summaryrefslogtreecommitdiffstats
path: root/ml/ml.cc
blob: 75f2e727a0722d9b97543c036b873aec7e1a3b9f (plain)
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
// SPDX-License-Identifier: GPL-3.0-or-later

#include <dlib/clustering.h>

#include "ml-private.h"

#include <random>

#include "ad_charts.h"
#include "database/sqlite/sqlite3.h"

#define WORKER_TRAIN_QUEUE_POP         0
#define WORKER_TRAIN_ACQUIRE_DIMENSION 1
#define WORKER_TRAIN_QUERY             2
#define WORKER_TRAIN_KMEANS            3
#define WORKER_TRAIN_UPDATE_MODELS     4
#define WORKER_TRAIN_RELEASE_DIMENSION 5
#define WORKER_TRAIN_UPDATE_HOST       6
#define WORKER_TRAIN_FLUSH_MODELS      7

static sqlite3 *db = NULL;
static netdata_mutex_t db_mutex = NETDATA_MUTEX_INITIALIZER;

/*
 * Functions to convert enums to strings
*/

__attribute__((unused)) static const char *
ml_machine_learning_status_to_string(enum ml_machine_learning_status mls)
{
    switch (mls) {
        case MACHINE_LEARNING_STATUS_ENABLED:
            return "enabled";
        case MACHINE_LEARNING_STATUS_DISABLED_DUE_TO_EXCLUDED_CHART:
            return "disabled-sp";
        default:
            return "unknown";
    }
}

__attribute__((unused)) static const char *
ml_metric_type_to_string(enum ml_metric_type mt)
{
    switch (mt) {
        case METRIC_TYPE_CONSTANT:
            return "constant";
        case METRIC_TYPE_VARIABLE:
            return "variable";
        default:
            return "unknown";
    }
}

__attribute__((unused)) static const char *
ml_training_status_to_string(enum ml_training_status ts)
{
    switch (ts) {
        case TRAINING_STATUS_PENDING_WITH_MODEL:
            return "pending-with-model";
        case TRAINING_STATUS_PENDING_WITHOUT_MODEL:
            return "pending-without-model";
        case TRAINING_STATUS_TRAINED:
            return "trained";
        case TRAINING_STATUS_UNTRAINED:
            return "untrained";
        case TRAINING_STATUS_SILENCED:
            return "silenced";
        default:
            return "unknown";
    }
}

__attribute__((unused)) static const char *
ml_training_result_to_string(enum ml_training_result tr)
{
    switch (tr) {
        case TRAINING_RESULT_OK:
            return "ok";
        case TRAINING_RESULT_INVALID_QUERY_TIME_RANGE:
            return "invalid-query";
        case TRAINING_RESULT_NOT_ENOUGH_COLLECTED_VALUES:
            return "missing-values";
        case TRAINING_RESULT_NULL_ACQUIRED_DIMENSION:
            return "null-acquired-dim";
        case TRAINING_RESULT_CHART_UNDER_REPLICATION:
            return "chart-under-replication";
        default:
            return "unknown";
    }
}

/*
 * Features
*/

// subtract elements that are `diff_n` positions apart
static void
ml_features_diff(ml_features_t *features)
{
    if (features->diff_n == 0)
        return;

    for (size_t idx = 0; idx != (features->src_n - features->diff_n); idx++) {
        size_t high = (features->src_n - 1) - idx;
        size_t low = high - features->diff_n;

        features->dst[low] = features->src[high] - features->src[low];
    }

    size_t n = features->src_n - features->diff_n;
    memcpy(features->src, features->dst, n * sizeof(calculated_number_t));

    for (size_t idx = features->src_n - features->diff_n; idx != features->src_n; idx++)
        features->src[idx] = 0.0;
}

// a function that computes the window average of an array inplace
static void
ml_features_smooth(ml_features_t *features)
{
    calculated_number_t sum = 0.0;

    size_t idx = 0;
    for (; idx != features->smooth_n - 1; idx++)
        sum += features->src[idx];

    for (; idx != (features->src_n - features->diff_n); idx++) {
        sum += features->src[idx];
        calculated_number_t prev_cn = features->src[idx - (features->smooth_n - 1)];
        features->src[idx - (features->smooth_n - 1)] = sum / features->smooth_n;
        sum -= prev_cn;
    }

    for (idx = 0; idx != features->smooth_n; idx++)
        features->src[(features->src_n - 1) - idx] = 0.0;
}

// create lag'd vectors out of the preprocessed buffer
static void
ml_features_lag(ml_features_t *features)
{
    size_t n = features->src_n - features->diff_n - features->smooth_n + 1 - features->lag_n;
    features->preprocessed_features.resize(n);

    unsigned target_num_samples = Cfg.max_train_samples * Cfg.random_sampling_ratio;
    double sampling_ratio = std::min(static_cast<double>(target_num_samples) / n, 1.0);

    uint32_t max_mt = std::numeric_limits<uint32_t>::max();
    uint32_t cutoff = static_cast<double>(max_mt) * sampling_ratio;

    size_t sample_idx = 0;

    for (size_t idx = 0; idx != n; idx++) {
        DSample &DS = features->preprocessed_features[sample_idx++];
        DS.set_size(features->lag_n);

        if (Cfg.random_nums[idx] > cutoff) {
            sample_idx--;
            continue;
        }

        for (size_t feature_idx = 0; feature_idx != features->lag_n + 1; feature_idx++)
            DS(feature_idx) = features->src[idx + feature_idx];
    }

    features->preprocessed_features.resize(sample_idx);
}

static void
ml_features_preprocess(ml_features_t *features)
{
    ml_features_diff(features);
    ml_features_smooth(features);
    ml_features_lag(features);
}

/*
 * KMeans
*/

static void
ml_kmeans_init(ml_kmeans_t *kmeans)
{
    kmeans->cluster_centers.reserve(2);
    kmeans->min_dist = std::numeric_limits<calculated_number_t>::max();
    kmeans->max_dist = std::numeric_limits<calculated_number_t>::min();
}

static void
ml_kmeans_train(ml_kmeans_t *kmeans, const ml_features_t *features, time_t after, time_t before)
{
    kmeans->after = (uint32_t) after;
    kmeans->before = (uint32_t) before;

    kmeans->min_dist = std::numeric_limits<calculated_number_t>::max();
    kmeans->max_dist  = std::numeric_limits<calculated_number_t>::min();

    kmeans->cluster_centers.clear();

    dlib::pick_initial_centers(2, kmeans->cluster_centers, features->preprocessed_features);
    dlib::find_clusters_using_kmeans(features->preprocessed_features, kmeans->cluster_centers, Cfg.max_kmeans_iters);

    for (const auto &preprocessed_feature : features->preprocessed_features) {
        calculated_number_t mean_dist = 0.0;

        for (const auto &cluster_center : kmeans->cluster_centers) {
            mean_dist += dlib::length(cluster_center - preprocessed_feature);
        }

        mean_dist /= kmeans->cluster_centers.size();

        if (mean_dist < kmeans->min_dist)
            kmeans->min_dist = mean_dist;

        if (mean_dist > kmeans->max_dist)
            kmeans->max_dist = mean_dist;
    }
}

static calculated_number_t
ml_kmeans_anomaly_score(const ml_kmeans_t *kmeans, const DSample &DS)
{
    calculated_number_t mean_dist = 0.0;
    for (const auto &CC: kmeans->cluster_centers)
        mean_dist += dlib::length(CC - DS);

    mean_dist /= kmeans->cluster_centers.size();

    if (kmeans->max_dist == kmeans->min_dist)
        return 0.0;

    calculated_number_t anomaly_score = 100.0 * std::abs((mean_dist - kmeans->min_dist) / (kmeans->max_dist - kmeans->min_dist));
    return (anomaly_score > 100.0) ? 100.0 : anomaly_score;
}

