-
Notifications
You must be signed in to change notification settings - Fork 9
Expand file tree
/
Copy pathnep-0026-missing-data-summary.html
More file actions
1247 lines (1059 loc) · 72.5 KB
/
nep-0026-missing-data-summary.html
File metadata and controls
1247 lines (1059 loc) · 72.5 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en" data-content_root="./" >
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" /><meta name="viewport" content="width=device-width, initial-scale=1" />
<title>NEP 26 — Summary of missing data NEPs and discussion — NumPy Enhancement Proposals</title>
<script data-cfasync="false">
document.documentElement.dataset.mode = localStorage.getItem("mode") || "";
document.documentElement.dataset.theme = localStorage.getItem("theme") || "";
</script>
<!--
this give us a css class that will be invisible only if js is disabled
-->
<noscript>
<style>
.pst-js-only { display: none !important; }
</style>
</noscript>
<!-- Loaded before other Sphinx assets -->
<link href="_static/styles/theme.css?digest=55c26ed414f5f8a31ebb" rel="stylesheet" />
<link href="_static/styles/pydata-sphinx-theme.css?digest=55c26ed414f5f8a31ebb" rel="stylesheet" />
<link rel="stylesheet" type="text/css" href="_static/pygments.css?v=03e43079" />
<!-- So that users can add custom icons -->
<script defer src="_static/scripts/fontawesome.js?digest=55c26ed414f5f8a31ebb"></script>
<!-- Pre-loaded scripts that we'll load fully later -->
<link rel="preload" as="script" href="_static/scripts/bootstrap.js?digest=55c26ed414f5f8a31ebb" />
<link rel="preload" as="script" href="_static/scripts/pydata-sphinx-theme.js?digest=55c26ed414f5f8a31ebb" />
<script src="_static/documentation_options.js?v=7f41d439"></script>
<script src="_static/doctools.js?v=888ff710"></script>
<script src="_static/sphinx_highlight.js?v=dc90522c"></script>
<script>DOCUMENTATION_OPTIONS.pagename = 'nep-0026-missing-data-summary';</script>
<script>DOCUMENTATION_OPTIONS.search_as_you_type = false;</script>
<link rel="icon" href="_static/favicon.ico"/>
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="NEP 30 — Duck typing for NumPy arrays - implementation" href="nep-0030-duck-array-protocol.html" />
<link rel="prev" title="NEP 25 — NA support via special dtypes" href="nep-0025-missing-data-3.html" />
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<meta name="docsearch:language" content="en"/>
<meta name="docsearch:version" content="" />
<meta name="docbuild:last-update" content="Apr 27, 2026"/>
<script src="_static/searchtools.js"></script>
<script src="_static/language_data.js"></script>
<script src="searchindex.js"></script>
</head>
<body data-default-mode="">
<div id="pst-skip-link" class="skip-link d-print-none"><a href="#main-content">Skip to main content</a></div>
<div id="pst-scroll-pixel-helper"></div>
<button type="button" class="btn rounded-pill" id="pst-back-to-top">
<i class="fa-solid fa-arrow-up"></i>Back to top</button>
<dialog id="pst-search-dialog">
<form class="bd-search d-flex align-items-center"
action="search.html"
method="get">
<i class="fa-solid fa-magnifying-glass"></i>
<input type="search"
class="form-control"
name="q"
placeholder="Search the docs ..."
aria-label="Search the docs ..."
autocomplete="off"
autocorrect="off"
autocapitalize="off"
spellcheck="false"/>
<span class="search-button__kbd-shortcut"><kbd class="kbd-shortcut__modifier">Ctrl</kbd>+<kbd>K</kbd></span>
</form>
</dialog>
<div class="pst-async-banner-revealer d-none">
<aside id="bd-header-version-warning" class="d-none d-print-none" aria-label="Version warning"></aside>
</div>
<header id="pst-header" class="bd-header navbar navbar-expand-lg bd-navbar d-print-none">
<div class="bd-header__inner bd-page-width">
<button class="pst-navbar-icon sidebar-toggle primary-toggle" aria-label="Site navigation">
<span class="fa-solid fa-bars"></span>
</button>
<div class="col-lg-3 navbar-header-items__start">
<div class="navbar-item">
<a class="navbar-brand logo" href="content.html">
<img src="_static/numpylogo.svg" class="logo__image only-light" alt="NumPy Enhancement Proposals - Home"/>
<img src="_static/numpylogo_dark.svg" class="logo__image only-dark pst-js-only" alt="NumPy Enhancement Proposals - Home"/>
</a></div>
</div>
<div class="col-lg-9 navbar-header-items">
<div class="me-auto navbar-header-items__center">
<div class="navbar-item">
<nav>
<ul class="bd-navbar-elements navbar-nav">
<li class="nav-item current active">
<a class="nav-link nav-internal" href="index.html">
Index
</a>
</li>
<li class="nav-item ">
<a class="nav-link nav-internal" href="scope.html">
The Scope of NumPy
</a>
</li>
<li class="nav-item ">
<a class="nav-link nav-internal" href="roadmap.html">
Current roadmap
</a>
</li>
<li class="nav-item ">
<a class="nav-link nav-external" href="https://github.com/numpy/numpy/issues?q=is%3Aopen+is%3Aissue+label%3A%2223+-+Wish+List%22">
Wishlist
</a>
</li>
</ul>
</nav></div>
</div>
<div class="navbar-header-items__end">
<div class="navbar-item navbar-persistent--container">
<button class="btn search-button-field search-button__button pst-js-only" title="Search" aria-label="Search" data-bs-placement="bottom" data-bs-toggle="tooltip">
<i class="fa-solid fa-magnifying-glass"></i>
<span class="search-button__default-text">Search</span>
<span class="search-button__kbd-shortcut"><kbd class="kbd-shortcut__modifier">Ctrl</kbd>+<kbd class="kbd-shortcut__modifier">K</kbd></span>
</button>
</div>
<div class="navbar-item">
<button class="btn btn-sm nav-link pst-navbar-icon theme-switch-button pst-js-only" aria-label="Color mode" data-bs-title="Color mode" data-bs-placement="bottom" data-bs-toggle="tooltip">
<i class="theme-switch fa-solid fa-sun fa-lg" data-mode="light" title="Light"></i>
<i class="theme-switch fa-solid fa-moon fa-lg" data-mode="dark" title="Dark"></i>
<i class="theme-switch fa-solid fa-circle-half-stroke fa-lg" data-mode="auto" title="System Settings"></i>
</button></div>
<div class="navbar-item"><ul class="navbar-icon-links"
aria-label="Icon Links">
<li class="nav-item">
<a href="https://github.com/numpy/numpy" title="GitHub" class="nav-link pst-navbar-icon" rel="noopener" target="_blank" data-bs-toggle="tooltip" data-bs-placement="bottom"><i class="fa-brands fa-square-github fa-lg" aria-hidden="true"></i>
<span class="sr-only">GitHub</span></a>
</li>
</ul></div>
</div>
</div>
<div class="navbar-persistent--mobile">
<button class="btn search-button-field search-button__button pst-js-only" title="Search" aria-label="Search" data-bs-placement="bottom" data-bs-toggle="tooltip">
