h5coro-hidefix

Compiled companion to h5coro: a tiny pyo3 binding over the hidefix crate (consumed from crates.io — no hidefix source lives in this tree). hidefix reads chunked HDF5 files through a pre-built chunk index instead of the HDF5 library's B-tree walk, decoding chunks in parallel (gzip + shuffle), which benchmarked ~6× faster than pure-Python h5coro on ICESat-2 ATL03 workloads — and ~15–20× once the index is built ahead of time and reloaded.

This package exists because the upstream hidefix Python binding cannot save, load, or enumerate its index, and squeezes length-1 dimensions on read. This binding fixes all three while keeping the surface minimal:

• Index(path) — build an index from a local HDF5 file (~0.2 s / ~0.6 MB for a ~2 GB ATL03 granule).

• Index.save(path) / Index.load(path, source=None) — persist / restore the index (bincode via hidefix's public serde impls) so reads need zero HDF5 metadata I/O. source points reads at the data file when it lives somewhere else than at index-build time.

• Index.datasets() — full dataset paths, recursing groups.

• Index.chunks(dataset) — the chunk table as numpy arrays (addrs, sizes, offsets): file byte offset, stored (compressed) byte count, and dataspace coordinates per chunk.

• Index.read(dataset, start=None, end=None) — rows [start, end) along dim 0 (remaining dims in full) as a numpy array with native dtype and the exact request shape. Never squeezes: a (1, 5) request returns (1, 5), byte-identical with h5py_dataset[start:end]. The GIL is released for the duration of the read.

• Index.read_plan(dataset, start=None, end=None) / Index.read_from_buffers(dataset, buffers, start=None, end=None) — the object-store read path: the plan lists the chunks a read needs (the same (addrs, sizes, offsets) triple as chunks(), restricted to the row range, ascending dataspace offset); the caller fetches those byte ranges itself and hands the raw stored bytes back for decode + assembly, byte-identical to read(). No file or network access happens inside the binding.

• Index.from_chunks(source, datasets) — construct a decode-capable index from an external chunk manifest (e.g. rows from a parquet sidecar store): no granule file access, no stored bincode. Index.filters(dataset) ({"gzip": int|None, "shuffle": bool, "byte_order": ...}) completes the extraction side, so Index(path) → getters → from_chunks round-trips byte-identically.

• zagg sidecar backend (h5coro_hidefix.zagg_backend.SidecarIndex) — registered under the zagg.index_backends entry-point group, so a zagg environment with this wheel installed discovers it automatically:

  data_source:
    index:
      backend: sidecar
      store: s3://bucket/zagg-index/ATL03/007/
      on_miss: fallback      # fallback | error | build
  

  Selection uses zagg's shared planned route; addressing reconstructs the granule's index from <store>/<granule_id>.parquet (zagg's inline write-back manifest schema) via from_chunks, fetches exactly the needed chunk ranges through the worker's own credentialed h5coro driver, and decodes with read_from_buffers. Written against the zagg virtual-index protocol at englacial/zagg PR #163 head 87b941e (see the module docstring). Importing h5coro_hidefix alone never pulls zagg/pandas — the backend module is only imported by zagg's entry-point discovery.

Design discussion: englacial/zagg#155, englacial/zagg#160 (parquet-primary sidecar store).

Install

pip install h5coro-hidefix

Or from source (needs a Rust toolchain and cmake — the bundled static libhdf5 is built at compile time; no system HDF5 required at build or run time):

pip install .

Quickstart

import h5coro_hidefix as hx

idx = hx.Index("ATL03_20190105163308_01260202_007_01.h5")   # ~0.2 s

idx.datasets()                        # ['/gt1l/heights/h_ph', ...]
idx.shape("/gt1l/heights/signal_conf_ph")   # (n_photons, 5)
idx.dtype("/gt1l/heights/h_ph")             # 'float64'

# persist the index; later reads skip all HDF5 metadata I/O
idx.save("granule.idx")
idx = hx.Index.load("granule.idx", source="ATL03_...h5")

arr = idx.read("/gt1l/heights/h_ph", 1_000_000, 2_000_000)  # float64 (1000000,)
one = idx.read("/gt1l/heights/signal_conf_ph", 7, 8)        # int8 (1, 5) -- not (5,)

addrs, sizes, offsets = idx.chunks("/gt1l/heights/h_ph")    # uint64 arrays

Obstore-fed reads (the Lambda worker flow)