/*
 * Queue
*/

static ml_queue_t *
ml_queue_init()
{
    ml_queue_t *q = new ml_queue_t();

    netdata_mutex_init(&q->mutex);
    pthread_cond_init(&q->cond_var, NULL);
    q->exit = false;
    return q;
}

static void
ml_queue_destroy(ml_queue_t *q)
{
    netdata_mutex_destroy(&q->mutex);
    pthread_cond_destroy(&q->cond_var);
    delete q;
}

static void
ml_queue_push(ml_queue_t *q, const ml_training_request_t req)
{
    netdata_mutex_lock(&q->mutex);
    q->internal.push(req);
    pthread_cond_signal(&q->cond_var);
    netdata_mutex_unlock(&q->mutex);
}

static ml_training_request_t
ml_queue_pop(ml_queue_t *q)
{
    netdata_mutex_lock(&q->mutex);

    ml_training_request_t req = {
        {'\0'}, // machine_guid
        NULL, // chart id
        NULL, // dimension id
        0, // current time
        0, // first entry
        0  // last entry
    };

    while (q->internal.empty()) {
        pthread_cond_wait(&q->cond_var, &q->mutex);

        if (q->exit) {
            netdata_mutex_unlock(&q->mutex);

            // We return a dummy request because the queue has been signaled
            return req;
        }
    }

    req = q->internal.front();
    q->internal.pop();

    netdata_mutex_unlock(&q->mutex);
    return req;
}

static size_t
ml_queue_size(ml_queue_t *q)
{
    netdata_mutex_lock(&q->mutex);
    size_t size = q->internal.size();
    netdata_mutex_unlock(&q->mutex);
    return size;
}

static void
ml_queue_signal(ml_queue_t *q)
{
    netdata_mutex_lock(&q->mutex);
    q->exit = true;
    pthread_cond_signal(&q->cond_var);
    netdata_mutex_unlock(&q->mutex);
}

/*
 * Dimension
*/

static std::pair<calculated_number_t *, ml_training_response_t>
ml_dimension_calculated_numbers(ml_training_thread_t *training_thread, ml_dimension_t *dim, const ml_training_request_t &training_request)
{
    ml_training_response_t training_response = {};

    training_response.request_time = training_request.request_time;
    training_response.first_entry_on_request = training_request.first_entry_on_request;
    training_response.last_entry_on_request = training_request.last_entry_on_request;

    training_response.first_entry_on_response = rrddim_first_entry_s_of_tier(dim->rd, 0);
    training_response.last_entry_on_response = rrddim_last_entry_s_of_tier(dim->rd, 0);

    size_t min_n = Cfg.min_train_samples;
    size_t max_n = Cfg.max_train_samples;

    // Figure out what our time window should be.
    training_response.query_before_t = training_response.last_entry_on_response;
    training_response.query_after_t = std::max(
        training_response.query_before_t - static_cast<time_t>((max_n - 1) * dim->rd->rrdset->update_every),
        training_response.first_entry_on_response
    );

    if (training_response.query_after_t >= training_response.query_before_t) {
        training_response.result = TRAINING_RESULT_INVALID_QUERY_TIME_RANGE;
        return { NULL, training_response };
    }

    if (rrdset_is_replicating(dim->rd->rrdset)) {
        training_response.result = TRAINING_RESULT_CHART_UNDER_REPLICATION;
        return { NULL, training_response };
    }

    /*
     * Execute the query
    */
    struct storage_engine_query_handle handle;

    storage_engine_query_init(dim->rd->tiers[0].backend, dim->rd->tiers[0].db_metric_handle, &handle,
              training_response.query_after_t, training_response.query_before_t,
              STORAGE_PRIORITY_BEST_EFFORT);

    size_t idx = 0;
    memset(training_thread->training_cns, 0, sizeof(calculated_number_t) * max_n * (Cfg.lag_n + 1));
    calculated_number_t last_value = std::numeric_limits<calculated_number_t>::quiet_NaN();

    while (!storage_engine_query_is_finished(&handle)) {
        if (idx == max_n)
            break;

        STORAGE_POINT sp = storage_engine_query_next_metric(&handle);

        time_t timestamp = sp.end_time_s;
        calculated_number_t value = sp.sum / sp.count;

        if (netdata_double_isnumber(value)) {
            if (!training_response.db_after_t)
                training_response.db_after_t = timestamp;
            training_response.db_before_t = timestamp;

            training_thread->training_cns[idx] = value;
            last_value = training_thread->training_cns[idx];
            training_response.collected_values++;
        } else
            training_thread->training_cns[idx] = last_value;

        idx++;
    }
    storage_engine_query_finalize(&handle);

    global_statistics_ml_query_completed(/* points_read */ idx);

    training_response.total_values = idx;
    if (training_response.collected_values < min_n) {
        training_response.result = TRAINING_RESULT_NOT_ENOUGH_COLLECTED_VALUES;
        return { NULL, training_response };
    }

    // Find first non-NaN value.
    for (idx = 0; std::isnan(training_thread->training_cns[idx]); idx++, training_response.total_values--) { }

    // Overwrite NaN values.
    if (idx != 0)
        memmove(training_thread->training_cns, &training_thread->training_cns[idx], sizeof(calculated_number_t) * training_response.total_values);

    training_response.result = TRAINING_RESULT_OK;
    return { training_thread->training_cns, training_response };
}

const char *db_models_create_table =
    "CREATE TABLE IF NOT EXISTS models("
    "    dim_id BLOB, after INT, before INT,"
    "    min_dist REAL, max_dist REAL,"
    "    c00 REAL, c01 REAL, c02 REAL, c03 REAL, c04 REAL, c05 REAL,"
    "    c10 REAL, c11 REAL, c12 REAL, c13 REAL, c14 REAL, c15 REAL,"
    "    PRIMARY KEY(dim_id, after)"
    ");";

const char *db_models_add_model =
    "INSERT OR REPLACE INTO models("
    "    dim_id, after, before,"
    "    min_dist, max_dist,"
    "    c00, c01, c02, c03, c04, c05,"
    "    c10, c11, c12, c13, c14, c15)"
    "VALUES("
    "    @dim_id, @after, @before,"
    "    @min_dist, @max_dist,"
    "    @c00, @c01, @c02, @c03, @c04, @c05,"
    "    @c10, @c11, @c12, @c13, @c14, @c15);";

const char *db_models_load =
    "SELECT * FROM models "
    "WHERE dim_id = @dim_id AND after >= @after ORDER BY before ASC;";

const char *db_models_delete =
    "DELETE FROM models "
    "WHERE dim_id = @dim_id AND before < @before;";

static int
ml_dimension_add_model(const uuid_t *metric_uuid, const ml_kmeans_t *km)
{
    static __thread sqlite3_stmt *res = NULL;
    int param = 0;
    int rc = 0;