<i class="fa-solid fa-magnifying-glass"></i>
<span class="search-button__default-text">Search</span>
<span class="search-button__kbd-shortcut"><kbd class="kbd-shortcut__modifier">Ctrl</kbd>+<kbd class="kbd-shortcut__modifier">K</kbd></span>
</button>
</div>
<button class="pst-navbar-icon sidebar-toggle secondary-toggle" aria-label="On this page">
<span class="fa-solid fa-outdent"></span>
</button>
</div>
</header>
<div class="bd-container">
<div class="bd-container__inner bd-page-width">
<dialog id="pst-primary-sidebar-modal"></dialog>
<div id="pst-primary-sidebar" class="bd-sidebar-primary bd-sidebar">
<div class="sidebar-header-items sidebar-primary__section">
<div class="sidebar-header-items__center">
<div class="navbar-item">
<nav>
<ul class="bd-navbar-elements navbar-nav">
<li class="nav-item current active">
<a class="nav-link nav-internal" href="index.html">
Index
</a>
</li>
<li class="nav-item ">
<a class="nav-link nav-internal" href="scope.html">
The Scope of NumPy
</a>
</li>
<li class="nav-item ">
<a class="nav-link nav-internal" href="roadmap.html">
Current roadmap
</a>
</li>
<li class="nav-item ">
<a class="nav-link nav-external" href="https://github.com/numpy/numpy/issues?q=is%3Aopen+is%3Aissue+label%3A%2223+-+Wish+List%22">
Wishlist
</a>
</li>
</ul>
</nav></div>
</div>
<div class="sidebar-header-items__end">
<div class="navbar-item">
<button class="btn btn-sm nav-link pst-navbar-icon theme-switch-button pst-js-only" aria-label="Color mode" data-bs-title="Color mode" data-bs-placement="bottom" data-bs-toggle="tooltip">
<i class="theme-switch fa-solid fa-sun fa-lg" data-mode="light" title="Light"></i>
<i class="theme-switch fa-solid fa-moon fa-lg" data-mode="dark" title="Dark"></i>
<i class="theme-switch fa-solid fa-circle-half-stroke fa-lg" data-mode="auto" title="System Settings"></i>
</button></div>
<div class="navbar-item"><ul class="navbar-icon-links"
aria-label="Icon Links">
<li class="nav-item">
<a href="https://github.com/numpy/numpy" title="GitHub" class="nav-link pst-navbar-icon" rel="noopener" target="_blank" data-bs-toggle="tooltip" data-bs-placement="bottom"><i class="fa-brands fa-square-github fa-lg" aria-hidden="true"></i>
<span class="sr-only">GitHub</span></a>
</li>
</ul></div>
</div>
</div>
<div class="sidebar-primary-items__start sidebar-primary__section">
<div class="sidebar-primary-item pst-sidebar-collapse"><button id="pst-collapse-sidebar-button" aria-expanded="true" aria-controls="pst-primary-sidebar"><svg class="pst-icon svg-inline--fa" role="img" aria-hidden="true" focusable="false" viewBox="0 0 16 16" xmlns="http://www.w3.org/2000/svg">
<path fill="currentColor" d="M3 15.5C2.36232 15.5 1.74874 15.2564 1.28478 14.8189C0.820828 14.3815 0.541576 13.7832 0.504167 13.1467L0.5 13L0.5 3C0.499965 2.36232 0.743605 1.74874 1.18107 1.28478C1.61854 0.820828 2.21676 0.541576 2.85333 0.504167L3 0.5L13 0.5C13.6377 0.499965 14.2513 0.743605 14.7152 1.18107C15.1792 1.61854 15.4584 2.21676 15.4958 2.85333L15.5 3L15.5 13C15.5 13.6377 15.2564 14.2513 14.8189 14.7152C14.3815 15.1792 13.7832 15.4584 13.1467 15.4958L13 15.5L3 15.5ZM3 13.8333L10.5 13.8333L10.5 2.16667L3 2.16667C2.79589 2.16669 2.59889 2.24163 2.44636 2.37726C2.29383 2.5129 2.19638 2.69979 2.1725 2.9025L2.16667 3L2.16667 13C2.16669 13.2041 2.24163 13.4011 2.37726 13.5536C2.5129 13.7062 2.69979 13.8036 2.9025 13.8275L3 13.8333ZM6.65583 10.325L6.5775 10.2558L4.91083 8.58917C4.76735 8.44567 4.68116 8.25476 4.66843 8.05223C4.65569 7.84971 4.71729 7.6495 4.84167 7.48917L4.91083 7.41083L6.5775 5.74417C6.72747 5.59471 6.9287 5.50794 7.14032 5.50148C7.35194 5.49502 7.55809 5.56935 7.7169 5.70937C7.8757 5.8494 7.97525 6.04463 7.99533 6.25539C8.01541 6.46616 7.95451 6.67667 7.825 6.84417L7.75583 6.9225L6.67917 8L7.75583 9.0775C7.89931 9.22099 7.98551 9.41191 7.99824 9.61443C8.01097 9.81695 7.94938 10.0172 7.825 10.1775L7.75583 10.2558C7.61234 10.3993 7.42142 10.4855 7.2189 10.4982C7.01638 10.511 6.81617 10.4494 6.65583 10.325Z"/>
</svg>
<span class="pst-collapse-sidebar-label">Collapse Sidebar</span>
<span class="pst-expand-sidebar-label">Expand Sidebar</span>
</button></div>
<div class="sidebar-primary-item">
<nav class="bd-docs-nav bd-links"
aria-label="Section Navigation">
<p class="bd-links__title" role="heading" aria-level="1">Section Navigation</p>
<div class="bd-toc-item navbar-nav"><ul class="nav bd-sidenav">
<li class="toctree-l1"><a class="reference internal" href="scope.html">The Scope of NumPy</a></li>
<li class="toctree-l1"><a class="reference internal" href="roadmap.html">Current roadmap</a></li>
</ul>
<ul class="current nav bd-sidenav">
<li class="toctree-l1 has-children"><a class="reference internal" href="meta.html">Meta-NEPs (NEPs about NEPs or active Processes)</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="nep-0000.html">NEP 0 — Purpose and process</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0023-backwards-compatibility.html">NEP 23 — Backwards compatibility and deprecation policy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0036-fair-play.html">NEP 36 — Fair play</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0045-c_style_guide.html">NEP 45 — C style guide</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0046-sponsorship-guidelines.html">NEP 46 — NumPy sponsorship guidelines</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0048-spending-project-funds.html">NEP 48 — Spending NumPy project funds</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-template.html">NEP X — Template and instructions</a></li>
</ul>
</details></li>
<li class="toctree-l1 has-children"><a class="reference internal" href="provisional.html">Provisional NEPs (provisionally accepted; interface may change)</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul class="simple">
</ul>
</details></li>
<li class="toctree-l1 has-children"><a class="reference internal" href="accepted.html">Accepted NEPs (implementation in progress)</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="nep-0041-improved-dtype-support.html">NEP 41 — First step towards a new datatype system</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0042-new-dtypes.html">NEP 42 — New and extensible DTypes</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0044-restructuring-numpy-docs.html">NEP 44 — Restructuring the NumPy documentation</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0051-scalar-representation.html">NEP 51 — Changing the representation of NumPy scalars</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0054-simd-cpp-highway.html">NEP 54 — SIMD infrastructure evolution: adopting Google Highway when moving to C++</a></li>
