Zerobus Python SDK

A high-performance Python client for streaming data ingestion into Databricks Delta tables using the Zerobus service.

Table of Contents

• Overview
• Features
• Installation
• Quick Start
  • JSON (Simplest)
  • Protocol Buffers
  • Acknowledgments and throughput
• Configuration
• Error Handling
• Handling Stream Failures
• Performance Tips
• API Reference
• Debugging
• Building from Source
• Community and Contributing
• License

Overview

The Zerobus Python SDK is a thin wrapper around the Zerobus Rust SDK, built using PyO3 bindings. It delivers native performance with a Python-friendly API supporting both synchronous and asynchronous usage.

What is Zerobus? See the project overview for details on the Zerobus service.

Prerequisites (workspace setup, table creation, service principal): See the top-level README.

Architecture

┌─────────────────────────────────────────┐
│         Python Application Code         │
└─────────────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────┐
│       Python SDK (Thin Wrapper)         │
│    • Sync and async APIs                │
│    • Python types & error handling      │
└─────────────────────────────────────────┘
                    │
                    ▼ (PyO3 bindings)
┌─────────────────────────────────────────┐
│         Rust Core Implementation        │
│    • gRPC communication                 │
│    • OAuth 2.0 authentication           │
│    • Stream management & recovery       │
└─────────────────────────────────────────┘

Features

• Rust-backed performance - Native Rust implementation via PyO3 bindings for maximum throughput
• Sync and Async support - Both synchronous and asynchronous Python APIs
• Automatic recovery - Built-in retry and reconnection for transient failures
• Multiple serialization formats - JSON (simple) and Protocol Buffers (type-safe)
• OAuth 2.0 authentication - Secure authentication with client credentials, automatically refreshed
• Acknowledgment callbacks - Receive notifications when records are acknowledged or encounter errors
• Flexible configuration - Fine-tune timeouts, retries, and recovery behavior

Installation

From PyPI (Recommended)

pip install databricks-zerobus-ingest-sdk

Pre-built wheels are available for:

• Linux: x86_64, aarch64 (manylinux)
• macOS: x86_64, arm64
• Windows: x86_64

Python Version

Requires Python 3.9-3.14. The SDK is tested on CPython 3.9 through 3.14.

The wheel uses the CPython stable ABI (abi3), so one wheel works on every supported version. The free-threaded builds (like 3.14t) are not supported, because the stable ABI does not cover them.

Dependencies

• protobuf >= 4.25.0, < 7.0 (for Protocol Buffer schema handling)
• requests >= 2.28.1, < 3 (only for the generate_proto utility tool)

All core ingestion functionality (gRPC, OAuth, stream management) is handled by the native Rust implementation.

Arrow Flight ingestion needs pyarrow. Install it with the arrow extra:

pip install "databricks-zerobus-ingest-sdk[arrow]"

The extra selects the pyarrow version for your interpreter, because no single pyarrow release covers Python 3.9 through 3.14. On Python 3.14 the extra installs pyarrow 22.0.0 or later, which is the first release with 3.14 wheels. On earlier versions it installs pyarrow below 22.0.0. Core ingestion (Protobuf and JSON) does not need pyarrow at all.

Quick Start

Choose Your Serialization Format

1. Protocol Buffers (Recommended) - Strongly-typed schemas with compact binary encoding. More efficient over the wire and the best choice for production and high-throughput workloads.
2. JSON - Simple, no schema compilation needed. Good for getting started or quick prototyping, but each record carries higher per-record overhead (text serialization plus UTF-8 validation), so it is slower than Protocol Buffers for high-volume ingestion.

