pypff - Python PANOSETI File Format I/O Library

A high-performance Python package for reading and analyzing data files generated by PANOSETI (PANOSETI File Format - PFF).

Features

• Streaming by default: iter_batches(size=256) and __iter__ yield zero-copy views without materializing the full sequence into RAM — safe for Jupyter and HPC pipelines alike.
• Distributed chunked reads: iter_byte_range(file_idx, byte_start, byte_end) lets Dask/Nextflow workers parse frame-aligned byte ranges in parallel without coordination.
• Zero-copy random access: seq[i] returns a strided mmap view; slicing and read_images(indices) use a sort + inverse-permutation for disk locality.
• Single-pass metadata: get_metadata_arrays(keys) extracts any number of header fields in one np.frombuffer pass per file via a composite NumPy structured dtype. Supports virtual unix_t_ns key.
• Nanosecond-precise timestamps: timestamps() returns int64 ns (no float precision loss); timestamps(as_datetime=True) returns a zero-copy datetime64[ns] view for matplotlib and pandas.
• PFF → Zarr v3 conversion: pypff[zarr] optional extra converts any .pffd run to Zarr v3 stores readable by xarray, dask, numpy, TensorStore, and Julia — lossless, compressed, HPC/ML-ready. See Zarr v3 spec.
• Bounded resources: LRU mmap handle cache (default 16 files); PFFSequence is a context manager.
• Multiprocessing-safe: pickle-compatible — file handles are dropped on serialisation and lazily reopened in workers.
• Pydantic validation: strict schema validation for all PFF headers and PANOSETI config files.
• Run discovery: PanosetiRun lazily scans a .pffd directory and exposes typed configs, housekeeping, and data products.

Installation

The package uses uv for dependency management.

# User install: pip
pip install pypff

# User install: add to existing uv environment:
uv add pypff

# Development Install options
cd pypff
uv sync                   # core library
uv sync --extra zarr      # + Zarr v3 conversion (zarr-python, xarray, dask)

Quick Start

Run discovery

from pypff import PanosetiRun

run = PanosetiRun("path/to/run_directory.pffd")
run.show()                          # pretty-print structure

print(run.list_products())          # ['dp_img16.bpp_2.module_1', ...]
seq = run.get_product("dp_img16.bpp_2.module_1")

obs_cfg  = run.get_config("obs_config")   # returns Pydantic model
data_cfg = run.get_config("data_config")
hk       = run.get_hk()                   # dict[device, dict[field, np.ndarray]]

Streaming (memory-bounded)

# Iterate frame-by-frame (zero-copy views into mmap)
for img in seq:
    process(img)

# Batch iteration — never holds more than one batch in RAM
for batch in seq.iter_batches(size=256):
    batch  # shape (256, H, W), dtype matches file

# With timestamps and headers in one pass
for batch, ts in seq.iter_batches(size=256, with_timestamps=True):
    ts  # int64 nanoseconds, shape (256,)

# Distributed byte-range reads (Dask / Nextflow workers)
for batch in seq.iter_byte_range(file_idx=0, byte_start=0, byte_end=size):
    ...

Random access and slicing

img   = seq[42]            # zero-copy view, shape (H, W)
imgs  = seq[0:100:2]       # every other frame, shape (50, H, W)
imgs  = seq.read_images(np.array([5, 0, 3]))   # unsorted — sorted internally for locality
imgs  = seq.read_images_range(start=0, count=500)

Timestamps — always int64 nanoseconds

# Full cached array — int64 ns, no float precision loss
ts = seq.timestamps()                      # np.ndarray[int64]

# Zero-copy datetime64[ns] view for matplotlib / pandas
ts_dt = seq.timestamps(as_datetime=True)   # np.ndarray[datetime64[ns]]