</ul>
</details></li>
<li class="toctree-l1 has-children"><a class="reference internal" href="open.html">Open NEPs (under consideration)</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="nep-0043-extensible-ufuncs.html">NEP 43 — Enhancing the extensibility of UFuncs</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0053-c-abi-evolution.html">NEP 53 — Evolving the NumPy C-API for NumPy 2.0</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0057-numpy-platform-support.html">NEP 57 — NumPy platform support</a></li>
</ul>
</details></li>
<li class="toctree-l1 has-children"><a class="reference internal" href="finished.html">Finished NEPs</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="nep-0001-npy-format.html">NEP 1 — A simple file format for NumPy arrays</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0005-generalized-ufuncs.html">NEP 5 — Generalized universal functions</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0007-datetime-proposal.html">NEP 7 — A proposal for implementing some date/time types in NumPy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0010-new-iterator-ufunc.html">NEP 10 — Optimizing iterator/UFunc performance</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0013-ufunc-overrides.html">NEP 13 — A mechanism for overriding Ufuncs</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0014-dropping-python2.7-proposal.html">NEP 14 — Plan for dropping Python 2.7 support</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0015-merge-multiarray-umath.html">NEP 15 — Merging multiarray and umath</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0018-array-function-protocol.html">NEP 18 — A dispatch mechanism for NumPy's high level array functions</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0019-rng-policy.html">NEP 19 — Random number generator policy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0020-gufunc-signature-enhancement.html">NEP 20 — Expansion of generalized universal function signatures</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0022-ndarray-duck-typing-overview.html">NEP 22 — Duck typing for NumPy arrays – high level overview</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0027-zero-rank-arrays.html">NEP 27 — Zero rank arrays</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0028-website-redesign.html">NEP 28 — numpy.org website redesign</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0029-deprecation_policy.html">NEP 29 — Recommend Python and NumPy version support as a community policy standard</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0032-remove-financial-functions.html">NEP 32 — Remove the financial functions from NumPy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0034-infer-dtype-is-object.html">NEP 34 — Disallow inferring ``dtype=object`` from sequences</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0035-array-creation-dispatch-with-array-function.html">NEP 35 — Array creation dispatching with __array_function__</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0038-SIMD-optimizations.html">NEP 38 — Using SIMD optimization instructions for performance</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0040-legacy-datatype-impl.html">NEP 40 — Legacy datatype implementation in NumPy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0049-data-allocation-strategies.html">NEP 49 — Data allocation strategies</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0050-scalar-promotion.html">NEP 50 — Promotion rules for Python scalars</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0052-python-api-cleanup.html">NEP 52 — Python API cleanup for NumPy 2.0</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0055-string_dtype.html">NEP 55 — Add a UTF-8 variable-width string DType to NumPy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0056-array-api-main-namespace.html">NEP 56 — Array API standard support in NumPy's main namespace</a></li>
</ul>
</details></li>
<li class="toctree-l1 current active has-children"><a class="reference internal" href="deferred.html">Deferred and Superseded NEPs</a><details open="open"><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul class="current">
<li class="toctree-l2"><a class="reference internal" href="nep-0002-warnfix.html">NEP 2 — A proposal to build numpy without warning with a big set of warning flags</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0003-math_config_clean.html">NEP 3 — Cleaning the math configuration of numpy.core</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0004-datetime-proposal3.html">NEP 4 — A (third) proposal for implementing some date/time types in NumPy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0006-newbugtracker.html">NEP 6 — Replacing Trac with a different bug tracker</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0008-groupby_additions.html">NEP 8 — A proposal for adding groupby functionality to NumPy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0009-structured_array_extensions.html">NEP 9 — Structured array extensions</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0011-deferred-ufunc-evaluation.html">NEP 11 — Deferred UFunc evaluation</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0012-missing-data.html">NEP 12 — Missing data functionality in NumPy</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0021-advanced-indexing.html">NEP 21 — Simplified and explicit advanced indexing</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0024-missing-data-2.html">NEP 24 — Missing data functionality - alternative 1 to NEP 12</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0025-missing-data-3.html">NEP 25 — NA support via special dtypes</a></li>
<li class="toctree-l2 current active"><a class="current reference internal" href="#">NEP 26 — Summary of missing data NEPs and discussion</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0030-duck-array-protocol.html">NEP 30 — Duck typing for NumPy arrays - implementation</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0031-uarray.html">NEP 31 — Context-local and global overrides of the NumPy API</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0037-array-module.html">NEP 37 — A dispatch protocol for NumPy-like modules</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0047-array-api-standard.html">NEP 47 — Adopting the array API standard</a></li>
