fpstreams

Documentation · Browser playground · Changelog · Contributing

Lazy Python pipelines for iterables and records, with bounded async concurrency.

Python 3.11 or newer is required.

The current stable release is 2.4.1. This maintenance release reorganizes planner and execution internals with no public API changes. Version 2.4.0 added stateful pipeline operations, composed aggregation, shared execution limits, file recovery, and versioned execution metadata. See the changelog for its scope and the performance guide for batch and executor examples. The roadmap lists the next optimization candidates and their acceptance criteria.

Install

Install the latest stable release:

python -m pip install fpstreams

Prebuilt wheels include the native extension. Building from source requires Rust 1.85 or newer.

The async extra installs aiofiles for async file adapters. Core async pipelines do not require an extra. Install other adapters only when needed:

python -m pip install "fpstreams[async]"   # aiofiles for async file adapters
python -m pip install "fpstreams[arrow]"   # PyArrow and Parquet
python -m pip install "fpstreams[data]"    # NumPy, pandas, and PyArrow
python -m pip install "fpstreams[polars]"  # Polars and PyArrow

Editor support

fpstreams includes inline type annotations, docstrings, and the py.typed marker. Select the Python interpreter where fpstreams is installed in your editor to get method completions, parameter hints, and hover documentation.

Version 2.2.1 fixes rows factory completion and keeps core iterable element types when optional data adapters are absent. For example, flow([1, 2, 3]) is inferred as Flow[int].

Quick start

This example filters paid orders, groups them by region, and reports both the number of orders and their revenue:

from fpstreams import agg, col, flow

orders = [
    {"region": "eu", "status": "paid", "amount": 24},
    {"region": "us", "status": "paid", "amount": 20},
    {"region": "eu", "status": "cancelled", "amount": 99},
    {"region": "eu", "status": "paid", "amount": 24},
]

result = (
    flow(orders)
    .filter(col("status") == "paid")
    .group_by("region")
    .aggregate(
        orders=agg.count(),
        revenue=agg.sum("amount"),
    )
    .sort_by("region")
    .to_list()
)

print(result)
# [{'region': 'eu', 'orders': 2, 'revenue': 48},
#  {'region': 'us', 'orders': 1, 'revenue': 20}]

When to use fpstreams

For a short, local transformation, a comprehension is often enough. fpstreams is useful when a pipeline needs grouped aggregations, early termination, resource cleanup, or async work with a concurrency limit. Small pipelines may be slower because planning and dispatch add overhead.

Lazy execution does not mean every operation uses constant memory. Sorting, grouping, and joins need additional state; use the documented limits and spill options when the input may not fit in memory.

fpstreams 2 replaces the v1 implementation and retains the compatibility aliases listed below.

What is in v2

• Flow[T]: the primary synchronous entry point for lazy value and record pipelines, including retained tabular sources.
• AsyncFlow[T]: asynchronous transforms with bounded concurrency, ordering, timeouts, merging, debouncing, and cleanup of tasks created by the pipeline.
• Rows[T]: a record view for expressions, joins, grouping, reshaping, and I/O.
• Pairs[K, V]: key/value transforms and per-key collection or aggregation.
• Collector and Aggregator: single-pass reductions, including named multi-aggregation.
• Option and Result: typed value and error containers.
• Automatic execution planning: fused Python loops, Arrow-native prefixes, native Rust kernels for supported scalar plans and guarded relational subpaths, and hybrid execution when only part of a plan is native.
• Configured memory and output limits for external sort and partitioned joins/grouping.

Version 2.1 adds execution reports, direct record operations from flow(), standard Arrow/dataframe protocol routing, retained NumPy execution, and new async queue, prefetch, window, and numeric terminal APIs. See the changelog for the release summary.

Release testing covers standard CPython 3.11 through 3.14. Free-threaded CPython 3.14t is exercised by an experimental, non-blocking job that builds the native extension; it is not currently a release-wheel target, and unsupported fast paths fall back conservatively.

Value pipelines

Pipelines are lazy. Transformations build a plan; terminal operations such as to_list(), aggregate(), first(), and count() execute it.

from fpstreams import flow, item

squares = (
    flow(range(1, 10))
    .filter(item % 2 == 0)  # Keep even values.
    .map(item * item)  # Square each value.
    .take(3)  # Stop after three results.
    .to_list()
)

assert squares == [4, 16, 36]

The placeholder expression above is equivalent to two lambdas, but it can also be compiled by the native engine when the complete plan is supported.

Named single-pass aggregation

from fpstreams import agg, flow

summary = flow([1, 2, 3, 4]).aggregate(
    count=agg.count(),
    total=agg.sum(),
    mean=agg.mean(),
)

assert summary == {"count": 4, "total": 10, "mean": 2.5}

All named aggregations share one traversal of the source.