    if (unlikely(!db)) {
        error_report("Database has not been initialized");
        return 1;
    }

    if (unlikely(!res)) {
        rc = prepare_statement(db, db_models_add_model, &res);
        if (unlikely(rc != SQLITE_OK)) {
            error_report("Failed to prepare statement to store model, rc = %d", rc);
            return 1;
        }
    }

    rc = sqlite3_bind_blob(res, ++param, metric_uuid, sizeof(*metric_uuid), SQLITE_STATIC);
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    rc = sqlite3_bind_int(res, ++param, (int) km->after);
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    rc = sqlite3_bind_int(res, ++param, (int) km->before);
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    rc = sqlite3_bind_double(res, ++param, km->min_dist);
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    rc = sqlite3_bind_double(res, ++param, km->max_dist);
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    if (km->cluster_centers.size() != 2)
        fatal("Expected 2 cluster centers, got %zu", km->cluster_centers.size());

    for (const DSample &ds : km->cluster_centers) {
        if (ds.size() != 6)
            fatal("Expected dsample with 6 dimensions, got %ld", ds.size());

        for (long idx = 0; idx != ds.size(); idx++) {
            calculated_number_t cn = ds(idx);
            int rc = sqlite3_bind_double(res, ++param, cn);
            if (unlikely(rc != SQLITE_OK))
                goto bind_fail;
        }
    }

    rc = execute_insert(res);
    if (unlikely(rc != SQLITE_DONE)) {
        error_report("Failed to store model, rc = %d", rc);
        return rc;
    }

    rc = sqlite3_reset(res);
    if (unlikely(rc != SQLITE_OK)) {
        error_report("Failed to reset statement when storing model, rc = %d", rc);
        return rc;
    }

    return 0;

bind_fail:
    error_report("Failed to bind parameter %d to store model, rc = %d", param, rc);
    rc = sqlite3_reset(res);
    if (unlikely(rc != SQLITE_OK))
        error_report("Failed to reset statement to store model, rc = %d", rc);
    return rc;
}

static int
ml_dimension_delete_models(const uuid_t *metric_uuid, time_t before)
{
    static __thread sqlite3_stmt *res = NULL;
    int rc = 0;
    int param = 0;

    if (unlikely(!db)) {
        error_report("Database has not been initialized");
        return 1;
    }

    if (unlikely(!res)) {
        rc = prepare_statement(db, db_models_delete, &res);
        if (unlikely(rc != SQLITE_OK)) {
            error_report("Failed to prepare statement to delete models, rc = %d", rc);
            return rc;
        }
    }

    rc = sqlite3_bind_blob(res, ++param, metric_uuid, sizeof(*metric_uuid), SQLITE_STATIC);
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    rc = sqlite3_bind_int(res, ++param, (int) before);
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    rc = execute_insert(res);
    if (unlikely(rc != SQLITE_DONE)) {
        error_report("Failed to delete models, rc = %d", rc);
        return rc;
    }

    rc = sqlite3_reset(res);
    if (unlikely(rc != SQLITE_OK)) {
        error_report("Failed to reset statement when deleting models, rc = %d", rc);
        return rc;
    }

    return 0;

bind_fail:
    error_report("Failed to bind parameter %d to delete models, rc = %d", param, rc);
    rc = sqlite3_reset(res);
    if (unlikely(rc != SQLITE_OK))
        error_report("Failed to reset statement to delete models, rc = %d", rc);
    return rc;
}

int ml_dimension_load_models(RRDDIM *rd) {
    ml_dimension_t *dim = (ml_dimension_t *) rd->ml_dimension;
    if (!dim)
        return 0;

    spinlock_lock(&dim->slock);
    bool is_empty = dim->km_contexts.empty();
    spinlock_unlock(&dim->slock);

    if (!is_empty)
        return 0;

    std::vector<ml_kmeans_t> V;

    static __thread sqlite3_stmt *res = NULL;
    int rc = 0;
    int param = 0;

    if (unlikely(!db)) {
        error_report("Database has not been initialized");
        return 1;
    }

    if (unlikely(!res)) {
        rc = prepare_statement(db, db_models_load, &res);
        if (unlikely(rc != SQLITE_OK)) {
            error_report("Failed to prepare statement to load models, rc = %d", rc);
            return 1;
        }
    }

    rc = sqlite3_bind_blob(res, ++param, &dim->rd->metric_uuid, sizeof(dim->rd->metric_uuid), SQLITE_STATIC);
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    rc = sqlite3_bind_int(res, ++param, now_realtime_usec() - (Cfg.num_models_to_use * Cfg.max_train_samples));
    if (unlikely(rc != SQLITE_OK))
        goto bind_fail;

    spinlock_lock(&dim->slock);

    dim->km_contexts.reserve(Cfg.num_models_to_use);
    while ((rc = sqlite3_step_monitored(res)) == SQLITE_ROW) {
        ml_kmeans_t km;

        km.after = sqlite3_column_int(res, 2);
        km.before = sqlite3_column_int(res, 3);

        km.min_dist = sqlite3_column_int(res, 4);
        km.max_dist = sqlite3_column_int(res, 5);

        km.cluster_centers.resize(2);

        km.cluster_centers[0].set_size(Cfg.lag_n + 1);
        km.cluster_centers[0](0) = sqlite3_column_double(res, 6);
        km.cluster_centers[0](1) = sqlite3_column_double(res, 7);
        km.cluster_centers[0](2) = sqlite3_column_double(res, 8);
        km.cluster_centers[0](3) = sqlite3_column_double(res, 9);
        km.cluster_centers[0](4) = sqlite3_column_double(res, 10);
        km.cluster_centers[0](5) = sqlite3_column_double(res, 11);

        km.cluster_centers[1].set_size(Cfg.lag_n + 1);
        km.cluster_centers[1](0) = sqlite3_column_double(res, 12);
        km.cluster_centers[1](1) = sqlite3_column_double(res, 13);
        km.cluster_centers[1](2) = sqlite3_column_double(res, 14);
        km.cluster_centers[1](3) = sqlite3_column_double(res, 15);
        km.cluster_centers[1](4) = sqlite3_column_double(res, 16);
        km.cluster_centers[1](5) = sqlite3_column_double(res, 17);

        dim->km_contexts.push_back(km);
    }

    if (!dim->km_contexts.empty()) {
        dim->ts = TRAINING_STATUS_TRAINED;
    }

    spinlock_unlock(&dim->slock);

    if (unlikely(rc != SQLITE_DONE))
        error_report("Failed to load models, rc = %d", rc);

    rc = sqlite3_reset(res);
    if (unlikely(rc != SQLITE_OK))
        error_report("Failed to reset statement when loading models, rc = %d", rc);

    return 0;

bind_fail:
    error_report("Failed to bind parameter %d to load models, rc = %d", param, rc);
    rc = sqlite3_reset(res);
    if (unlikely(rc != SQLITE_OK))
        error_report("Failed to reset statement to load models, rc = %d", rc);
    return 1;
}

static enum ml_training_result
ml_dimension_train_model(ml_training_thread_t *training_thread, ml_dimension_t *dim, const ml_training_request_t &training_request)
{
    worker_is_busy(WORKER_TRAIN_QUERY);
    auto P = ml_dimension_calculated_numbers(training_thread, dim, training_request);
    ml_training_response_t training_response = P.second;

