v0.8.5

Add scipy as required dependency

See 0.8.4 release notes

v0.8.4

What's Changed

Another round of absolutely massive performance improvements for method="cohorts". This should bigly improve many Xarray groupby workloads (which use cohorts by default). Resampling in particular should be much better

Benchmarks improvements undersell the changes, since the core loop is approximately quadratic. Graph construction time for the example in https://xarray.dev/blog/flox with "cohorts" specified now drops from 30s to 3s 😱

| Before [15324a7a] <v0.8.3>   | After [666d45e9] <main>   |   Ratio | Benchmark (Parameter)                                   |
|------------------------------|---------------------------|---------|---------------------------------------------------------|

#### Larger cohorts, lower tasks
| 5180                         | 3600                      |    0.69 | cohorts.NWMMidwest.track_num_tasks                      |
| 4891                         | 3385                      |    0.69 | cohorts.NWMMidwest.track_num_tasks_optimized            |
| 505                          | 345                       |    0.68 | cohorts.NWMMidwest.track_num_layers                     |

#### Much faster algorithm for detecting cohorts, that should scale better.
| 3.19±0.07ms                  | 2.42±0.05ms               |    0.76 | cohorts.ERA5Google.time_find_group_cohorts              |
| 1.04±0.01ms                  | 782±70μs                  |    0.75 | cohorts.PerfectMonthly.time_find_group_cohorts          |
| 1.06±0.01ms                  | 781±70μs                  |    0.73 | cohorts.PerfectMonthlyRechunked.time_find_group_cohorts |
| 29.7±2ms                     | 12.6±0.9ms                |    0.43 | cohorts.ERA5DayOfYear.time_find_group_cohorts           |
| 7.76±1ms                     | 2.90±0.2ms                |    0.37 | cohorts.ERA5MonthHour.time_find_group_cohorts           |
| 8.17±0.8ms                   | 2.75±0.2ms                |    0.34 | cohorts.ERA5MonthHourRechunked.time_find_group_cohorts  |
| 242±5ms                      | 47.3±2ms                  |    0.2  | cohorts.NWMMidwest.time_find_group_cohorts              |
| 28.8±3ms                     | 4.11±0.3ms                |    0.14 | cohorts.ERA5DayOfYearRechunked.time_find_group_cohorts  |

##### Total time is not too different, we have some overhead in constructing the graphs
| 162±5ms                      | 144±9ms                   |    0.89 | cohorts.ERA5DayOfYearRechunked.time_graph_construct     |
| 20.7±0.2ms                   | 18.3±0.4ms                |    0.89 | cohorts.ERA5Google.time_graph_construct                 |
| 3.21±0.2ms                   | 2.40±0.04ms               |    0.75 | cohorts.PerfectMonthly.time_graph_construct             |
| 181±10ms                     | 129±10ms                  |    0.71 | cohorts.NWMMidwest.time_graph_construct                 |

Changes

• More cohorts speedups by @dcherian in #290
• typing fixes. by @dcherian in #292
• Use set containment instead of perfect subsets by @dcherian in #291

Full Changelog: v0.8.3...v0.8.4

v0.8.3

What's Changed

• Fix reordering of dataarray dimensions inside dataset by @eendebakpt in #289

New Contributors

• @eendebakpt made their first contribution in #289

Full Changelog: v0.8.2...v0.8.3

v0.8.2

Major performance improvements (yet again!) Thanks to @max-sixty for prompting these.

What's Changed

• Properly dispatch to numbagg when we can by @dcherian in #282
• Actually optimize out multiple "nanlen" by @dcherian in #283
• Set order='F' when raveling group_idx after broadcast by @dcherian in #286

Full Changelog: v0.8.1...v0.8.2

v0.8.1

Fix packaging of v0.8.0

See v0.8.0 release notes for all changes

v0.8.0

What's Changed

Major performance improvements!!!

1. Support numbagg throughengine="numbagg" for many common nan-skipping reductions in #72. Using numbagg appears to be a major speedup (2x-3x in general, 6X for nanmean). Special thanks to @max-sixty for major work on numbagg's grouped aggregations! Here are timings for reducing a 2D array along the last axis with ordered group labels.

   func	engine
   nansum	flox	70.3±0.2ms
   	numpy	122±0.2ms
   	numbagg	18.4±0.04ms
   nanmean	flox	144±0.4ms
   	numpy	196±0.5ms
   	numbagg	23.7±0.2ms
   nanmax	flox	93.4±0.8ms
   	numpy	953±2ms
   	numbagg	20.3±0.2ms
   count	flox	59.8±1ms
   	numpy	114±0.2ms
   	numbagg	29.3±0.1ms

2. Support engine=None in #266. This will

   • Use numbagg if available
   • If not, use flox if the group labels are sorted
   • Fallback to numpy otherwise.
     Thanks to @mathause for kicking off this work.

3. Significant speed to detecting "cohorts" of groups in #272

Other Major Changes

1. Test and support for python 3.12 (note numba does not support 3.12 yet)
2. Bump minimum numpy version to 1.22.
3. New Aggregations : Support quantile, median, mode with method="blockwise". by @dcherian in #269
4. Add multidimensional bins demo notebook by @dcherian in #203 . This is useful for prediction/forecasting problems.

Minor Changes

• Delete resample_reduce by @dcherian in #246
• Fix test failure on i386 by @avalentino in #248
• typing fixes by @Illviljan in #235, #253
• replace the deprecated provision-with-micromamba with setup-micromamba by @keewis in #258
• compatibility with numpy>=2.0 by @keewis in #257
• convert datetime: micro-optimizations by @mathause in #261

New Contributors

• @keewis made their first contribution in #258
• @mathause made their first contribution in #261

Full Changelog: v0.7.2...v0.8.0

v0.7.2

New reductions: nanfirst, nanlast, nanargmax, nanargmin .

Please test before using in production

What's Changed

• Support nanfirst, nanlast with simple combine algo by @dcherian in #240
• Enable nanargmax, nanargmin by @dcherian in #171

Full Changelog: v0.7.1...v0.7.2

v0.7.1

What's Changed

Bugfix!

• Check method only for dask reductions. by @dcherian in #241

Full Changelog: v0.7.0...v0.7.1

v0.7.0

What's Changed

Optimizations and performance improvements.

• Always factorize early by @dcherian in #234
• Handle min_count=0 by @dcherian in #238
• Optimize broadcasting by @dcherian in #230

Full Changelog: v0.6.10...v0.7.0

v0.6.10

Small performance improvement.

What's Changed

• Always reindex=True for all numpy inputs by @dcherian in #228

Full Changelog: v0.6.9...v0.6.10