icclim 7.3.4

icclim 7.3.4

on Python PyPI

date: 2026-11-08

• [fix] Support xclim 1.0 (PR #579). xclim 1.0.0 removed xclim.indices
  (moved to xclim.compute), renamed the op argument of
  growing_season_length to condition, and returns a Dataset from
  its indicators. icclim 7.3.3 and earlier declare xclim>=0.45.0 without
  an upper bound, so a fresh install picks up xclim 1.0.0. With it, CDD
  raised ModuleNotFoundError: No module named 'xclim.indices' and GSL
  raised TypeError: select_time() got an unexpected keyword argument 'op'
  (run on 7.3.3 with xclim 1.0.0). The other 7.x releases were not tested with
  xclim 1.0.0. The test suite passes with xclim 0.60.0 and xclim 1.0.0.
• [fix] difference_of_means no longer returns NaN everywhere when
  sampling_method="groupby_ref_and_resample_study" is used with a
  reference given as a separate input that covers another period (issue #571,
  PR #572). The completeness check combined the masks of the study and of the
  reference, so with disjoint periods every output period was masked, without
  any error or warning. Now only the study variable decides which output
  periods are complete. Affected: 7.2.0, 7.3.0, 7.3.1, 7.3.2 and 7.3.3 (run on
  a study 2002-2003 against a reference 2000-2001; 7.1.10 returns the expected
  values). Results computed with these releases for such a call should be
  recomputed; allow_missing_periods=True restored the values.
• [fix] Threshold queries that cannot be tokenized, such as
  "Coco l'asticot", now raise InvalidIcclimArgumentError on
  Python 3.13. The tokenizer of Python 3.13 makes pint raise
  tokenize.TokenError, which was not caught (PR #570).
• [perf] Keep the Dask chunks of caller-provided arrays when they are
  reasonably large, instead of rechunking with "auto" (PR #518). Arrays
  whose largest chunk is smaller than 1/8 of array.chunk-size (16 MiB with
  the default), and eager arrays, are still chunked with "auto": with
  one chunk per daily file the preserved chunks made some indices 3 to 7 times
  slower. Results are unchanged.
• [enh] Add opt-in NetCDF4 compression and stored chunk encoding through
  ICCLIM_OUTPUT_COMPRESSION_LEVEL and ICCLIM_OUTPUT_CHUNKS. The default
  output is unchanged (PR #518).
• [perf] Compute the per-period reducers with vectorized reductions instead of
  Python loops over the periods: maximum and minimum with date_event,
  count with date_event and consecutive occurrences with dates (PR #573).
  On the ACCESS-CM2 benchmark the date-event indices go from about 2 hours to
  about 1 minute, with identical results. date_event with txx and
  tnn now also works on non-standard (cftime) calendars. The dated
  max_consecutive_occurrence on cftime calendars still fails (issue #574).
• [enh] New option split_runs_at_period_boundary for max_consecutive_occurrence
  and the ECA&D indices CDD, CWD, CFD and CSU (issue #543, PRs #547 and #585).
  No change by default (False): a run of consecutive days that crosses the end
  of an output period is still counted whole in the period where it starts
  (run_index="first", the default) or where it ends (run_index="last"), and
  the other period gets no credit for the days of that run, so its value can be 0.
  With split_runs_at_period_boundary=True the runs are cut at the boundary of
  every output period and each period gets the longest run found inside it. For
  daily data from 2001 to 2002 with frost days from 2001-10-01 to 2002-04-30 and a
  yearly output, CFD is [212, 0] by default ([0, 212] with
  run_index="last") and [92, 120] with the option. The output dimension
  order is the same with and without the option. sum_of_spell_lengths and the
  indices built on it (WSDI, CSDI) are not concerned by the option.
• [doc] Explain in the completeness how-to why counting indices returned 0 over
  cells without any data before 7.2 (a comparison with NaN is never true), why
  they now return NaN, and how to keep the old behaviour with
  completeness="none" (PR #575).
• [test] Add an isolated non-bootstrap performance matrix reporting graph
  construction, task count, scheduler calls, wall time and peak RSS across
  workloads, input layouts and chunking policies.
• [test] Add matching Kraken real-data workloads and a slow synthetic-cube
  memory regression test.
• [test] Add metamorphic property tests (PR #578).