# Indexed subset
ts_sub = seq.timestamps(indices=np.arange(0, len(seq), 100))

# Single frame
t_ns = seq.timestamp_at(42)               # int, nanoseconds

# Time-based navigation
idx = seq.seek_time(t_ns + 1_000_000_000)  # 1 s later

# Arithmetic rule: subtract epoch in integer space before dividing
rel_s = (ts - ts[0]) / 1e9               # CORRECT — small values, no precision loss
# ts / 1e9 - t0_s                        # WRONG  — loses ns precision at ~1.7e9 s

Metadata extraction (single pass)

# One np.frombuffer call per file regardless of number of keys
meta = seq.get_metadata_arrays(["pkt_num", "tv_sec", "unix_t_ns"])
meta["unix_t_ns"]   # int64 ns, same as timestamps()
meta["pkt_num"]     # int64

# All fields at once
all_meta = seq.get_all_metadata()

Headers and Pydantic validation

# Fast path — raw dict, no Pydantic overhead
header, img = seq.get_frame(0)
header["pkt_num"]                 # quabo file
header["quabo_0"]["pkt_tai"]      # module file

# Validated path — full Pydantic model
from pypff.models import ModuleHeader
hdr, img = seq.get_frame_validated(0)
assert isinstance(hdr, ModuleHeader)

Context manager and multiprocessing

# Deterministic handle cleanup
with run.get_product("dp_ph256.bpp_2.module_254") as seq:
    data = seq.read_images_range(0, 1000)

# PFFSequence is pickle-safe — handles are dropped on serialisation
import concurrent.futures
with concurrent.futures.ProcessPoolExecutor() as ex:
    futures = [ex.submit(lambda s, i: s[i].sum(), seq, i) for i in range(len(seq))]

PFF → Zarr v3 conversion (pypff[zarr])

Convert a .pffd observation run to Zarr v3 stores readable by xarray, dask, numpy, TensorStore, Julia, and Rust. See docs/zarr_v3_spec.md for the full layout specification.

Write

from pypff.io2 import PanosetiRun
from pypff.zarr import convert_run

run = PanosetiRun("path/to/obs.pffd")
stores = convert_run(run, "output/L0_zarr")
# → one .zarr per (data_product, module)
# → one .panoseti-meta/ sidecar bundle (configs, logs, hk.pff)

Or via the CLI:

uv run pypff zarr path/to/obs.pffd output/L0_zarr

Read

from pypff.zarr import PanosetiZarrRun
import xarray as xr

# High-level wrapper (mirrors PanosetiRun)
zrun = PanosetiZarrRun("output/L0_zarr")
store = zrun.get_product("dp_ph256.bpp_2.module_254")
ts    = store.timestamps()           # int64 ns array
ds    = store.to_dataset()           # xarray.Dataset
cfgs  = zrun.configs                 # parsed run configs

# Or open directly with xarray — all arrays visible as variables
ds = xr.open_zarr(str(stores[0]), consolidated=False)
# ds.images      (T, H, W)   int16 / uint16
# ds.unix_t_ns   (T,)        int64  — nanosecond timestamps
# ds.pkt_num     (T,)        uint32 ─┐ per-frame header fields
# ds.pkt_nsec    (T,)        uint32  │ (single-level: ph256)
# ds.quabo_num   (T,)        uint8  ─┘
# ds.quabo_0_pkt_num  (T,)   uint32 ─┐ module-level headers
# …                                  ─┘ (img16, ph1024)

Zarr store layout

All arrays live at the root of each store (no sub-groups), so xr.open_zarr(store) surfaces every variable — images, timestamps, and all header fields — as a single Dataset aligned on the shared time dimension. Logical grouping of headers is expressed via header_fields and quabo_fields root attributes. Full specification: docs/zarr_v3_spec.md.

Testing

Run the test suite via the built-in CLI:

uv run pypff test all

The test suite includes:

• Tier 1 (Unit): Basic logic and timing tests.
• Tier 2 (Logic): Higher-level I/O, slicing, and concurrency tests.
• Legacy Integration: The original pypff test suite using provided sample data.

Dockerized CI

Build the CI environment:

docker build -t pypff-ci -f src/ci/Dockerfile.ci .