    if (training_response.result != TRAINING_RESULT_OK) {
        spinlock_lock(&dim->slock);

        dim->mt = METRIC_TYPE_CONSTANT;

        switch (dim->ts) {
            case TRAINING_STATUS_PENDING_WITH_MODEL:
                dim->ts = TRAINING_STATUS_TRAINED;
                break;
            case TRAINING_STATUS_PENDING_WITHOUT_MODEL:
                dim->ts = TRAINING_STATUS_UNTRAINED;
                break;
            default:
                break;
        }

        dim->suppression_anomaly_counter = 0;
        dim->suppression_window_counter = 0;
        dim->tr = training_response;

        dim->last_training_time = training_response.last_entry_on_response;
        enum ml_training_result result = training_response.result;

        spinlock_unlock(&dim->slock);

        return result;
    }

    // compute kmeans
    worker_is_busy(WORKER_TRAIN_KMEANS);
    {
        memcpy(training_thread->scratch_training_cns, training_thread->training_cns,
               training_response.total_values * sizeof(calculated_number_t));

        ml_features_t features = {
            Cfg.diff_n, Cfg.smooth_n, Cfg.lag_n,
            training_thread->scratch_training_cns, training_response.total_values,
            training_thread->training_cns, training_response.total_values,
            training_thread->training_samples
        };
        ml_features_preprocess(&features);

        ml_kmeans_init(&dim->kmeans);
        ml_kmeans_train(&dim->kmeans, &features, training_response.query_after_t, training_response.query_before_t);
    }

    // update models
    worker_is_busy(WORKER_TRAIN_UPDATE_MODELS);
    {
        spinlock_lock(&dim->slock);

        if (dim->km_contexts.size() < Cfg.num_models_to_use) {
            dim->km_contexts.push_back(std::move(dim->kmeans));
        } else {
            bool can_drop_middle_km = false;

            if (Cfg.num_models_to_use > 2) {
                const ml_kmeans_t *old_km = &dim->km_contexts[dim->km_contexts.size() - 1];
                const ml_kmeans_t *middle_km = &dim->km_contexts[dim->km_contexts.size() - 2];
                const ml_kmeans_t *new_km = &dim->kmeans;

                can_drop_middle_km = (middle_km->after < old_km->before) &&
                                     (middle_km->before > new_km->after);
            }

            if (can_drop_middle_km) {
                dim->km_contexts.back() = dim->kmeans;
            } else {
                std::rotate(std::begin(dim->km_contexts), std::begin(dim->km_contexts) + 1, std::end(dim->km_contexts));
                dim->km_contexts[dim->km_contexts.size() - 1] = std::move(dim->kmeans);
            }
        }

        dim->mt = METRIC_TYPE_CONSTANT;
        dim->ts = TRAINING_STATUS_TRAINED;

        dim->suppression_anomaly_counter = 0;
        dim->suppression_window_counter = 0;

        dim->tr = training_response;
        dim->last_training_time = rrddim_last_entry_s(dim->rd);

        // Add the newly generated model to the list of pending models to flush
        ml_model_info_t model_info;
        uuid_copy(model_info.metric_uuid, dim->rd->metric_uuid);
        model_info.kmeans = dim->km_contexts.back();
        training_thread->pending_model_info.push_back(model_info);

        spinlock_unlock(&dim->slock);
    }

    return training_response.result;
}

static void
ml_dimension_schedule_for_training(ml_dimension_t *dim, time_t curr_time)
{
    switch (dim->mt) {
    case METRIC_TYPE_CONSTANT:
        return;
    default:
        break;
    }

    bool schedule_for_training = false;

    switch (dim->ts) {
    case TRAINING_STATUS_PENDING_WITH_MODEL:
    case TRAINING_STATUS_PENDING_WITHOUT_MODEL:
        schedule_for_training = false;
        break;
    case TRAINING_STATUS_UNTRAINED:
        schedule_for_training = true;
        dim->ts = TRAINING_STATUS_PENDING_WITHOUT_MODEL;
        break;
    case TRAINING_STATUS_SILENCED:
    case TRAINING_STATUS_TRAINED:
        if ((dim->last_training_time + (Cfg.train_every * dim->rd->rrdset->update_every)) < curr_time) {
            schedule_for_training = true;
            dim->ts = TRAINING_STATUS_PENDING_WITH_MODEL;
        }
        break;
    }

    if (schedule_for_training) {
        ml_training_request_t req;

        memcpy(req.machine_guid, dim->rd->rrdset->rrdhost->machine_guid, GUID_LEN + 1);
        req.chart_id = string_dup(dim->rd->rrdset->id);
        req.dimension_id = string_dup(dim->rd->id);
        req.request_time = curr_time;
        req.first_entry_on_request = rrddim_first_entry_s(dim->rd);
        req.last_entry_on_request = rrddim_last_entry_s(dim->rd);

        ml_host_t *host = (ml_host_t *) dim->rd->rrdset->rrdhost->ml_host;
        ml_queue_push(host->training_queue, req);
    }
}

static bool
ml_dimension_predict(ml_dimension_t *dim, time_t curr_time, calculated_number_t value, bool exists)
{
    // Nothing to do if ML is disabled for this dimension
    if (dim->mls != MACHINE_LEARNING_STATUS_ENABLED)
        return false;

    // Don't treat values that don't exist as anomalous
    if (!exists) {
        dim->cns.clear();
        return false;
    }

    // Save the value and return if we don't have enough values for a sample
    unsigned n = Cfg.diff_n + Cfg.smooth_n + Cfg.lag_n;
    if (dim->cns.size() < n) {
        dim->cns.push_back(value);
        return false;
    }

    // Push the value and check if it's different from the last one
    bool same_value = true;
    std::rotate(std::begin(dim->cns), std::begin(dim->cns) + 1, std::end(dim->cns));
    if (dim->cns[n - 1] != value)
        same_value = false;
    dim->cns[n - 1] = value;

    // Create the sample
    assert((n * (Cfg.lag_n + 1) <= 128) &&
           "Static buffers too small to perform prediction. "
           "This should not be possible with the default clamping of feature extraction options");
    calculated_number_t src_cns[128];
    calculated_number_t dst_cns[128];

    memset(src_cns, 0, n * (Cfg.lag_n + 1) * sizeof(calculated_number_t));
    memcpy(src_cns, dim->cns.data(), n * sizeof(calculated_number_t));
    memcpy(dst_cns, dim->cns.data(), n * sizeof(calculated_number_t));

    ml_features_t features = {
        Cfg.diff_n, Cfg.smooth_n, Cfg.lag_n,
        dst_cns, n, src_cns, n,
        dim->feature
    };
    ml_features_preprocess(&features);

    /*
     * Lock to predict and possibly schedule the dimension for training
    */
    if (spinlock_trylock(&dim->slock) == 0)
        return false;

    // Mark the metric time as variable if we received different values
    if (!same_value)
        dim->mt = METRIC_TYPE_VARIABLE;

    // Decide if the dimension needs to be scheduled for training
    ml_dimension_schedule_for_training(dim, curr_time);