</ul>
</details></li>
<li class="toctree-l1 has-children"><a class="reference internal" href="rejected.html">Rejected and Withdrawn NEPs</a><details><summary><span class="toctree-toggle" role="presentation"><i class="fa-solid fa-chevron-down"></i></span></summary><ul>
<li class="toctree-l2"><a class="reference internal" href="nep-0016-abstract-array.html">NEP 16 — An abstract base class for identifying "duck arrays"</a></li>
<li class="toctree-l2"><a class="reference internal" href="nep-0017-split-out-maskedarray.html">NEP 17 — Split out masked arrays</a></li>
</ul>
</details></li>
</ul>
</div>
</nav></div>
</div>
<div class="sidebar-primary-items__end sidebar-primary__section">
<div class="sidebar-primary-item">
<div id="ethical-ad-placement"
class="flat"
data-ea-publisher="readthedocs"
data-ea-type="readthedocs-sidebar"
data-ea-manual="true">
</div></div>
</div>
</div>
<main id="main-content" class="bd-main" role="main">
<div class="bd-content">
<div class="bd-article-container">
<div class="bd-header-article d-print-none">
<div class="header-article-items header-article__inner">
<div class="header-article-items__start">
<div class="header-article-item">
<nav aria-label="Breadcrumb" class="d-print-none">
<ul class="bd-breadcrumbs">
<li class="breadcrumb-item breadcrumb-home">
<a href="content.html" class="nav-link" aria-label="Home">
<i class="fa-solid fa-home"></i>
</a>
</li>
<li class="breadcrumb-item"><a href="index.html" class="nav-link">Roadmap & NumPy enhancement proposals</a></li>
<li class="breadcrumb-item"><a href="deferred.html" class="nav-link">Deferred and Superseded NEPs</a></li>
<li class="breadcrumb-item active" aria-current="page"><span class="ellipsis">NEP 26 — Summary of missing data NEPs and discussion</span></li>
</ul>
</nav>
</div>
</div>
</div>
</div>
<div id="searchbox"></div>
<article class="bd-article">
<section id="nep-26-summary-of-missing-data-neps-and-discussion">
<span id="nep26"></span><h1>NEP 26 — Summary of missing data NEPs and discussion<a class="headerlink" href="#nep-26-summary-of-missing-data-neps-and-discussion" title="Link to this heading">#</a></h1>
<dl class="field-list simple">
<dt class="field-odd">Author<span class="colon">:</span></dt>
<dd class="field-odd"><p>Mark Wiebe <<a class="reference external" href="mailto:mwwiebe%40gmail.com">mwwiebe<span>@</span>gmail<span>.</span>com</a>>, Nathaniel J. Smith <<a class="reference external" href="mailto:njs%40pobox.com">njs<span>@</span>pobox<span>.</span>com</a>></p>
</dd>
<dt class="field-even">Status<span class="colon">:</span></dt>
<dd class="field-even"><p>Deferred</p>
</dd>
<dt class="field-odd">Type<span class="colon">:</span></dt>
<dd class="field-odd"><p>Standards Track</p>
</dd>
<dt class="field-even">Created<span class="colon">:</span></dt>
<dd class="field-even"><p>2012-04-22</p>
</dd>
</dl>
<p><em>Context</em>: this NEP was written as summary of the large number of discussions
and proposals (<a class="reference internal" href="nep-0012-missing-data.html#nep12"><span class="std std-ref">NEP 12 — Missing data functionality in NumPy</span></a>, <a class="reference internal" href="nep-0024-missing-data-2.html#nep24"><span class="std std-ref">NEP 24 — Missing data functionality - alternative 1 to NEP 12</span></a>, <a class="reference internal" href="nep-0025-missing-data-3.html#nep25"><span class="std std-ref">NEP 25 — NA support via special dtypes</span></a>), regarding missing data
functionality.</p>
<p>The debate about how NumPy should handle missing data, a subject with
many preexisting approaches, requirements, and conventions, has been long and
contentious. There has been more than one proposal for how to implement
support into NumPy, and there is a testable implementation which is
merged into NumPy’s current main. The vast number of emails and differing
points of view has made it difficult for interested parties to understand
the issues and be comfortable with the direction NumPy is going.</p>
<p>Here is our (Mark and Nathaniel’s) attempt to summarize the
problem, proposals, and points of agreement/disagreement in a single
place, to help the community move towards consensus.</p>
<section id="the-numpy-developers-problem">
<h2>The NumPy developers’ problem<a class="headerlink" href="#the-numpy-developers-problem" title="Link to this heading">#</a></h2>
<p>For this discussion, “missing data” means array elements
which can be indexed (e.g. A[3] in an array A with shape (5,)),
but have, in some sense, no value.</p>
<p>It does not refer to compressed or sparse storage techniques where
the value for A[3] is not actually stored in memory, but still has a
well-defined value like 0.</p>
<p>This is still vague, and to create an actual implementation,
it is necessary to answer such questions as:</p>
<ul class="simple">
<li><p>What values are computed when doing element-wise ufuncs.</p></li>
<li><p>What values are computed when doing reductions.</p></li>
<li><p>Whether the storage for an element gets overwritten when marking
that value missing.</p></li>
<li><p>Whether computations resulting in NaN automatically treat in the
same way as a missing value.</p></li>
<li><p>Whether one interacts with missing values using a placeholder object
(e.g. called “NA” or “masked”), or through a separate boolean array.</p></li>
<li><p>Whether there is such a thing as an array object that cannot hold
missing array elements.</p></li>
<li><p>How the (C and Python) API is expressed, in terms of dtypes,
masks, and other constructs.</p></li>
<li><p>If we decide to answer some of these questions in multiple ways,
then that creates the question of whether that requires multiple
systems, and if so how they should interact.</p></li>
</ul>
<p>There’s clearly a very large space of missing-data APIs that <em>could</em>
be implemented. There is likely at least one user, somewhere, who
would find any possible implementation to be just the thing they
need to solve some problem. On the other hand, much of NumPy’s power
and clarity comes from having a small number of orthogonal concepts,
such as strided arrays, flexible indexing, broadcasting, and ufuncs,
and we’d like to preserve that simplicity.</p>
<p>There has been dissatisfaction among several major groups of NumPy users
about the existing status quo of missing data support. In particular,
neither the numpy.ma component nor use of floating-point NaNs as a
missing data signal fully satisfy the performance requirements and
