package kafka-eio

Changes

0.1.1

• Add /usr/local and Homebrew C include/library paths for librdkafka on FreeBSD and macOS.
• Skip the /proc/self/fd fd-leak regression test on platforms without procfs.

0.1.0

• Initial standalone OPAM package: Kafka_producer and Kafka_consumer on top of librdkafka, extracted from the Sun platform after real in-tree usage. Idempotent/ transactional delivery, consumer-group streaming with explicit ack, partition-aware concurrency, and typed transport security (Kafka_security.t: plaintext/SSL/SASL) on every producer, consumer, and service config.
• Delivery receipts flow through a Unix pipe from the C delivery callback (thread-safe, no OCaml runtime needed from C); a background Eio fiber reads the pipe and resolves pending promises.
• Blocking C calls (ocaml_rd_kafka_flush, ocaml_rd_kafka_consumer_close, ocaml_rd_kafka_consumer_poll) release the OCaml domain lock at the FFI boundary around the blocking librdkafka call, so the GC can run and Eio's cancellation delivers cleanly once the call returns — not worked around with Eio_unix.run_in_systhread or polling loops at the OCaml level.

Post-tag fix (#1): mutex-unlock skipped under Eio cancellation

Every Mutex.lock t.mutex; ...; Mutex.unlock t.mutex pair in Kafka_producer (delivery callback dispatch, pending-promise bookkeeping in produce_await, and cleanup in close) and the poll_exit_r resolution in Kafka_consumer's poll_fiber skipped the unlock/resolve step if an Eio.Cancel.Cancelled (or any other exception) was raised between the lock and unlock — leaving the mutex held forever, or a fiber awaiting poll_exit_r hanging forever after close. Fixed by wrapping every such critical section in Fun.protect ~finally, so the unlock/resolve always runs regardless of how the protected code exits.