    // Nothing to do if we don't have a model
    switch (dim->ts) {
        case TRAINING_STATUS_UNTRAINED:
        case TRAINING_STATUS_PENDING_WITHOUT_MODEL: {
        case TRAINING_STATUS_SILENCED:
            spinlock_unlock(&dim->slock);
            return false;
        }
        default:
            break;
    }

    dim->suppression_window_counter++;

    /*
     * Use the KMeans models to check if the value is anomalous
    */

    size_t sum = 0;
    size_t models_consulted = 0;

    for (const auto &km_ctx : dim->km_contexts) {
        models_consulted++;

        calculated_number_t anomaly_score = ml_kmeans_anomaly_score(&km_ctx, features.preprocessed_features[0]);
        if (anomaly_score == std::numeric_limits<calculated_number_t>::quiet_NaN())
            continue;

        if (anomaly_score < (100 * Cfg.dimension_anomaly_score_threshold)) {
            global_statistics_ml_models_consulted(models_consulted);
            spinlock_unlock(&dim->slock);
            return false;
        }

        sum += 1;
    }

    dim->suppression_anomaly_counter += sum ? 1 : 0;

    if ((dim->suppression_anomaly_counter >= Cfg.suppression_threshold) &&
        (dim->suppression_window_counter >= Cfg.suppression_window)) {
        dim->ts = TRAINING_STATUS_SILENCED;
    }

    spinlock_unlock(&dim->slock);

    global_statistics_ml_models_consulted(models_consulted);
    return sum;
}

/*
 * Chart
*/

static bool
ml_chart_is_available_for_ml(ml_chart_t *chart)
{
    return rrdset_is_available_for_exporting_and_alarms(chart->rs);
}

void
ml_chart_update_dimension(ml_chart_t *chart, ml_dimension_t *dim, bool is_anomalous)
{
    switch (dim->mls) {
        case MACHINE_LEARNING_STATUS_DISABLED_DUE_TO_EXCLUDED_CHART:
            chart->mls.num_machine_learning_status_disabled_sp++;
            return;
        case MACHINE_LEARNING_STATUS_ENABLED: {
            chart->mls.num_machine_learning_status_enabled++;

            switch (dim->mt) {
                case METRIC_TYPE_CONSTANT:
                    chart->mls.num_metric_type_constant++;
                    chart->mls.num_training_status_trained++;
                    chart->mls.num_normal_dimensions++;
                    return;
                case METRIC_TYPE_VARIABLE:
                    chart->mls.num_metric_type_variable++;
                    break;
            }

            switch (dim->ts) {
                case TRAINING_STATUS_UNTRAINED:
                    chart->mls.num_training_status_untrained++;
                    return;
                case TRAINING_STATUS_PENDING_WITHOUT_MODEL:
                    chart->mls.num_training_status_pending_without_model++;
                    return;
                case TRAINING_STATUS_TRAINED:
                    chart->mls.num_training_status_trained++;

                    chart->mls.num_anomalous_dimensions += is_anomalous;
                    chart->mls.num_normal_dimensions += !is_anomalous;
                    return;
                case TRAINING_STATUS_PENDING_WITH_MODEL:
                    chart->mls.num_training_status_pending_with_model++;

                    chart->mls.num_anomalous_dimensions += is_anomalous;
                    chart->mls.num_normal_dimensions += !is_anomalous;
                    return;
                case TRAINING_STATUS_SILENCED:
                    chart->mls.num_training_status_silenced++;
                    chart->mls.num_training_status_trained++;

                    chart->mls.num_anomalous_dimensions += is_anomalous;
                    chart->mls.num_normal_dimensions += !is_anomalous;
                    return;
            }

            return;
        }
    }
}

/*
 * Host detection & training functions
*/

#define WORKER_JOB_DETECTION_COLLECT_STATS 0
#define WORKER_JOB_DETECTION_DIM_CHART 1
#define WORKER_JOB_DETECTION_HOST_CHART 2
#define WORKER_JOB_DETECTION_STATS 3

static void
ml_host_detect_once(ml_host_t *host)
{
    worker_is_busy(WORKER_JOB_DETECTION_COLLECT_STATS);

    host->mls = {};
    ml_machine_learning_stats_t mls_copy = {};

    if (host->ml_running) {
        netdata_mutex_lock(&host->mutex);

        /*
         * prediction/detection stats
        */
        void *rsp = NULL;
        rrdset_foreach_read(rsp, host->rh) {
            RRDSET *rs = static_cast<RRDSET *>(rsp);

            ml_chart_t *chart = (ml_chart_t *) rs->ml_chart;
            if (!chart)
                continue;

            if (!ml_chart_is_available_for_ml(chart))
                continue;

            ml_machine_learning_stats_t chart_mls = chart->mls;

            host->mls.num_machine_learning_status_enabled += chart_mls.num_machine_learning_status_enabled;
            host->mls.num_machine_learning_status_disabled_sp += chart_mls.num_machine_learning_status_disabled_sp;

            host->mls.num_metric_type_constant += chart_mls.num_metric_type_constant;
            host->mls.num_metric_type_variable += chart_mls.num_metric_type_variable;

            host->mls.num_training_status_untrained += chart_mls.num_training_status_untrained;
            host->mls.num_training_status_pending_without_model += chart_mls.num_training_status_pending_without_model;
            host->mls.num_training_status_trained += chart_mls.num_training_status_trained;
            host->mls.num_training_status_pending_with_model += chart_mls.num_training_status_pending_with_model;
            host->mls.num_training_status_silenced += chart_mls.num_training_status_silenced;

            host->mls.num_anomalous_dimensions += chart_mls.num_anomalous_dimensions;
            host->mls.num_normal_dimensions += chart_mls.num_normal_dimensions;
        }
        rrdset_foreach_done(rsp);

        host->host_anomaly_rate = 0.0;
        size_t NumActiveDimensions = host->mls.num_anomalous_dimensions + host->mls.num_normal_dimensions;
        if (NumActiveDimensions)
              host->host_anomaly_rate = static_cast<double>(host->mls.num_anomalous_dimensions) / NumActiveDimensions;

        mls_copy = host->mls;

        netdata_mutex_unlock(&host->mutex);
    } else {
        host->host_anomaly_rate = 0.0;
    }

    worker_is_busy(WORKER_JOB_DETECTION_DIM_CHART);
    ml_update_dimensions_chart(host, mls_copy);

    worker_is_busy(WORKER_JOB_DETECTION_HOST_CHART);
    ml_update_host_and_detection_rate_charts(host, host->host_anomaly_rate * 10000.0);
}

typedef struct {
    RRDHOST_ACQUIRED *acq_rh;
    RRDSET_ACQUIRED *acq_rs;
    RRDDIM_ACQUIRED *acq_rd;
    ml_dimension_t *dim;
} ml_acquired_dimension_t;

static ml_acquired_dimension_t
ml_acquired_dimension_get(char *machine_guid, STRING *chart_id, STRING *dimension_id)
{
    RRDHOST_ACQUIRED *acq_rh = NULL;
    RRDSET_ACQUIRED *acq_rs = NULL;
    RRDDIM_ACQUIRED *acq_rd = NULL;
    ml_dimension_t *dim = NULL;

    rrd_rdlock();