ease of use for these users. The example of R, where missing data
is treated via an NA placeholder and is deeply integrated into all
computation, is where many of these users point to indicate what
functionality they would like. Doing a deep integration of missing
data like in R must be considered carefully, it must be clear it
is not being done in a way which sacrifices existing performance
or functionality.</p>
<p>Our problem is, how can we choose some incremental additions to
NumPy that will make a large class of users happy, be
reasonably elegant, complement the existing design, and that we’re
comfortable we won’t regret being stuck with in the long term.</p>
</section>
<section id="prior-art">
<h2>Prior art<a class="headerlink" href="#prior-art" title="Link to this heading">#</a></h2>
<p>So a major (maybe <em>the</em> major) problem is figuring out how ambitious
the project to add missing data support to NumPy should be, and which
kinds of problems are in scope. Let’s start with the
best understood situation where “missing data” comes into play:</p>
<section id="statistical-missing-data">
<h3>“Statistical missing data”<a class="headerlink" href="#statistical-missing-data" title="Link to this heading">#</a></h3>
<p>In statistics, social science, etc., “missing data” is a term of art
referring to a specific (but extremely common and important)
situation: we have tried to gather some measurements according to some
scheme, but some of these measurements are missing. For example, if we
have a table listing the height, age, and income of a number of
individuals, but one person did not provide their income, then we need
some way to represent this:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">Person</span> <span class="o">|</span> <span class="n">Height</span> <span class="o">|</span> <span class="n">Age</span> <span class="o">|</span> <span class="n">Income</span>
<span class="o">------------------------------</span>
<span class="mi">1</span> <span class="o">|</span> <span class="mi">63</span> <span class="o">|</span> <span class="mi">25</span> <span class="o">|</span> <span class="mi">15000</span>
<span class="mi">2</span> <span class="o">|</span> <span class="mi">58</span> <span class="o">|</span> <span class="mi">32</span> <span class="o">|</span> <span class="o"><</span><span class="n">missing</span><span class="o">></span>
<span class="mi">3</span> <span class="o">|</span> <span class="mi">71</span> <span class="o">|</span> <span class="mi">45</span> <span class="o">|</span> <span class="mi">30000</span>
</pre></div>
</div>
<p>The traditional way is to record that income as, say, “-99”, and
document this in the README along with the data set. Then, you have to
remember to check for and handle such incomes specially; if you
forget, you’ll get superficially reasonable but completely incorrect
results, like calculating the average income on this data set as
14967. If you’re in one of these fields, then such missing-ness is
routine and inescapable, and if you use the “-99” approach then it’s a
pitfall you have to remember to check for explicitly on literally
<em>every</em> calculation you ever do. This is, obviously, an unpleasant way
to live.</p>
<p>Let’s call this situation the “statistical missing data” situation,
just to have a convenient handle for it. (As mentioned, practitioners
just call this “missing data”, and what to do about it is literally an
entire sub-field of statistics; if you google “missing data” then
every reference is on how to handle it.) NumPy isn’t going to do
automatic imputation or anything like that, but it could help a great
deal by providing some standard way to at least represent data which
is missing in this sense.</p>
<p>The main prior art for how this could be done comes from the S/S+/R
family of languages. Their strategy is, for each type they support,
to define a special value called “NA”. (For ints this is INT_MAX,
for floats it’s a special NaN value that’s distinguishable from
other NaNs, …) Then, they arrange that in computations, this
value has a special semantics that we will call “NA semantics”.</p>
</section>
<section id="na-semantics">
<h3>NA semantics<a class="headerlink" href="#na-semantics" title="Link to this heading">#</a></h3>
<p>The idea of NA semantics is that any computations involving NA
values should be consistent with what would have happened if we
had known the correct value.</p>
<p>For example, let’s say we want to compute the mean income, how might
we do this? One way would be to just ignore the missing entry, and
compute the mean of the remaining entries. This gives us (15000 +
30000)/2, or 22500.</p>
<p>Is this result consistent with discovering the income of person 2?
Let’s say we find out that person 2’s income is 50000. This means
the correct answer is (15000 + 50000 + 30000)/3, or 31666.67,
indicating clearly that it is not consistent. Therefore, the mean
income is NA, i.e. a specific number whose value we are unable
to compute.</p>
<p>This motivates the following rules, which are how R implements NA:</p>
<dl>
<dt>Assignment:</dt><dd><p>NA values are understood to represent specific
unknown values, and thus should have value-like semantics with
respect to assignment and other basic data manipulation
operations. Code which does not actually look at the values involved
should work the same regardless of whether some of them are
missing. For example, one might write:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">income</span><span class="p">[:]</span> <span class="o">=</span> <span class="n">income</span><span class="p">[</span><span class="n">np</span><span class="o">.</span><span class="n">argsort</span><span class="p">(</span><span class="n">height</span><span class="p">)]</span>
</pre></div>
</div>
<p>to perform an in-place sort of the <code class="docutils literal notranslate"><span class="pre">income</span></code> array, and know that
the shortest person’s income would end up being first. It turns out
that the shortest person’s income is not known, so the array should
end up being <code class="docutils literal notranslate"><span class="pre">[NA,</span> <span class="pre">15000,</span> <span class="pre">30000]</span></code>, but there’s nothing
special about NAness here.</p>
</dd>
<dt>Propagation:</dt><dd><p>In the example above, we concluded that an operation like <code class="docutils literal notranslate"><span class="pre">mean</span></code>
should produce NA when one of its data values was NA.