Option 1: JSON (Simplest)

Synchronous:

from zerobus.sdk.sync import ZerobusSdk
from zerobus.sdk.shared import TableProperties

server_endpoint = "https://1234567890123456.zerobus.us-west-2.cloud.databricks.com"
workspace_url = "https://dbc-a1b2c3d4-e5f6.cloud.databricks.com"

sdk = ZerobusSdk(server_endpoint, workspace_url)
table_properties = TableProperties("main.default.air_quality")
stream = sdk.create_stream(client_id, client_secret, table_properties)

try:
    for i in range(100):
        offset = stream.ingest_record_offset({
            "device_name": f"sensor-{i % 10}",
            "temp": 20 + (i % 15),
            "humidity": 50 + (i % 40)
        })
    stream.flush()
finally:
    stream.close()

Asynchronous:

import asyncio
from zerobus.sdk.aio import ZerobusSdk
from zerobus.sdk.shared import TableProperties

async def main():
    server_endpoint = "https://1234567890123456.zerobus.us-west-2.cloud.databricks.com"
    workspace_url = "https://dbc-a1b2c3d4-e5f6.cloud.databricks.com"

    sdk = ZerobusSdk(server_endpoint, workspace_url)
    table_properties = TableProperties("main.default.air_quality")
    stream = await sdk.create_stream(client_id, client_secret, table_properties)

    try:
        for i in range(100):
            offset = await stream.ingest_record_offset({
                "device_name": f"sensor-{i % 10}",
                "temp": 20 + (i % 15),
                "humidity": 50 + (i % 40)
            })
        await stream.flush()
    finally:
        await stream.close()

asyncio.run(main())

Acknowledgments and throughput

Ingestion is asynchronous. ingest_record_offset() returns as soon as the record is queued; the SDK sends it and tracks its acknowledgment in the background. To confirm records are durably committed, call flush() — it returns once everything queued so far is acknowledged. The idiomatic flow is ingest in a loop, then flush() (once for a bounded batch, or periodically for a long-running stream); or register an AckCallback to be notified as records commit.

Each ingest also returns the record's offset, and wait_for_offset(offset) blocks until that offset is acknowledged — handy when a specific record must be confirmed before continuing (acks are ordered, so waiting on the last offset confirms the whole run). Just avoid calling wait_for_offset() after every record in a tight loop, since that limits throughput to one record per round-trip.

Option 2: Protocol Buffers

First, define a protobuf schema. Use proto2 syntax with optional fields to match Delta table columns:

// record.proto
syntax = "proto2";
message AirQuality {
    optional string device_name = 1;
    optional int32 temp = 2;
    optional int64 humidity = 3;
}

See the Delta → Protobuf type mappings in the top-level README.

Compile the schema to generate a Python module:

pip install "grpcio-tools>=1.60.0,<2.0"
python -m grpc_tools.protoc --python_out=. --proto_path=. record.proto
# Generates record_pb2.py

Load the descriptor from the generated module and pass it to TableProperties:

import record_pb2

# The DESCRIPTOR is the compiled schema — pass it so the SDK can validate records
table_properties = TableProperties("main.default.air_quality", record_pb2.AirQuality.DESCRIPTOR)

Alternatively, generate the schema automatically from an existing Unity Catalog table:

python -m zerobus.tools.generate_proto \
    --uc-endpoint "https://dbc-a1b2c3d4-e5f6.cloud.databricks.com" \
    --client-id "your-client-id" \
    --client-secret "your-client-secret" \
    --table "main.default.air_quality" \
    --output "record.proto" \
    --proto-msg "AirQuality"

# Then compile the generated file the same way:
python -m grpc_tools.protoc --python_out=. --proto_path=. record.proto

Synchronous:

from zerobus.sdk.sync import ZerobusSdk
from zerobus.sdk.shared import TableProperties
import record_pb2

sdk = ZerobusSdk(server_endpoint, workspace_url)
table_properties = TableProperties("main.default.air_quality", record_pb2.AirQuality.DESCRIPTOR)
stream = sdk.create_stream(client_id, client_secret, table_properties)

try:
    for i in range(100):
        record = record_pb2.AirQuality(
            device_name=f"sensor-{i % 10}",
            temp=20 + (i % 15),
            humidity=50 + (i % 40)
        )
        stream.ingest_record_offset(record)
    stream.flush()
finally:
    stream.close()

