v1.1.1

MARIO v1.1.1

This bugfix release rebuilds the GTAP Power MRIO parser around a record-based
assembly engine shared by the CSV and GDX backends, fixing crashes and silent
data losses on recent GTAP exports, and adds an opt-in structured row layout
for the GTAP factor and satellite blocks.

GTAP parser rewrite

• The parser now maps every record onto integer axis positions and accumulates
  values directly into dense blocks, so the build cost scales with the number
  of records instead of the size of the cartesian label space. A full GTAP
  2023 Power MRIO CSV bundle (~7 GB, ~140M records) parses end-to-end in
  about 80–90 seconds, where the previous implementation ran for 20+ minutes
  before failing.
• CSV reading prefers the multithreaded pyarrow engine (with a silent
  fallback to the default engine) and logs per-file and per-block progress.

Fixes

• Duplicated record keys no longer crash the CSV path. Recent GTAP CSV
  exports flatten several GDX symbols into one file (e.g. combustion and
  non-combustion emissions), so the same key can legitimately appear more than
  once: duplicated keys are now summed instead of raising
  ValueError: cannot handle a non-unique multi-index.
• Output- and value-added-based emission accounts are captured. The CSV
  export encodes them with the SRC="TOT" placeholder; they were silently
  dropped — about one third of the emission record mass in the GTAP 2023
  bundle — and now land in the domestic satellite rows of the destination
  region (e.g. EMI_CH4_dms_QO), mirroring how the GDX bundles store them.
• No more spurious EMI_0_* satellite rows produced by the fill step of
  the previous CSV implementation (which also raised a TypeError on the
  pandas 3 string dtype).
• E and EY expose one dense, aligned satellite row set on both
  backends, instead of the data-dependent asymmetric row sets of the previous
  CSV path.
• Any satellite record mass the parser cannot attribute is now reported with a
  parser warning instead of being silently dropped.

New: optional structured row layouts for GTAP

• parse_gtap gained an opt-in matrix_layouts argument:

  db = mario.parse_gtap(
      path="path/to/bundle",
      input_format="auto",
      matrix_layouts={"V": ("Region", "Sector"), "E": ("Region", "Sector")},
  )

  Factor and satellite rows are exposed as a semantic
  (Region, Sector, item) MultiIndex instead of the historical flat string
  names (MTAX_AUS_GRO becomes ("AUS", "GRO", "MTAX")), and the matching
  block specifications are registered on the database. Values are identical to
  the flat parse — only the index representation changes.

• Row families without a region or sector of their own (PTAX, VAAD/VTAX
  endowments, ETAX, DTAX/ITAX, domestic satellite accounts) carry the
  "TOTAL" sentinel on the levels that do not apply. The sentinel is not
  part of the Region/Sector sets and passes through aggregation untouched, so
  no extra entry is needed in aggregation mapping files.

• The default (matrix_layouts=None) keeps the historical flat rows, so
  existing workflows are unaffected.

Notes

• The GDX backend requires the GAMS Python API; the pip package
  gamsapi[transfer] is sufficient to read GDX bundles.
• GTAP databases parsed with earlier MARIO versions differ from v1.1.1 output
  by design: the emission block now includes the previously dropped output-
  and value-added-based accounts, and E/EY share one aligned row set.

Compatibility

• Requires Python 3.11–3.13.
• No API breaking changes since v1.1.0; the matrix_layouts argument is
  additive and off by default.

Full changelog: v1.1.0...v1.1.1