If you ask me, “what is 3 plus x?”, then my only possible answer is
“I don’t know what x is, so I don’t know what 3 + x is either”. NA
means “I don’t know”, so 3 + NA is NA.</p>
<p>This is important for safety when analyzing data: missing data often
requires special handling for correctness – the fact that you are
missing information might mean that something you wanted to compute
cannot actually be computed, and there are whole books written on
how to compensate in various situations. Plus, it’s easy to not
realize that you have missing data, and write code that assumes you
have all the data. Such code should not silently produce the wrong
answer.</p>
<p>There is an important exception to characterizing this as propagation,
in the case of boolean values. Consider the calculation:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">v</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">any</span><span class="p">([</span><span class="kc">False</span><span class="p">,</span> <span class="kc">False</span><span class="p">,</span> <span class="n">NA</span><span class="p">,</span> <span class="kc">True</span><span class="p">])</span>
</pre></div>
</div>
<p>If we strictly propagate, <code class="docutils literal notranslate"><span class="pre">v</span></code> will become NA. However, no
matter whether we place True or False into the third array position,
<code class="docutils literal notranslate"><span class="pre">v</span></code> will then get the value True. The answer to the question
“Is the result True consistent with later discovering the value
that was missing?” is yes, so it is reasonable to not propagate here,
and instead return the value True. This is what R does:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="o">></span> <span class="nb">any</span><span class="p">(</span><span class="n">c</span><span class="p">(</span><span class="n">F</span><span class="p">,</span> <span class="n">F</span><span class="p">,</span> <span class="n">NA</span><span class="p">,</span> <span class="n">T</span><span class="p">))</span>
<span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="n">TRUE</span>
<span class="o">></span> <span class="nb">any</span><span class="p">(</span><span class="n">c</span><span class="p">(</span><span class="n">F</span><span class="p">,</span> <span class="n">F</span><span class="p">,</span> <span class="n">NA</span><span class="p">,</span> <span class="n">F</span><span class="p">))</span>
<span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="n">NA</span>
</pre></div>
</div>
</dd>
<dt>Other:</dt><dd><p>NaN and NA are conceptually distinct. 0.0/0.0 is not a mysterious,
unknown value – it’s defined to be NaN by IEEE floating point, Not
a Number. NAs are numbers (or strings, or whatever), just unknown
ones. Another small but important difference is that in Python, <code class="docutils literal notranslate"><span class="pre">if</span>
<span class="pre">NaN:</span> <span class="pre">...</span></code> treats NaN as True (NaN is “truthy”); but <code class="docutils literal notranslate"><span class="pre">if</span> <span class="pre">NA:</span> <span class="pre">...</span></code>
would be an error.</p>
<p>In R, all reduction operations implement an alternative semantics,
activated by passing a special argument (<code class="docutils literal notranslate"><span class="pre">na.rm=TRUE</span></code> in R).
<code class="docutils literal notranslate"><span class="pre">sum(a)</span></code> means “give me the sum of all the
values” (which is NA if some of the values are NA);
<code class="docutils literal notranslate"><span class="pre">sum(a,</span> <span class="pre">na.rm=True)</span></code> means “give me the sum of all the non-NA
values”.</p>
</dd>
</dl>
</section>
<section id="other-prior-art">
<h3>Other prior art<a class="headerlink" href="#other-prior-art" title="Link to this heading">#</a></h3>
<p>Once we move beyond the “statistical missing data” case, the correct
behavior for missing data becomes less clearly defined. There are many
cases where specific elements are singled out to be treated specially
or excluded from computations, and these could often be conceptualized
as involving ‘missing data’ in some sense.</p>
<p>In image processing, it’s common to use a single image together with
one or more boolean masks to e.g. composite subsets of an image. As
Joe Harrington pointed out on the list, in the context of processing
astronomical images, it’s also common to generalize to a
floating-point valued mask, or alpha channel, to indicate degrees of
“missingness”. We think this is out of scope for the present design,
but it is an important use case, and ideally NumPy should support
natural ways of manipulating such data.</p>
<p>After R, numpy.ma is probably the most mature source of
experience on missing-data-related APIs. Its design is quite different
from R; it uses different semantics – reductions skip masked values
by default and NaNs convert to masked – and it uses a different
storage strategy via a separate mask. While it seems to be generally
considered sub-optimal for general use, it’s hard to pin down whether
this is because the API is immature but basically good, or the API
is fundamentally broken, or the API is great but the code should be
faster, or what. We looked at some of those users to try and get a
better idea.</p>
<p>Matplotlib is perhaps the best known package to rely on numpy.ma. It
seems to use it in two ways. One is as a way for users to indicate
what data is missing when passing it to be graphed. (Other ways are
also supported, e.g., passing in NaN values gives the same result.) In
this regard, matplotlib treats np.ma.masked and NaN values in the same way
that R’s plotting routines handle NA and NaN values. For these purposes,
matplotlib doesn’t really care what semantics or storage strategy is
used for missing data.</p>
<p>Internally, matplotlib uses numpy.ma arrays to store and pass around
separately computed boolean masks containing ‘validity’ information
for each input array in a cheap and non-destructive fashion. Mark’s
impression from some shallow code review is that mostly it works
directly with the data and mask attributes of the masked arrays,
not extensively using the particular computational semantics of
numpy.ma. So, for this usage they do rely on the non-destructive
mask-based storage, but this doesn’t say much about what semantics
are needed.</p>
<p>Paul Hobson <a class="reference external" href="https://mail.scipy.org/pipermail/numpy-discussion/2012-April/061743.html">posted some code</a> on the list that uses numpy.ma for
storing arrays of contaminant concentration measurements. Here the
mask indicates whether the corresponding number represents an actual
measurement, or just the estimated detection limit for a concentration
which was too small to detect. Nathaniel’s impression from reading
through this code is that it also mostly uses the .data and .mask
attributes in preference to performing operations on the MaskedArray
directly.</p>
<p>So, these examples make it clear that there is demand for a convenient
way to keep a data array and a mask array (or even a floating point
array) bundled up together and “aligned”. But they don’t tell us much
about what semantics the resulting object should have with respect to
ufuncs and friends.</p>
</section>
</section>
<section id="semantics-storage-api-oh-my">
<h2>Semantics, storage, API, oh my!<a class="headerlink" href="#semantics-storage-api-oh-my" title="Link to this heading">#</a></h2>
<p>We think it’s useful to draw a clear line between use cases,
semantics, and storage. Use cases are situations that users encounter,
regardless of what NumPy does; they’re the focus of the previous
section. When we say <em>semantics</em>, we mean the result of different
operations as viewed from the Python level without regard to the
underlying implementation.</p>
<p><em>NA semantics</em> are the ones described above and used by R:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="mi">1</span> <span class="o">+</span> <span class="n">NA</span> <span class="o">=</span> <span class="n">NA</span>