    acq_rh = rrdhost_find_and_acquire(machine_guid);
    if (acq_rh) {
        RRDHOST *rh = rrdhost_acquired_to_rrdhost(acq_rh);
        if (rh && !rrdhost_flag_check(rh, RRDHOST_FLAG_ORPHAN | RRDHOST_FLAG_ARCHIVED)) {
            acq_rs = rrdset_find_and_acquire(rh, string2str(chart_id));
            if (acq_rs) {
                RRDSET *rs = rrdset_acquired_to_rrdset(acq_rs);
                if (rs && !rrdset_flag_check(rs, RRDSET_FLAG_ARCHIVED | RRDSET_FLAG_OBSOLETE)) {
                    acq_rd = rrddim_find_and_acquire(rs, string2str(dimension_id));
                    if (acq_rd) {
                        RRDDIM *rd = rrddim_acquired_to_rrddim(acq_rd);
                        if (rd)
                            dim = (ml_dimension_t *) rd->ml_dimension;
                    }
                }
            }
        }
    }

    rrd_unlock();

    ml_acquired_dimension_t acq_dim = {
        acq_rh, acq_rs, acq_rd, dim
    };

    return acq_dim;
}

static void
ml_acquired_dimension_release(ml_acquired_dimension_t acq_dim)
{
    if (acq_dim.acq_rd)
        rrddim_acquired_release(acq_dim.acq_rd);

    if (acq_dim.acq_rs)
        rrdset_acquired_release(acq_dim.acq_rs);

    if (acq_dim.acq_rh)
        rrdhost_acquired_release(acq_dim.acq_rh);
}

static enum ml_training_result
ml_acquired_dimension_train(ml_training_thread_t *training_thread, ml_acquired_dimension_t acq_dim, const ml_training_request_t &tr)
{
    if (!acq_dim.dim)
        return TRAINING_RESULT_NULL_ACQUIRED_DIMENSION;

    return ml_dimension_train_model(training_thread, acq_dim.dim, tr);
}

static void *
ml_detect_main(void *arg)
{
    UNUSED(arg);

    worker_register("MLDETECT");
    worker_register_job_name(WORKER_JOB_DETECTION_COLLECT_STATS, "collect stats");
    worker_register_job_name(WORKER_JOB_DETECTION_DIM_CHART, "dim chart");
    worker_register_job_name(WORKER_JOB_DETECTION_HOST_CHART, "host chart");
    worker_register_job_name(WORKER_JOB_DETECTION_STATS, "training stats");

    heartbeat_t hb;
    heartbeat_init(&hb);

    while (!Cfg.detection_stop) {
        worker_is_idle();
        heartbeat_next(&hb, USEC_PER_SEC);

        RRDHOST *rh;
        rrd_rdlock();
        rrdhost_foreach_read(rh) {
            if (!rh->ml_host)
                continue;

            ml_host_detect_once((ml_host_t *) rh->ml_host);
        }
        rrd_unlock();

        if (Cfg.enable_statistics_charts) {
            // collect and update training thread stats
            for (size_t idx = 0; idx != Cfg.num_training_threads; idx++) {
                ml_training_thread_t *training_thread = &Cfg.training_threads[idx];

                netdata_mutex_lock(&training_thread->nd_mutex);
                ml_training_stats_t training_stats = training_thread->training_stats;
                training_thread->training_stats = {};
                netdata_mutex_unlock(&training_thread->nd_mutex);

                // calc the avg values
                if (training_stats.num_popped_items) {
                    training_stats.queue_size /= training_stats.num_popped_items;
                    training_stats.allotted_ut /= training_stats.num_popped_items;
                    training_stats.consumed_ut /= training_stats.num_popped_items;
                    training_stats.remaining_ut /= training_stats.num_popped_items;
                } else {
                    training_stats.queue_size = ml_queue_size(training_thread->training_queue);
                    training_stats.consumed_ut = 0;
                    training_stats.remaining_ut = training_stats.allotted_ut;

                    training_stats.training_result_ok = 0;
                    training_stats.training_result_invalid_query_time_range = 0;
                    training_stats.training_result_not_enough_collected_values = 0;
                    training_stats.training_result_null_acquired_dimension = 0;
                    training_stats.training_result_chart_under_replication = 0;
                }

                ml_update_training_statistics_chart(training_thread, training_stats);
            }
        }
    }

    return NULL;
}

/*
 * Public API
*/

bool ml_capable()
{
    return true;
}

bool ml_enabled(RRDHOST *rh)
{
    if (!rh)
        return false;
    
    if (!Cfg.enable_anomaly_detection)
        return false;

    if (simple_pattern_matches(Cfg.sp_host_to_skip, rrdhost_hostname(rh)))
        return false;

    return true;
}

bool ml_streaming_enabled()
{
    return Cfg.stream_anomaly_detection_charts;
}

void ml_host_new(RRDHOST *rh)
{
    if (!ml_enabled(rh))
        return;

    ml_host_t *host = new ml_host_t();

    host->rh = rh;
    host->mls = ml_machine_learning_stats_t();
    host->host_anomaly_rate = 0.0;

    static std::atomic<size_t> times_called(0);
    host->training_queue = Cfg.training_threads[times_called++ % Cfg.num_training_threads].training_queue;

    netdata_mutex_init(&host->mutex);

    host->ml_running = true;
    rh->ml_host = (rrd_ml_host_t *) host;
}

void ml_host_delete(RRDHOST *rh)
{
    ml_host_t *host = (ml_host_t *) rh->ml_host;
    if (!host)
        return;

    netdata_mutex_destroy(&host->mutex);

    delete host;
    rh->ml_host = NULL;
}

void ml_host_start(RRDHOST *rh) {
    ml_host_t *host = (ml_host_t *) rh->ml_host;
    if (!host)
        return;

    host->ml_running = true;
}

void ml_host_stop(RRDHOST *rh) {
    ml_host_t *host = (ml_host_t *) rh->ml_host;
    if (!host || !host->ml_running)
        return;

    netdata_mutex_lock(&host->mutex);

    // reset host stats
    host->mls = ml_machine_learning_stats_t();

    // reset charts/dims
    void *rsp = NULL;
    rrdset_foreach_read(rsp, host->rh) {
        RRDSET *rs = static_cast<RRDSET *>(rsp);

        ml_chart_t *chart = (ml_chart_t *) rs->ml_chart;
        if (!chart)
            continue;

        // reset chart
        chart->mls = ml_machine_learning_stats_t();

        void *rdp = NULL;
        rrddim_foreach_read(rdp, rs) {
            RRDDIM *rd = static_cast<RRDDIM *>(rdp);

            ml_dimension_t *dim = (ml_dimension_t *) rd->ml_dimension;
            if (!dim)
                continue;

            spinlock_lock(&dim->slock);

            // reset dim
            // TODO: should we drop in-mem models, or mark them as stale? Is it
            // okay to resume training straight away?