Workers never open the HDF5 file: indices are built once at catalog time, and at read time the worker loads the index, asks which byte ranges a row-slice needs, fetches those ranges itself (obstore/boto3 ranged GETs), and hands the raw bytes back for decode:

import obstore
import h5coro_hidefix as hx

idx = hx.Index.load("granule.idx")          # ~1 ms; zero HDF5 metadata I/O
name = "/gt1l/heights/h_ph"

addrs, sizes, _ = idx.read_plan(name, row0, row1)
buffers = obstore.get_ranges(               # caller owns the fetch policy:
    store, "ATL03_...h5",                   # coalescing, concurrency, retries
    starts=addrs.tolist(),
    ends=(addrs + sizes).tolist(),
)
arr = idx.read_from_buffers(name, buffers, row0, row1)
# arr is byte-identical to idx.read(name, row0, row1) on a local copy

buffers must line up 1:1, in order, with read_plan's chunks; any bytes-like objects work (bytes is zero-copy, others are copied once). ValueError is raised on a wrong buffer count or a buffer whose length differs from the chunk's stored size; the GIL is released during decode. This path removes any dependency on hidefix-side S3 support.

Manifest-built indices: from_chunks (parquet-primary store)

When the chunk manifest lives in a durable store (parquet sidecars — see englacial/zagg#160), the index is reconstructed from it on demand; the granule file is never opened and no bincode is ever stored:

# caller side: read the manifest (pyarrow / obstore) -- this package stays
# numpy-only at runtime, so parquet decoding happens outside it
idx = hx.Index.from_chunks("s3-key-or-local-path.h5", {
    "/gt1l/heights/h_ph": dict(
        dtype="<f8",              # numpy dtype str; byte order honored
        shape=(23_692_855,),
        chunk_shape=(100_000,),
        gzip=6,                   # deflate level, bool, or None (see below)
        shuffle=True,
        addrs=addrs,              # u64[k]   chunk byte offsets in the file
        sizes=sizes,              # u64[k]   stored (compressed) byte counts
        offsets=offsets,          # u64[k, ndim] dataspace chunk coordinates
        filter_mask=masks,        # optional; must be all zero (see below)
    ),
    ...
})
arr = idx.read_from_buffers(name, buffers, row0, row1)  # as above

Chunk rows need not be pre-sorted. The result is a real index — read, read_plan, read_from_buffers, save, chunks all behave exactly as if it had been built from the file.

Metadata contract for extractors (what a manifest must carry, per dataset): dtype, shape, chunk_shape, gzip (level int | bool | null — manifests that cannot see the deflate level, e.g. h5coro's metadata parse, may emit a boolean: decode only checks presence, so True records a placeholder level), shuffle, plus the chunk table (addrs, sizes, and per-chunk dataspace offsets — for 1-D data the offset equals the element start index; for N-D it must be stored explicitly). All of it is available from a real index via datasets(), shape(), chunk_shape(), dtype(), filters() and chunks().

Per-chunk filter masks are unrepresentable: hidefix models filters at dataset level (shuffle, gzip), and its chunk records carry only (addr, size, offset). A nonzero HDF5 filter_mask (bit i = filter i skipped for that chunk) therefore raises ValueError in from_chunks. ATL03 masks are all zero (verified in englacial/zagg PR #159).

Scope and limitations

• The binding itself performs only local-file I/O (read()); object-store reads are the caller's job via read_plan/read_from_buffers above.
• Filters: gzip (deflate) + shuffle + byte-order — full coverage for ATL03 and most NASA Earthdata HDF5. Other filters (szip, lzf, scaleoffset) fail at index time.
• Hyperslabs are row ranges on dimension 0; remaining dimensions are read in full.
• The serialized index format is hidefix's bincode encoding of its Rust types — treat it as an opaque cache keyed by (file, hidefix version), not as a stable interchange format.

Licensing

The binding code in this repository is MIT. Published wheels statically embed third-party components; their license texts ship inside the wheel (*.dist-info/licenses/):

• hidefix — MIT, Copyright 2023 Gaute Hope (LICENSE-hidefix). The earlier upstream metadata ambiguity is resolved: MIT was confirmed in gauteh/hidefix#48 and the stale Cargo.toml field fixed in gauteh/hidefix#49.
• HDF5, built and bundled by hdf5-metno-src — The HDF Group's BSD-style license (LICENSE-hdf5).

Remaining statically-linked Rust dependencies are MIT/Apache-2.0 dual-licensed or similarly permissive.

Development

python -m venv .venv && . .venv/bin/activate
pip install maturin pytest h5py numpy
maturin develop --release
pytest -v          # hermetic tests generate their own HDF5 files

A second, local-only test tier runs automatically when real ATL03 granules are present (default ~/ignore/zagg_neon_atl03_test_shard/granules, override with H5CORO_HIDEFIX_GRANULE_DIR).

CI builds wheels for manylinux x86_64 + aarch64 (static libhdf5, cmake installed in the manylinux image) and macOS arm64, abi3 (>=3.9), one wheel per platform. The publish job triggers on version tags and publishes to PyPI via Trusted Publishing.