kafkars

Rust-based, Arrow-powered Python Kafka client for high-throughput data pipelines.

Motivation

Python's Global Interpreter Lock (GIL) and memory management create bottlenecks when consuming high-volume Kafka streams. Traditional Python Kafka clients process messages one at a time, requiring serialization/deserialization overhead for each message and limiting throughput.

kafkars solves this by:

• Rust core: All Kafka operations (polling, buffering, ordering) happen in Rust, bypassing the GIL
• Batch processing: Messages are accumulated and returned as Apache Arrow RecordBatches, not individual Python objects
• Zero-copy where possible: Arrow's columnar format enables efficient data transfer between Rust and Python
• Vectorized operations: Process thousands of messages at once with pandas, polars, or any Arrow-compatible library

This architecture is ideal for:

• Real-time analytics pipelines
• ML feature stores consuming from Kafka
• High-volume event processing
• Data lake ingestion

Important: Analytics-Focused Design

kafkars does not commit offsets. It is designed for analytics and high-throughput batch processing, not transactional workloads.

• No exactly-once semantics: Messages may be reprocessed if your application restarts
• No offset tracking: You control where to start reading via offset policies
• Stateless consumers: Each consumer instance starts fresh based on the configured policy

If you need exactly-once processing, transactional guarantees, or automatic offset management, use a traditional Kafka client like confluent-kafka-python.

Features

• Ordered delivery: Messages released in timestamp order across all partitions
• Flexible offset policies: Start from earliest, latest, or any timestamp
• Backpressure management: Automatically pauses fast partitions to prevent memory overflow
• Arrow-native output: Returns PyArrow RecordBatch for efficient downstream processing

Installation

pip install kafkars

Quick Start

from kafkars import ConsumerManager, SourceTopic

# Define source topics with offset policies
topics = [
    SourceTopic.from_earliest("events"),
    SourceTopic.from_relative_time("metrics", 3600_000),  # 1 hour ago
]

# Create consumer
manager = ConsumerManager(
    config={"bootstrap.servers": "localhost:9092"},
    topics=topics,
)

# Poll returns PyArrow RecordBatch
while True:
    batch = manager.poll(timeout_ms=1000)
    if batch.num_rows > 0:
        # Convert to pandas/polars for processing
        df = batch.to_pandas()
        print(f"Received {len(df)} messages")

    if manager.is_live():
        print("Caught up to real-time")
        break

Message Schema

Each poll returns a RecordBatch with the following schema:

Column	Type	Description
key	binary	Message key (nullable)
value	binary	Message payload (nullable)
topic	utf8	Source topic name
partition	int32	Partition number
offset	int64	Message offset
timestamp	timestamp[ms, UTC]	Message timestamp

Offset Policies

Policy	Description
from_earliest(topic)	Start from the beginning
from_latest(topic)	Start from the end (new messages only)
from_relative_time(topic, ms)	Start from N milliseconds ago
from_absolute_time(topic, ms)	Start from specific Unix timestamp

Documentation

Full documentation is available at kafkars.readthedocs.io.

License

Apache 2.0