Record pipelines

flow() is the main synchronous entry point for both values and records. Record operations with no conflicting Flow meaning, such as with_columns() and group_by(), enter a Rows view automatically. Records can be dictionaries, dataclasses, named tuples, or objects with attributes.

from fpstreams import agg, col, flow

orders = [
    {"region": "eu", "status": "paid", "price": 12, "quantity": 2},
    {"region": "us", "status": "paid", "price": 20, "quantity": 1},
    {"region": "eu", "status": "cancelled", "price": 99, "quantity": 1},
    {"region": "eu", "status": "paid", "price": 8, "quantity": 3},
]

revenue = (
    flow(orders)
    .filter(col("status") == "paid")
    .with_columns(revenue=col("price") * col("quantity"))
    .group_by("region")
    .aggregate(
        orders=agg.count(),
        revenue=agg.sum("revenue"),
    )
    .sort_by("region")
    .to_list()
)

assert revenue == [
    {"region": "eu", "orders": 2, "revenue": 48},
    {"region": "us", "orders": 1, "revenue": 20},
]

Four relation-building names already have general Flow meanings. Call .rows() first when you intend these relational versions:

Flow call	Flow meaning	Explicit Rows call	Rows meaning
drop(count)	Skip leading items	rows().drop(*columns)	Remove record fields
join(separator)	Join string representations	rows().join(other, ...)	Relational join
aggregate(...)	Execute and return a dictionary	rows().aggregate(...)	Build a lazy one-row relation
where(predicate)	Alias of filter	rows().where(predicate, **equalities)	Filter records, including field equalities

Flow also keeps its own output methods. Enter .rows() before same-named methods when you need Rows-specific options, for example to_csv(fieldnames=...) or to_pandas(batch_size=..., schema=...).

flow(source) automatically retains concrete PyArrow tables, batches and readers, pandas DataFrames, Polars frames, and standard __arrow_c_stream__/__dataframe__ providers. Arrow takes priority when a custom object implements both protocols, and a pandas index is not emitted as a record column. Explicit Flow factories cover Arrow, equal-length named columns, NumPy arrays, dataframe, Polars, typed CSV, and Parquet inputs. The rows namespace continues to provide compatibility CSV, JSONL, SQLite, DB-API, and record-oriented output methods. CSV and JSONL inputs accept paths, caller-owned open handles, or replayable opener functions:

from fpstreams import col, flow

active = (
    flow.from_parquet("accounts.parquet", columns=["id", "status", "balance"])
    .filter(col("status") == "active")
    .select("id", "balance")
)

active.to_parquet("active-accounts.parquet")

Asynchronous pipelines

map_async accepts synchronous or asynchronous callables. concurrency bounds in-flight tasks, while ordered=True preserves input order.

The core async API needs no optional extra. Install fpstreams[async] when you also need the aiofiles-based async file adapters.

import asyncio

from fpstreams import aflow


async def fetch(value: int) -> int:
    await asyncio.sleep(0.01)
    return value * 10


async def main() -> None:
    result = await (
        aflow([1, 2, 3, 4])
        .map_async(fetch, concurrency=2, ordered=True)
        .filter(lambda value: value >= 20)
        .to_list()
    )
    print(result)
    # [20, 30, 40]


asyncio.run(main())

Async iterators consumed by a pipeline are closed, and tasks created by the pipeline are cancelled, when it finishes, errors, times out, or short-circuits.

Key/value pipelines

from fpstreams import agg, pairs

totals = pairs([("a", 2), ("b", 5), ("a", 3)]).aggregate_values(
    total=agg.sum(),
    average=agg.mean(),
)

assert totals == {
    "a": {"total": 5, "average": 2.5},
    "b": {"total": 5, "average": 5.0},
}

Execution engines

The default auto engine chooses Python, Rust, Arrow, NumPy, or a combination for each supported plan. Relational plans may use native shortcuts while keeping a Python fallback. Pass the terminal you intend to call to explain() so it can include that terminal in its planning decisions:

from fpstreams import flow, item

pipeline = flow([1, 2, 3])
plan = pipeline.explain(terminal="count").to_dict()

assert plan["selected_engine"] == "python"
assert plan["complexity"] == "O(1)"
assert plan["data_movement"] == {
    "scans_source": False,
    "copies_source": False,
    "materializes": False,
}

You can request an engine explicitly when testing parity or diagnosing a plan:

python_result = pipeline.with_engine("python").to_list()
native_result = pipeline.with_engine("native").to_list()

Use run_with_report() to get a terminal's value, recorded route, and query-owned resource counts in one execution. Pairs supports terminals such as flow([("a", 1)]).pairs().run_with_report("group_values"). Version 2.4.0 also includes up to 64 child plan and route observations in report.events. See the execution report guide for supported terminals and the limits of route reporting.

In 2.4.0, report.to_dict() exports versioned metadata with metric units, resource scope, and event limits. It excludes the terminal value and does not rerun the query. See the metadata export guide for JSON storage and caller-managed OpenLineage integration. Report export requires 2.4.0 or newer.