Asynchronous:

import asyncio
from zerobus.sdk.aio import ZerobusSdk
from zerobus.sdk.shared import TableProperties
import record_pb2

async def main():
    sdk = ZerobusSdk(server_endpoint, workspace_url)
    table_properties = TableProperties("main.default.air_quality", record_pb2.AirQuality.DESCRIPTOR)
    stream = await sdk.create_stream(client_id, client_secret, table_properties)

    try:
        for i in range(100):
            record = record_pb2.AirQuality(
                device_name=f"sensor-{i % 10}",
                temp=20 + (i % 15),
                humidity=50 + (i % 40)
            )
            await stream.ingest_record_offset(record)
        await stream.flush()
    finally:
        await stream.close()

asyncio.run(main())

See the examples/ directory for complete runnable examples.

Configuration

Configure stream behavior by passing a StreamConfigurationOptions object to create_stream():

from zerobus.sdk.shared import AckCallback, StreamConfigurationOptions

class MyCallback(AckCallback):
    def on_ack(self, offset: int):
        print(f"Acknowledged offset: {offset}")

    def on_error(self, offset: int, error_message: str):
        print(f"Error at offset {offset}: {error_message}")

options = StreamConfigurationOptions(
    max_inflight_records=10000,
    recovery=True,
    ack_callback=MyCallback()
)

stream = sdk.create_stream(client_id, client_secret, table_properties, options)

Available Options

The record format is inferred from TableProperties: omitting descriptor_proto selects JSON, while providing a Protobuf descriptor selects Protobuf. record_type is retained for backward compatibility but does not select the format.

Option	Type	Default	Description
record_type	RecordType	RecordType.PROTO	Retained for backward compatibility; format comes from TableProperties.descriptor_proto
max_inflight_records	int	1000000	Maximum number of unacknowledged records
recovery	bool	True	Enable automatic stream recovery
recovery_timeout_ms	int	15000	Timeout for recovery operations (ms)
recovery_backoff_ms	int	2000	Delay between recovery attempts (ms)
recovery_retries	int	4	Maximum number of recovery attempts
flush_timeout_ms	int	300000	Timeout for flush operations (ms)
server_lack_of_ack_timeout_ms	int	60000	Server acknowledgment timeout (ms)
stream_paused_max_wait_time_ms	Optional[int]	None	Max wait during graceful stream close. None = full server duration, 0 = immediate, x = min(x, server_duration)
callback_max_wait_time_ms	Optional[int]	5000	Max wait for callbacks after close(). None = wait forever
ack_callback	AckCallback	None	Callback invoked once per successfully queued ingest submission that later acknowledges or fails

Error Handling

The SDK raises two categories of exception. Handle them differently:

• Retriable (ZerobusException): transient conditions such as network issues or temporary server errors. Safe to retry, ideally with backoff. The SDK's built-in recovery already handles many of these for you.
• Non-retriable (NonRetriableException): fatal conditions such as invalid credentials or a missing table. Retrying won't help. Fix the underlying problem.

NonRetriableException is a subclass of ZerobusException, so a bare except ZerobusException still catches both.

from zerobus.sdk.shared import ZerobusException, NonRetriableException

try:
    stream.ingest_record_offset(record)
except NonRetriableException as e:
    # Fatal: do not retry; fix the cause (credentials, table, schema)
    raise
except ZerobusException as e:
    # Transient: retry with backoff
    ...