<span class="nb">sum</span><span class="p">([</span><span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="n">NA</span><span class="p">])</span> <span class="o">=</span> <span class="n">NA</span>
<span class="n">NA</span> <span class="o">|</span> <span class="kc">False</span> <span class="o">=</span> <span class="n">NA</span>
<span class="n">NA</span> <span class="o">|</span> <span class="kc">True</span> <span class="o">=</span> <span class="kc">True</span>
</pre></div>
</div>
<p>With <code class="docutils literal notranslate"><span class="pre">na.rm=TRUE</span></code> or <code class="docutils literal notranslate"><span class="pre">skipNA=True</span></code>, this switches to:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="mi">1</span> <span class="o">+</span> <span class="n">NA</span> <span class="o">=</span> <span class="n">illegal</span> <span class="c1"># in R, only reductions take na.rm argument</span>
<span class="nb">sum</span><span class="p">([</span><span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="n">NA</span><span class="p">],</span> <span class="n">skipNA</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span> <span class="o">=</span> <span class="mi">3</span>
</pre></div>
</div>
<p>There’s also been discussion of what we’ll call <em>ignore
semantics</em>. These are somewhat underdefined:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="nb">sum</span><span class="p">([</span><span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="n">IGNORED</span><span class="p">])</span> <span class="o">=</span> <span class="mi">3</span>
<span class="c1"># Several options here:</span>
<span class="mi">1</span> <span class="o">+</span> <span class="n">IGNORED</span> <span class="o">=</span> <span class="mi">1</span>
<span class="c1"># or</span>
<span class="mi">1</span> <span class="o">+</span> <span class="n">IGNORED</span> <span class="o">=</span> <span class="o"><</span><span class="n">leaves</span> <span class="n">output</span> <span class="n">array</span> <span class="n">untouched</span><span class="o">></span>
<span class="c1"># or</span>
<span class="mi">1</span> <span class="o">+</span> <span class="n">IGNORED</span> <span class="o">=</span> <span class="n">IGNORED</span>
</pre></div>
</div>
<p>The numpy.ma semantics are:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="nb">sum</span><span class="p">([</span><span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="n">masked</span><span class="p">])</span> <span class="o">=</span> <span class="mi">3</span>
<span class="mi">1</span> <span class="o">+</span> <span class="n">masked</span> <span class="o">=</span> <span class="n">masked</span>
</pre></div>
</div>
<p>If either NA or ignore semantics are implemented with masks, then there
is a choice of what should be done to the value in the storage
for an array element which gets assigned a missing value. Three
possibilities are:</p>
<ul class="simple">
<li><p>Leave that memory untouched (the choice made in the NEP).</p></li>
<li><p>Do the calculation with the values independently of the mask
(perhaps the most useful option for Paul Hobson’s use-case above).</p></li>
<li><p>Copy whatever value is stored behind the input missing value into
the output (this is what numpy.ma does. Even that is ambiguous in
the case of <code class="docutils literal notranslate"><span class="pre">masked</span> <span class="pre">+</span> <span class="pre">masked</span></code> – in this case numpy.ma copies the
value stored behind the leftmost masked value).</p></li>
</ul>
<p>When we talk about <em>storage</em>, we mean the debate about whether missing
values should be represented by designating a particular value of the
underlying data-type (the <em>bitpattern dtype</em> option, as used in R), or
by using a separate <em>mask</em> stored alongside the data itself.</p>
<p>For mask-based storage, there is also an important question about what
the API looks like for accessing the mask, modifying the mask, and
“peeking behind” the mask.</p>
</section>
<section id="designs-that-have-been-proposed">
<h2>Designs that have been proposed<a class="headerlink" href="#designs-that-have-been-proposed" title="Link to this heading">#</a></h2>
<p>One option is to just copy R, by implementing a mechanism whereby
dtypes can arrange for certain bitpatterns to be given NA semantics.</p>
<p>One option is to copy numpy.ma closely, but with a more optimized
implementation. (Or to simply optimize the existing implementation.)</p>
<p>One option is that described in <cite>NEP12</cite>, for which an implementation
of mask-based missing data exists. This system is roughly:</p>
<ul class="simple">
<li><p>There is both bitpattern and mask-based missing data, and both
have identical interoperable NA semantics.</p></li>
<li><p>Masks are modified by assigning np.NA or values to array elements.
The way to peek behind the mask or to unmask values is to keep a
view of the array that shares the data pointer but not the mask pointer.</p></li>
<li><p>Mark would like to add a way to access and manipulate the mask more
directly, to be used in addition to this view-based API.</p></li>
<li><p>If an array has both a bitpattern dtype and a mask, then assigning
np.NA writes to the mask, rather than to the array itself. Writing
a bitpattern NA to an array which supports both requires accessing
the data by “peeking under the mask”.</p></li>
</ul>
<p>Another option is that described in <cite>NEP24</cite>, which is to implement
bitpattern dtypes with NA semantics for the “statistical missing data”
use case, and to also implement a totally independent API for masked
arrays with ignore semantics and all mask manipulation done explicitly
through a .mask attribute.</p>
<p>Another option would be to define a minimalist aligned array container
that holds multiple arrays and that can be used to pass them around
together. It would support indexing (to help with the common problem
of wanting to subset several arrays together without their becoming
unaligned), but all arithmetic etc. would be done by accessing the
underlying arrays directly via attributes. The “prior art” discussion
above suggests that something like this holding a .data and a .mask
array might actually be solve a number of people’s problems without
requiring any major architectural changes to NumPy. This is similar to
a structured array, but with each field in a separately stored array
instead of packed together.</p>
<p>Several people have suggested that there should be a single system
that has multiple missing values that each have different semantics,
e.g., a MISSING value that has NA semantics, and a separate IGNORED
value that has ignored semantics.</p>
<p>None of these options are necessarily exclusive.</p>
</section>
<section id="the-debate">
<h2>The debate<a class="headerlink" href="#the-debate" title="Link to this heading">#</a></h2>
<p>We both are dubious of using ignored semantics as a default missing
data behavior. <strong>Nathaniel</strong> likes NA semantics because he is most
interested in the “statistical missing data” use case, and NA semantics
are exactly right for that. <strong>Mark</strong> isn’t as interested in that use
case in particular, but he likes the NA computational abstraction
because it is unambiguous and well-defined in all cases, and has a
lot of existing experience to draw from.</p>
<p>What <strong>Nathaniel</strong> thinks, overall:</p>
<ul class="simple">
<li><p>The “statistical missing data” use case is clear and compelling; the
other use cases certainly deserve our attention, but it’s hard to say what
they <em>are</em> exactly yet, or even if the best way to support them is
by extending the ndarray object.</p></li>
<li><p>The “statistical missing data” use case is best served by an R-style
system that uses bitpattern storage to implement NA semantics. The
main advantage of bitpattern storage for this use case is that it