            dim->mt = METRIC_TYPE_CONSTANT;
            dim->ts = TRAINING_STATUS_UNTRAINED;
            dim->last_training_time = 0;
            dim->suppression_anomaly_counter = 0;
            dim->suppression_window_counter = 0;
            dim->cns.clear();

            ml_kmeans_init(&dim->kmeans);

            spinlock_unlock(&dim->slock);
        }
        rrddim_foreach_done(rdp);
    }
    rrdset_foreach_done(rsp);

    netdata_mutex_unlock(&host->mutex);

    host->ml_running = false;
}

void ml_host_get_info(RRDHOST *rh, BUFFER *wb)
{
    ml_host_t *host = (ml_host_t *) rh->ml_host;
    if (!host) {
        buffer_json_member_add_boolean(wb, "enabled", false);
        return;
    }

    buffer_json_member_add_uint64(wb, "version", 1);

    buffer_json_member_add_boolean(wb, "enabled", Cfg.enable_anomaly_detection);

    buffer_json_member_add_uint64(wb, "min-train-samples", Cfg.min_train_samples);
    buffer_json_member_add_uint64(wb, "max-train-samples", Cfg.max_train_samples);
    buffer_json_member_add_uint64(wb, "train-every", Cfg.train_every);

    buffer_json_member_add_uint64(wb, "diff-n", Cfg.diff_n);
    buffer_json_member_add_uint64(wb, "smooth-n", Cfg.smooth_n);
    buffer_json_member_add_uint64(wb, "lag-n", Cfg.lag_n);

    buffer_json_member_add_double(wb, "random-sampling-ratio", Cfg.random_sampling_ratio);
    buffer_json_member_add_uint64(wb, "max-kmeans-iters", Cfg.random_sampling_ratio);

    buffer_json_member_add_double(wb, "dimension-anomaly-score-threshold", Cfg.dimension_anomaly_score_threshold);

    buffer_json_member_add_string(wb, "anomaly-detection-grouping-method",
                                  time_grouping_method2string(Cfg.anomaly_detection_grouping_method));

    buffer_json_member_add_int64(wb, "anomaly-detection-query-duration", Cfg.anomaly_detection_query_duration);

    buffer_json_member_add_string(wb, "hosts-to-skip", Cfg.hosts_to_skip.c_str());
    buffer_json_member_add_string(wb, "charts-to-skip", Cfg.charts_to_skip.c_str());
}

void ml_host_get_detection_info(RRDHOST *rh, BUFFER *wb)
{
    ml_host_t *host = (ml_host_t *) rh->ml_host;
    if (!host)
        return;

    netdata_mutex_lock(&host->mutex);

    buffer_json_member_add_uint64(wb, "version", 2);
    buffer_json_member_add_uint64(wb, "ml-running", host->ml_running);
    buffer_json_member_add_uint64(wb, "anomalous-dimensions", host->mls.num_anomalous_dimensions);
    buffer_json_member_add_uint64(wb, "normal-dimensions", host->mls.num_normal_dimensions);
    buffer_json_member_add_uint64(wb, "total-dimensions", host->mls.num_anomalous_dimensions +
                                                          host->mls.num_normal_dimensions);
    buffer_json_member_add_uint64(wb, "trained-dimensions", host->mls.num_training_status_trained +
                                                            host->mls.num_training_status_pending_with_model);
    netdata_mutex_unlock(&host->mutex);
}

bool ml_host_get_host_status(RRDHOST *rh, struct ml_metrics_statistics *mlm) {
    ml_host_t *host = (ml_host_t *) rh->ml_host;
    if (!host) {
        memset(mlm, 0, sizeof(*mlm));
        return false;
    }

    netdata_mutex_lock(&host->mutex);

    mlm->anomalous = host->mls.num_anomalous_dimensions;
    mlm->normal = host->mls.num_normal_dimensions;
    mlm->trained = host->mls.num_training_status_trained + host->mls.num_training_status_pending_with_model;
    mlm->pending = host->mls.num_training_status_untrained + host->mls.num_training_status_pending_without_model;
    mlm->silenced = host->mls.num_training_status_silenced;

    netdata_mutex_unlock(&host->mutex);

    return true;
}

bool ml_host_running(RRDHOST *rh) {
    ml_host_t *host = (ml_host_t *) rh->ml_host;
    if(!host)
        return false;

    return true;
}

void ml_host_get_models(RRDHOST *rh, BUFFER *wb)
{
    UNUSED(rh);
    UNUSED(wb);

    // TODO: To be implemented
    netdata_log_error("Fetching KMeans models is not supported yet");
}

void ml_chart_new(RRDSET *rs)
{
    ml_host_t *host = (ml_host_t *) rs->rrdhost->ml_host;
    if (!host)
        return;

    ml_chart_t *chart = new ml_chart_t();

    chart->rs = rs;
    chart->mls = ml_machine_learning_stats_t();

    rs->ml_chart = (rrd_ml_chart_t *) chart;
}

void ml_chart_delete(RRDSET *rs)
{
    ml_host_t *host = (ml_host_t *) rs->rrdhost->ml_host;
    if (!host)
        return;

    ml_chart_t *chart = (ml_chart_t *) rs->ml_chart;

    delete chart;
    rs->ml_chart = NULL;
}

bool ml_chart_update_begin(RRDSET *rs)
{
    ml_chart_t *chart = (ml_chart_t *) rs->ml_chart;
    if (!chart)
        return false;

    chart->mls = {};
    return true;
}

void ml_chart_update_end(RRDSET *rs)
{
    ml_chart_t *chart = (ml_chart_t *) rs->ml_chart;
    if (!chart)
        return;
}

void ml_dimension_new(RRDDIM *rd)
{
    ml_chart_t *chart = (ml_chart_t *) rd->rrdset->ml_chart;
    if (!chart)
        return;

    ml_dimension_t *dim = new ml_dimension_t();

    dim->rd = rd;

    dim->mt = METRIC_TYPE_CONSTANT;
    dim->ts = TRAINING_STATUS_UNTRAINED;
    dim->last_training_time = 0;
    dim->suppression_anomaly_counter = 0;
    dim->suppression_window_counter = 0;

    ml_kmeans_init(&dim->kmeans);

    if (simple_pattern_matches(Cfg.sp_charts_to_skip, rrdset_name(rd->rrdset)))
        dim->mls = MACHINE_LEARNING_STATUS_DISABLED_DUE_TO_EXCLUDED_CHART;
    else
        dim->mls = MACHINE_LEARNING_STATUS_ENABLED;

    spinlock_init(&dim->slock);

    dim->km_contexts.reserve(Cfg.num_models_to_use);

    rd->ml_dimension = (rrd_ml_dimension_t *) dim;

    metaqueue_ml_load_models(rd);
}

void ml_dimension_delete(RRDDIM *rd)
{
    ml_dimension_t *dim = (ml_dimension_t *) rd->ml_dimension;
    if (!dim)
        return;

    delete dim;
    rd->ml_dimension = NULL;
}

bool ml_dimension_is_anomalous(RRDDIM *rd, time_t curr_time, double value, bool exists)
{
    ml_dimension_t *dim = (ml_dimension_t *) rd->ml_dimension;
    if (!dim)
        return false;

    ml_host_t *host = (ml_host_t *) rd->rrdset->rrdhost->ml_host;
    if (!host->ml_running)
        return false;

    ml_chart_t *chart = (ml_chart_t *) rd->rrdset->ml_chart;

    bool is_anomalous = ml_dimension_predict(dim, curr_time, value, exists);
    ml_chart_update_dimension(chart, dim, is_anomalous);

    return is_anomalous;
}

static void ml_flush_pending_models(ml_training_thread_t *training_thread) {
    int rc = db_execute(db, "BEGIN TRANSACTION;");
    int op_no = 1;

    if (!rc) {
        op_no++;

        for (const auto &pending_model: training_thread->pending_model_info) {
            if (!rc)
                rc = ml_dimension_add_model(&pending_model.metric_uuid, &pending_model.kmeans);

            if (!rc)
                rc = ml_dimension_delete_models(&pending_model.metric_uuid, pending_model.kmeans.before - (Cfg.num_models_to_use * Cfg.train_every));
        }
    }

    if (!rc) {
        op_no++;
        rc = db_execute(db, "COMMIT TRANSACTION;");
    }

