of confluent-kafka on every produce path (roughly 1.1x to 1.4x) and well ahead on Avro, at parity on plain consume, and an order of magnitude past the pure-Python clients. The Rust core does the work a Python produce/consume loop pays for per record. Full numbers and method on the Benchmarks page.

Wheels ship for Linux glibc and musl on x86_64 and aarch64, macOS on Apple silicon and Intel, and Windows x64 and arm64, so pip install pyrula-kafka is the whole install on Alpine and ARM too. The package depends on pyrula and nothing else; AWS MSK IAM auth is an optional extra.

Pages

• Quickstart: produce a batch, poll it back.
• Producer and Consumer: delivery, batching, groups, offsets.
• Transactions: exactly-once produce and consume.
• Formats: raw bytes, Avro, JSON.
• Admin: topics, partitions, configs.
• Configuration and Authentication (SASL, mTLS, OAuth).
• Examples: producer+consumer round trips, EOS, Avro.
• Client guarantees: what the client guarantees, capability by capability, every cell proven by a test.

Python defaults, the type stubs, and the Rust fallbacks are kept in sync, so what the docs say matches what the runtime does.