Skip to content

v4.0.0

Choose a tag to compare

@dheerajchand dheerajchand released this 11 Oct 22:45
· 880 commits to main since this release
9c3463b

[4.0.0] - 2026-10-11

Major release. The public surface changes incompatibly (the data.world
connector and its exports are removed, and several packaging extras changed),
so this supersedes the never-published 3.24.0 in-repo version and bumps the
major per Semantic Versioning.

BREAKING

  • Removed the data.world connector (siege_utilities.analytics.datadotworld_connector
    and its exports: DataDotWorldConnector, get_datadotworld_connector,
    search_datadotworld_datasets, load_datadotworld_dataset,
    query_datadotworld_dataset, search_datasets, list_datasets,
    DATADOTWORLD_AVAILABLE). The upstream open-source datadotworld product was
    discontinued, so the connector is removed rather than repaired. The
    datadotworld dependency is dropped from the analytics extra. Code importing
    these symbols must remove those imports. (C15/C16, #1358.)

Added

  • Dual-source Census retrieval framework with provenance. A CensusDataSource
    protocol abstracts a source; ApiSource wraps the Census API and exposes an
    exhausted-timeout cause so the fallback policy recognizes it;
    DualSourceCensusClient retries the primary on configured errors (rate limit,
    timeout) up to a budget, then tries the secondary once, recording per-request
    Provenance (source, location, attempts, degraded). If both sources fail the
    last error propagates rather than returning an empty frame. AlternationPolicy
    is pure, validated configuration. (#1375.)
  • ACS table-based Summary File bulk source (BulkFileSource), the secondary for
    the dual-source client: streamed per-table caching, geography filtering, API
    variable names, numeric estimates and MOEs, sentinel normalization, and
    fallback provenance. Malformed records (width drift, embedded control bytes,
    duplicate or BOM-masked headers, quoted delimiters, non-finite tokens) raise
    rather than returning misaligned or NaN data. Geographic name enrichment is
    deferred; bulk NAME values are missing. (#1385.)
  • First-class yearless Census timeseries support. CensusCatalogPopulator.populate
    accepts year=None to discover a Census timeseries dataset, building
    {base}/timeseries/<program>/<dataset>/ variables and groups with no injected
    year; a _TIMESERIES_PATHS alias registry maps names like eits/resconst.
    CensusCatalogDataset.year becomes int or None and gains
    timeseries_variables and timeseries_groups, retained through the disk-cache
    round trip. Year-based discovery is unchanged and builds byte-identical URLs.
    (#1374.)
  • Python 3.13 is now a blocking CI gate (previously allowed to fail). The full
    Linux suite passes on 3.13. Python 3.14 is tested as a non-blocking CI gate
    while its recently released native stack stabilizes. (#1365.)

Changed

  • Capped pandas>=2.0.0,<3.0 across every extra that declares pandas (data,
    geo, reporting, analytics, survey, all; geo pins it explicitly because
    geopandas pulls pandas transitively). pandas 3.0 makes
    the default string dtype Arrow-backed, so .str operations dispatch into
    pyarrow.compute and segfault on 3.13 with the current native stack. The cap
    keeps strings on the stable backend; lift it once pandas 3.x + pyarrow string
    interop is verified green on the CI matrix. (#1365.)

Fixed

  • Docstring hygiene tool works on Python 3.14: it matches ast.Constant instead
    of the removed ast.Str alias, and renders rewritten source with stdlib
    ast.unparse rather than the unmaintained astor (which still reached for
    ast.Str when rendering f-strings and raised AttributeError on 3.14). The
    astor dependency is dropped from the dev extra (#1365).

  • EPSG:2163 round-trip property test compares the minimal angular longitude
    difference, so the antimeridian (lon 180 emitted as -180 by newer PROJ) is no
    longer read as a 360-degree drift (#1365).

  • Engine aggregation set honors advertised stddev/variance on the
    pandas-family engines (C17, #1359).

  • Boundary normalization raises instead of silently returning un-normalized data
    on crosswalk failure (C3, #1351).

  • Census sample datasets are flagged unavailable instead of advertised as working
    (C9/C10, #1360).

  • spark_to_geopandas rejects a mixed-CRS metadata column instead of silently
    unifying it (C6, #1354).

  • Snowflake connector is enabled on an installed SDK and its config-file factory
    works (C13/C14, #1357).

  • DuckDB engine query() returns an eager, materialized frame with stable
    snapshot and lifetime semantics, converts native GEOMETRY to WKB by column
    position for every result-producing statement (including DML RETURNING),
    loads the spatial extension only when a result carries geometry so ordinary
    SQL runs under a restricted configuration, preserves DML affected-row counts,
    and matches prepared-statement names on engine token boundaries with a binary
    collation (#1369).

  • Crosswalk intensive variables area-weight by true overlap area with
    real-tobler parity, exclude missing values per variable, treat zero-overlap
    sources as non-contributing instead of a false merge rejection, and areal
    interpolation leaves a no-valid-contributor target undefined rather than
    fabricating 0.0 (#1369).