avoids the extra memory and speed overhead of storing and checking a
mask (especially for the common case of floating point data, where
some tricks with NaNs allow us to effectively hardware-accelerate
most NA operations). These concerns alone appears to make a
mask-based implementation unacceptable to many NA users,
particularly in areas like neuroscience (where memory is tight) or
financial modeling (where milliseconds are critical). In addition,
the bit-pattern approach is less confusing conceptually (e.g.,
assignment really is just assignment, no magic going on behind the
curtain), and it’s possible to have in-memory compatibility with R
for inter-language calls via rpy2. The main disadvantage of the
bitpattern approach is the need to give up a value to represent NA,
but this is not an issue for the most important data types (float,
bool, strings, enums, objects); really, only integers are
affected. And even for integers, giving up a value doesn’t really
matter for statistical problems. (Occupy Wall Street
notwithstanding, no-one’s income is 2**63 - 1. And if it were, we’d
be switching to floats anyway to avoid overflow.)</p></li>
<li><p>Adding new dtypes requires some cooperation with the ufunc and
casting machinery, but doesn’t require any architectural changes or
violations of NumPy’s current orthogonality.</p></li>
<li><p>His impression from the mailing list discussion, esp. the <a class="reference external" href="http://thread.gmane.org/gmane.comp.python.numeric.general/46704">“what can
we agree on?” thread</a>, is that many numpy.ma users specifically
like the combination of masked storage, the mask being easily
accessible through the API, and ignored semantics. He could be
wrong, of course. But he cannot remember seeing anybody besides Mark
advocate for the specific combination of masked storage and NA
semantics, which makes him nervous.</p>
</li>
<li><p>Also, he personally is not very happy with the idea of having two
storage implementations that are almost-but-not-quite identical at
the Python level. While there likely are people who would like to
temporarily pretend that certain data is “statistically missing
data” without making a copy of their array, it’s not at all clear
that they outnumber the people who would like to use bitpatterns and
masks simultaneously for distinct purposes. And honestly he’d like
to be able to just ignore masks if he wants and stick to
bitpatterns, which isn’t possible if they’re coupled together
tightly in the API. So he would say the jury is still very much out
on whether this aspect of the NEP design is an advantage or a
disadvantage. (Certainly he’s never heard of any R users complaining
that they really wish they had an option of making a different
trade-off here.)</p></li>
<li><p>R’s NA support is a <a class="reference external" href="http://www.sr.bham.ac.uk/~ajrs/R/why_R.html">headline feature</a> and its target audience
consider it a compelling advantage over other platforms like Matlab
or Python. Working with statistical missing data is very painful
without platform support.</p>
</li>
<li><p>By comparison, we clearly have much more uncertainty about the use
cases that require a mask-based implementation, and it doesn’t seem
like people will suffer too badly if they are forced for now to
settle for using NumPy’s excellent mask-based indexing, the new
where= support, and even numpy.ma.</p></li>
<li><p>Therefore, bitpatterns with NA semantics seem to meet the criteria
of making a large class of users happy, in an elegant way, that fits
into the original design, and where we can have reasonable certainty
that we understand the problem and use cases well enough that we’ll
be happy with them in the long run. But no mask-based storage
proposal does, yet.</p></li>
</ul>
<p>What <strong>Mark</strong> thinks, overall:</p>
<ul class="simple">
<li><p>The idea of using NA semantics by default for missing data, inspired
by the “statistical missing data” problem, is better than all the
other default behaviors which were considered. This applies equally
to the bitpattern and the masked approach.</p></li>
<li><p>For NA-style functionality to get proper support by all NumPy
features and eventually all third-party libraries, it needs to be
in the core. How to correctly and efficiently handle missing data
differs by algorithm, and if thinking about it is required to fully
support NumPy, NA support will be broader and higher quality.</p></li>
<li><p>At the same time, providing two different missing data interfaces,
one for masks and one for bitpatterns, requires NumPy developers
and third-party NumPy plugin developers to separately consider the
question of what to do in either case, and do two additional
implementations of their code. This complicates their job,
and could lead to inconsistent support for missing data.</p></li>
<li><p>Providing the ability to work with both masks and bitpatterns through
the same C and Python programming interface makes missing data support
cleanly orthogonal with all other NumPy features.</p></li>
<li><p>There are many trade-offs of memory usage, performance, correctness, and
flexibility between masks and bitpatterns. Providing support for both
approaches allows users of NumPy to choose the approach which is
most compatible with their way of thinking, or has characteristics
which best match their use-case. Providing them through the same
interface further allows them to try both with minimal effort, and
choose the one which performs better or uses the least memory for
their programs.</p></li>
<li><p>Memory Usage</p>
<ul>
<li><p>With bitpatterns, less memory is used for storing a single array
containing some NAs.</p></li>
<li><p>With masks, less memory is used for storing multiple arrays that
are identical except for the location of their NAs. (In this case a
single data array can be re-used with multiple mask arrays;
bitpattern NAs would need to copy the whole data array.)</p></li>
</ul>
</li>
<li><p>Performance</p>
<ul>
<li><p>With bitpatterns, the floating point type can use native hardware
operations, with nearly correct behavior. For fully correct floating
point behavior and with other types, code must be written which
specially tests for equality with the missing-data bitpattern.</p></li>
<li><p>With masks, there is always the overhead of accessing mask memory
and testing its truth value. The implementation that currently exists
has no performance tuning, so it is only good to judge a minimum
performance level. Optimal mask-based code is in general going to
be slower than optimal bitpattern-based code.</p></li>
</ul>
</li>
<li><p>Correctness</p>
<ul>
<li><p>Bitpattern integer types must sacrifice a valid value to represent NA.
For larger integer types, there are arguments that this is ok, but for
8-bit types there is no reasonable choice. In the floating point case,
if the performance of native floating point operations is chosen,
there is a small inconsistency that NaN+NA and NA+NaN are different.</p></li>
<li><p>With masks, it works correctly in all cases.</p></li>
</ul>
</li>
<li><p>Generality</p>
<ul>
<li><p>The bitpattern approach can work in a fully general way only when
there is a specific value which can be given up from the
data type. For IEEE floating point, a NaN is an obvious choice,
and for booleans represented as a byte, there are plenty of choices.
For integers, a valid value must be sacrificed to use this approach.
Third-party dtypes which plug into NumPy will also have to
make a bitpattern choice to support this system, something which
may not always be possible.</p></li>
<li><p>The mask approach works universally with all data types.</p></li>
</ul>
</li>
</ul>
</section>
<section id="recommendations-for-moving-forward">
<h2>Recommendations for moving forward<a class="headerlink" href="#recommendations-for-moving-forward" title="Link to this heading">#</a></h2>
<p><strong>Nathaniel</strong> thinks we should:</p>
<ul class="simple">
<li><p>Go ahead and implement bitpattern NAs.</p></li>