Version 2.4.0 also includes BoundedExecutor for a shared submission limit across queries and FileCheckpoint for verified file-partition recovery. See the resource controls and file-partition recipe for ownership, memory, and retry semantics. These additions require 2.4.0 or newer.

In 2.4.0, flow.unfold / aflow.unfold generate values from explicit state transitions. Flow.map_accum and AsyncFlow.map_accum update state while mapping each input to one output. See the state examples for termination, tuple ordering, and mutable-state ownership. These APIs require 2.4.0 or newer.

A forced native plan raises NativeUnsupportedError if its complete types or operations cannot run natively. In particular, the presence of an internal native relational specialization does not make the complete relation eligible for with_engine("native"). An unsupported forced relational plan fails before claiming its one-shot sources; auto selects a legal fallback path.

Identity list and tuple materialization stays in Python. Large, supported integer sums and numeric range reductions may use Rust in auto mode. count() uses a known exact size in O(1) when no operation changes cardinality and the source is safely reiterable. Use explain(terminal=...) to inspect the planned route.

Resource and file-safety controls

For data larger than memory, use external_sort(..., buffer_size=...), a partitioned join, or spilled grouping. Spill processing is bounded by its configuration; it is not a promise that every skewed or many-to-many input will finish. The defaults are 100,000 rows and 64 MiB per partition, 100,000 matches per key, 1,000,000 output rows, and three repartition levels.

from fpstreams import SpillLimits, rows

limits = SpillLimits(max_output_rows=250_000, max_matches_per_key=10_000)
result = rows(left).join(right, on="id", partitions=32, limits=limits)
grouped = rows(records).group_by("account_id").spill(partitions=32, limits=limits)

Exceeding a configured partition, match, state, record, or output limit raises BufferLimitError and cleans up temporary files.

CSV output is raw by default for machine interchange. For values supplied by untrusted users and later opened in spreadsheet software, pass spreadsheet_safe=True; strings beginning (after whitespace) with =, +, -, or @ are prefixed with a single quote. JSONL reads accept at most 8 MiB per physical record by default and reject larger lines before JSON parsing; binary inputs are checked before decoding as well. Pass max_record_bytes=None only for trusted local input when an unlimited record is intentional.

Source and resource semantics

• Reiterable inputs such as lists can execute more than once.
• Iterators and async iterators are one-shot and raise FlowConsumedError on a second execution.
• Ordinary generators are not opened or sampled when flow() is constructed.
• Generic __dataframe__ conversion and Polars LazyFrame collection remain deferred until execution. A custom __arrow_c_stream__ provider is imported once during construction and treated as one-shot; a PyArrow RecordBatchReader is also one-shot.
• flow.defer(factory) opens a fresh source for every execution.
• Terminal operations close owned iterators, database cursors, temporary files, and asynchronous tasks, including on errors and early termination.

v1 compatibility

Stream remains an alias of Flow, AsyncStream remains an alias of AsyncFlow, and ParallelStream remains an alias of Flow to ease imports. New synchronous code should start with flow; use .rows() or the rows factory namespace when an explicit relational view or record-specific adapter is required. Use aflow and pairs for their respective domains.

v2 breaks parts of the v1 API. The standalone core and ParallelStream implementations are gone. ParallelStream remains an alias, and Flow.parallel() remains as a compatibility strategy for following maps. New code can call map_parallel() directly or use map_async() for asynchronous work.

Development

uv sync --extra arrow --extra data --extra polars \
  --group build --group test --group lint --group type --group docs

uv run maturin develop --release
uv run pytest -W error --cov=src/fpstreams --cov-branch --cov-report=term-missing
uv run coverage report
uv run ruff check src tests scripts benchmarks benchmark.py
uv run ruff format --check src tests scripts benchmarks benchmark.py
uv run mypy src/fpstreams
cargo test --manifest-path rust/Cargo.toml
uv run python scripts/build_browser_wheel.py
uv run mkdocs build --strict -f fpstreams/mkdocs.yml

Run ./run_benchmark.sh for the configurable benchmark matrix. Its settings are kept at the top of the script, including input size, repeats, presets, case filters, and optional JSON output. Competitive mode checks equivalent tasks against Python, NumPy, and pandas and prints the percentage difference.

The source tree is organized by domain under src/fpstreams/: streams, planning, execution, collecting, tabular, expressions, and primitives. Small top-level modules are compatibility facades, not duplicate implementations. The Chinese code-reading guide walks through the main sync, relational, spill, native, and async execution paths in those domains.

Contributions are welcome. The contributing guide contains the development setup, validation commands, and performance-change expectations. Use the issue forms for reproducible bugs and scoped proposals; report security problems through GitHub's private vulnerability reporting flow.

License

MIT. See the license.