Handling Stream Failures

The SDK automatically handles retries for transient errors. Enqueue, flush, and close failures surface as ZerobusException or NonRetriableException. get_unacked_records() and recreate_stream() succeed only after the stream has already closed, which a terminal failure does. An enqueue failure leaves the stream active, so those calls fail; raise the original error and keep the stream. recreate_stream() re-queues records that were already accepted; it does not retry a payload that failed to enqueue.

from zerobus.sdk.shared import ZerobusException

try:
    for i in range(10000):
        stream.ingest_record_offset(record)
    stream.flush()
except ZerobusException as e:
    print(f"Ingestion failed: {e}")
    try:
        unacked = list(stream.get_unacked_records())
    except ZerobusException:
        raise e
    print(f"{len(unacked)} previously queued records were unacknowledged.")
    try:
        new_stream = sdk.recreate_stream(stream)
        try:
            new_stream.flush()
        finally:
            new_stream.close()
    except ZerobusException:
        raise e
else:
    stream.close()

Use get_unacked_batches() to inspect the original batch grouping after the stream closes:

unacked_batches = list(stream.get_unacked_batches())
print(f"{len(unacked_batches)} batches remain unacknowledged")

Decoding unacked records:

• JSON mode: json.loads(record_bytes.decode('utf-8'))
• Protobuf mode: YourMessage.FromString(record_bytes)

Performance Tips

The reliable bulk path is ingest_records_offset() plus one flush(). That call amortizes the Python-to-Rust crossing and returns an offset after the batch is queued. For single records, use ingest_record_offset() in a loop and flush() once. Ingest calls queue immediately and the SDK acknowledges records in the background, so a single flush() confirms everything queued so far. The ack watermark is monotonic, so if you want a durability checkpoint mid-stream, waiting on the last offset returned confirms every prior record. In async code, an AckCallback tracks durability without blocking. Calling wait_for_offset() after every record in a tight loop limits throughput to one record per round-trip, so save it for confirming a specific record.

ingest_record_nowait() and ingest_records_nowait() spawn detached tasks and discard enqueue errors. flush() can complete before those tasks allocate offsets, so they are not a safe durability path. Prefer the offset APIs.

Method	Throughput	Use case
ingest_records_offset()	Highest	Recommended bulk path: queue a batch, then flush() once
ingest_record_offset()	Medium	Recommended for single records: ingest in a loop, then flush() once
ingest_record()	Low	Deprecated; prefer offset-based APIs
ingest_record_nowait()	Unsafe	Detached fire-and-forget; enqueue errors can be lost and are not synchronized with flush()
ingest_records_nowait()	Unsafe	Detached batch fire-and-forget; same durability caveats as ingest_record_nowait()

Idiomatic flow:

async def ingest_all(stream, records):
    await stream.ingest_records_offset(records)     # queues the batch, no round-trip
    await stream.flush()                            # one wait for everything

Confirming a specific record (waiting on the last offset confirms all prior records):

async def ingest_and_confirm(stream, records):
    offset = None
    for record in records:
        offset = await stream.ingest_record_offset(record)
    if offset is not None:
        await stream.wait_for_offset(offset)        # confirm the run before continuing

API Reference

ZerobusSdk

Main entry point. Sync: from zerobus.sdk.sync import ZerobusSdk / Async: from zerobus.sdk.aio import ZerobusSdk

sdk = ZerobusSdk(
    host="https://<workspace>.zerobus.<region>.cloud.databricks.com",
    unity_catalog_url="https://<workspace-host>",
    application_name="my-app/1.0",
)

application_name is optional; when set it is appended to the user-agent header on gRPC requests to the Zerobus server (not on the OAuth token requests to the login service). It follows the "<product>/<version>" convention (e.g. my-app/1.0).

