agora-etl 0.3.3

Agora — async-first ETL framework for Python.

Agora ETL Framework

Async-first ETL framework for Python.

agora-etl is the core Agora runtime: a builder-first ETL framework organized around Source -> Middleware chain -> Sink(s).

It stays focused on runtime semantics and extension contracts:

• immutable pipeline composition
• per-record, batched, and Arrow-aware execution
• checkpointing and replay
• DLQ and sink failure policies
• backpressure, tracing, metrics, and health surfaces
• scheduled workers and config-driven runs

Official backend integrations such as Redis, Kafka, PostgreSQL, cron scheduling, and distributed coordination live in agora-etl-plugins. Optional acceleration belongs in agora-etl-rs.

Install

pip install agora-etl
pip install "agora-etl[file]"          # pyarrow + faster JSONL paths
pip install "agora-etl-plugins[all]"   # official first-party integrations

Quick Start

import asyncio

from agora import DeliveryConfig, IterableSource, Pipeline
from agora.core.dlq import SQLiteDLQSink
from agora.sinks.file.jsonlines import JsonLinesSink

records = [
    {"id": 1, "city": "Ho Chi Minh City", "confidence": 0.92},
    {"id": 2, "city": "Da Nang",          "confidence": 0.41},
    {"id": 3, "city": "Hanoi",            "confidence": 0.88},
]


async def main() -> None:
    summary = await (
        Pipeline(IterableSource(records))
        .filter(lambda row: row["confidence"] >= 0.5, name="confidence_gate")
        .build(
            JsonLinesSink(path="cities.jsonl"),
            config=DeliveryConfig(
                dlq=SQLiteDLQSink(".agora_dlq.db"),
                batch_size=100,
            ),
        )
        .run()
    )
    print(summary)


asyncio.run(main())

What this does:

• reads from an in-memory source
• filters low-confidence rows
• writes the survivors to cities.jsonl
• sends failed records to a local SQLite DLQ
• returns a PipelineRunSummary with counts, timing, and runtime signals

Why 0.3.0 Matters

Agora now exposes execution shape as a first-class contract.

Three shared data planes:

• python_rows
• python_batches
• arrow_batches

That vocabulary now shows up consistently across sources, middleware, writer planning, sink boundaries, runtime summaries, and tracing.

You can inspect the plan before the run starts:

from agora import Pipeline
from agora.sinks.file.csv import CsvSink

bound = Pipeline(source).build(CsvSink(path="out.csv", row_mapper=lambda row: row))

plan = bound.explain(max_records=1_000)
print(plan)
print(plan.to_dict())

PipelineExplain surfaces:

• selected execution lane
• source and writer data planes
• middleware compatibility matrix
• per-sink selected plane and downgrade markers

Execution Shapes

Agora supports more than one runtime shape, but keeps them explicit:

• row pipelines: standard Middleware / MapMiddleware / FilterMiddleware
• Python batch pipelines: stream_batches() plus BatchMiddleware
• Arrow pipelines: pyarrow.RecordBatch plus Arrow-native middleware and sinks

The important rule is that one middleware chain must stay internally compatible:

• Python-row chain is valid
• all-Arrow chain is valid
• mixed Arrow plus Python-row/list-batch chain is rejected during planning

For Arrow fan-out, Agora chooses the best path per sink:

• Arrow-capable sinks receive the original RecordBatch
• non-Arrow sinks downgrade only at the sink boundary

Builder Model

The pipeline builder is immutable.

base = (
    Pipeline(source)
    .pipe(NormalizeMiddleware())
    .filter(lambda row: row["confidence"] >= 0.5)
)

csv_pipeline = base.build(CsvSink(path="clean.csv", row_mapper=lambda row: row))
jsonl_pipeline = base.build(JsonLinesSink(path="clean.jsonl"))

Common composition patterns:

• .build(sink) for one destination
• .fan_out([...]) for writing the same record to many sinks
• .route(router) for sending each record to one matching sink
• .run(max_records=N) for bounded runs, now enforced at the source boundary

Delivery options are passed via DeliveryConfig, including:

• dlq
• checkpoint
• batch_size
• sink_failure_policy
• sink_concurrency
• backpressure
• tracer

Public Data-Plane API

The execution-shape contract is now public API:

from agora import DataPlane, SinkDataPlaneSpec, SourceDataPlaneSpec

source_spec = source.data_plane_spec()
sink_spec = sink.data_plane_spec()

Legacy data-plane booleans still work in 0.3.x, but they are now a compatibility bridge and emit DeprecationWarning when Agora has to infer a non-row plane from them.

Preferred direction:

• sources override data_plane_spec()
• sinks advertise accepted_data_planes and native_data_planes

Core vs Plugins

Keep this mental model:

• agora-etl: runtime semantics, pipeline contracts, CLI, docs
• agora-etl-plugins: official integrations
• agora-etl-rs: optional acceleration that preserves the same semantics

If a capability depends on Redis, Kafka, PostgreSQL, cron parsing, or distributed lease ownership, it likely belongs in the plugin bundle rather than the core package.

Documentation Map

• Quickstart
• Pipelines
• Failure Handling
• Checkpointing
• Observability
• Scheduling
• Plugins
• Architecture
• CLI
• Change Log

Reference sections:

• Sources
• Sinks
• Middlewares

License

Apache 2.0 - see LICENSE.