# Sync
stream = sdk.create_stream(client_id, client_secret, table_properties, options=None, headers_provider=None)
# Async
async def create_async_stream(sdk):
    return await sdk.create_stream(client_id, client_secret, table_properties, options=None, headers_provider=None)

ZerobusStream

Single record ingestion:

Method	Sync	Async	Notes
ingest_record_offset(record)	→ int	await → int	Recommended for single records; returns offset after queueing
ingest_record(record)	→ RecordAcknowledgment	await → Awaitable	Deprecated since v0.3.0
ingest_record_nowait(record)	→ None	→ None (not async)	Detached fire-and-forget; enqueue errors are not synchronized with flush()

Batch ingestion:

Method	Sync	Async	Notes
ingest_records_offset(records)	→ int	await → int	Recommended bulk path; returns the batch's final offset
ingest_records_nowait(records)	→ None	→ None (not async)	Detached fire-and-forget; same durability caveats as ingest_record_nowait()

Accepted record types:

• JSON mode: dict (SDK serializes) or str (pre-serialized JSON)
• Protobuf mode: Message object (SDK serializes) or bytes (pre-serialized)

Offset tracking (use to confirm a specific record before continuing; for bulk durability, ingest in a loop and flush() once):

# Sync
offset = stream.ingest_record_offset(record)
# ... do other work ...
stream.wait_for_offset(offset)  # Block until durably written

# Async
async def confirm_async(stream, record):
    offset = await stream.ingest_record_offset(record)
    # ... do other work ...
    await stream.wait_for_offset(offset)  # Block until durably written

Acks are ordered, so waiting on the last offset returned confirms all prior records too.

Stream management:

# Sync
stream.flush()   # Wait for all pending records to be acknowledged
stream.close()   # Flush and close gracefully (always call in finally)

# Async
async def close_async(stream):
    await stream.flush()
    await stream.close()

Unacknowledged records:

# Sync
records = stream.get_unacked_records()   # Iterator[bytes]
batches = stream.get_unacked_batches()  # Iterator[List[bytes]]

# Async
async def get_unacked_async(stream):
    records = await stream.get_unacked_records()
    batches = await stream.get_unacked_batches()
    return records, batches

TableProperties

# JSON mode
TableProperties("catalog.schema.table")

# Protobuf mode
TableProperties("catalog.schema.table", descriptor_proto=MyMessage.DESCRIPTOR)

StreamConfigurationOptions

See Configuration for full parameter list.

AckCallback

from zerobus.sdk.shared import AckCallback

class MyCallback(AckCallback):
    def on_ack(self, offset: int) -> None:
        # Called once for each acknowledged single-record or batch submission
        pass

    def on_error(self, offset: int, error_message: str) -> None:
        # Called once when a single-record or batch submission encounters an error
        pass

close() waits at most callback_max_wait_time_ms (default 5000 ms) for in-flight callbacks. A callback per queued submission is not guaranteed if that budget expires.

HeadersProvider

For custom authentication (e.g. custom token providers), implement HeadersProvider and pass it to create_stream(). Must include both authorization and x-databricks-zerobus-table-name headers. See examples/ for implementation details.

RecordAcknowledgment (Sync only, deprecated)

ack.wait_for_ack(timeout_sec=None)  # Block until acknowledged
ack.is_done() -> bool

Exceptions

• ZerobusException(message, cause=None) - Base exception; retryable SDK failures are raised as this type
• NonRetriableException(message, cause=None) - Subclass for fatal errors (ZerobusError::is_retryable() is false)

Debugging

The SDK uses Rust's tracing framework. Control log levels via RUST_LOG:

export RUST_LOG=info           # Default
export RUST_LOG=debug          # Detailed debugging
export RUST_LOG=trace          # Very verbose
export RUST_LOG=zerobus_sdk=debug  # Only SDK components

Building from Source

Building from source requires the Rust toolchain 1.88 or newer (install from rustup.rs).

git clone https://github.com/databricks/zerobus-sdk.git
cd zerobus-sdk/python
make dev    # Set up venv and install in editable mode
make test   # Run tests
make build  # Build release wheel

For development workflows and detailed instructions, see CONTRIBUTING.md.

Community and Contributing

We are keen to hear feedback. Please file issues.

See CONTRIBUTING.md for development setup and contribution guidelines.

License

This project is licensed under the Apache License 2.0. See LICENSE for the full text.