    // try to rollback transaction if we got any failures
    if (rc) {
        netdata_log_error("Trying to rollback ML transaction because it failed with rc=%d, op_no=%d", rc, op_no);
        op_no++;
        rc = db_execute(db, "ROLLBACK;");
        if (rc)
            netdata_log_error("ML transaction rollback failed with rc=%d", rc);
    }

    training_thread->pending_model_info.clear();
}

static void *ml_train_main(void *arg) {
    ml_training_thread_t *training_thread = (ml_training_thread_t *) arg;

    char worker_name[1024];
    snprintfz(worker_name, 1024, "training_thread_%zu", training_thread->id);
    worker_register("MLTRAIN");

    worker_register_job_name(WORKER_TRAIN_QUEUE_POP, "pop queue");
    worker_register_job_name(WORKER_TRAIN_ACQUIRE_DIMENSION, "acquire");
    worker_register_job_name(WORKER_TRAIN_QUERY, "query");
    worker_register_job_name(WORKER_TRAIN_KMEANS, "kmeans");
    worker_register_job_name(WORKER_TRAIN_UPDATE_MODELS, "update models");
    worker_register_job_name(WORKER_TRAIN_RELEASE_DIMENSION, "release");
    worker_register_job_name(WORKER_TRAIN_UPDATE_HOST, "update host");
    worker_register_job_name(WORKER_TRAIN_FLUSH_MODELS, "flush models");

    while (!Cfg.training_stop) {
        worker_is_busy(WORKER_TRAIN_QUEUE_POP);

        ml_training_request_t training_req = ml_queue_pop(training_thread->training_queue);

        // we know this thread has been cancelled, when the queue starts
        // returning "null" requests without blocking on queue's pop().
        if (training_req.chart_id == NULL)
            break;

        size_t queue_size = ml_queue_size(training_thread->training_queue) + 1;

        usec_t allotted_ut = (Cfg.train_every * USEC_PER_SEC) / queue_size;
        if (allotted_ut > USEC_PER_SEC)
            allotted_ut = USEC_PER_SEC;

        usec_t start_ut = now_monotonic_usec();

        enum ml_training_result training_res;
        {
            worker_is_busy(WORKER_TRAIN_ACQUIRE_DIMENSION);
            ml_acquired_dimension_t acq_dim = ml_acquired_dimension_get(
                training_req.machine_guid,
                training_req.chart_id,
                training_req.dimension_id);

            training_res = ml_acquired_dimension_train(training_thread, acq_dim, training_req);

            string_freez(training_req.chart_id);
            string_freez(training_req.dimension_id);

            worker_is_busy(WORKER_TRAIN_RELEASE_DIMENSION);
            ml_acquired_dimension_release(acq_dim);
        }

        usec_t consumed_ut = now_monotonic_usec() - start_ut;

        usec_t remaining_ut = 0;
        if (consumed_ut < allotted_ut)
            remaining_ut = allotted_ut - consumed_ut;

        if (Cfg.enable_statistics_charts) {
            worker_is_busy(WORKER_TRAIN_UPDATE_HOST);

            netdata_mutex_lock(&training_thread->nd_mutex);

            training_thread->training_stats.queue_size += queue_size;
            training_thread->training_stats.num_popped_items += 1;

            training_thread->training_stats.allotted_ut += allotted_ut;
            training_thread->training_stats.consumed_ut += consumed_ut;
            training_thread->training_stats.remaining_ut += remaining_ut;

            switch (training_res) {
                case TRAINING_RESULT_OK:
                    training_thread->training_stats.training_result_ok += 1;
                    break;
                case TRAINING_RESULT_INVALID_QUERY_TIME_RANGE:
                    training_thread->training_stats.training_result_invalid_query_time_range += 1;
                    break;
                case TRAINING_RESULT_NOT_ENOUGH_COLLECTED_VALUES:
                    training_thread->training_stats.training_result_not_enough_collected_values += 1;
                    break;
                case TRAINING_RESULT_NULL_ACQUIRED_DIMENSION:
                    training_thread->training_stats.training_result_null_acquired_dimension += 1;
                    break;
                case TRAINING_RESULT_CHART_UNDER_REPLICATION:
                    training_thread->training_stats.training_result_chart_under_replication += 1;
                    break;
            }

            netdata_mutex_unlock(&training_thread->nd_mutex);
        }

        if (training_thread->pending_model_info.size() >= Cfg.flush_models_batch_size) {
            worker_is_busy(WORKER_TRAIN_FLUSH_MODELS);
            netdata_mutex_lock(&db_mutex);
            ml_flush_pending_models(training_thread);
            netdata_mutex_unlock(&db_mutex);
            continue;
        }

        worker_is_idle();
        std::this_thread::sleep_for(std::chrono::microseconds{remaining_ut});
    }

    return NULL;
}

void ml_init()
{
    // Read config values
    ml_config_load(&Cfg);

    if (!Cfg.enable_anomaly_detection)
        return;

    // Generate random numbers to efficiently sample the features we need
    // for KMeans clustering.
    std::random_device RD;
    std::mt19937 Gen(RD());

    Cfg.random_nums.reserve(Cfg.max_train_samples);
    for (size_t Idx = 0; Idx != Cfg.max_train_samples; Idx++)
        Cfg.random_nums.push_back(Gen());

    // init training thread-specific data
    Cfg.training_threads.resize(Cfg.num_training_threads);
    for (size_t idx = 0; idx != Cfg.num_training_threads; idx++) {
        ml_training_thread_t *training_thread = &Cfg.training_threads[idx];

        size_t max_elements_needed_for_training = (size_t) Cfg.max_train_samples * (size_t) (Cfg.lag_n + 1);
        training_thread->training_cns = new calculated_number_t[max_elements_needed_for_training]();
        training_thread->scratch_training_cns = new calculated_number_t[max_elements_needed_for_training]();

        training_thread->id = idx;
        training_thread->training_queue = ml_queue_init();
        training_thread->pending_model_info.reserve(Cfg.flush_models_batch_size);
        netdata_mutex_init(&training_thread->nd_mutex);
    }

    // open sqlite db
    char path[FILENAME_MAX];
    snprintfz(path, FILENAME_MAX - 1, "%s/%s", netdata_configured_cache_dir, "ml.db");
    int rc = sqlite3_open(path, &db);
    if (rc != SQLITE_OK) {
        error_report("Failed to initialize database at %s, due to \"%s\"", path, sqlite3_errstr(rc));
        sqlite3_close(db);
        db = NULL;
    }

    if (db) {
        char *err = NULL;
        int rc = sqlite3_exec(db, db_models_create_table, NULL, NULL, &err);
        if (rc != SQLITE_OK) {
            error_report("Failed to create models table (%s, %s)", sqlite3_errstr(rc), err ? err : "");
            sqlite3_close(db);
            sqlite3_free(err);
            db = NULL;
        }
    }
}

void ml_fini() {
    if (!Cfg.enable_anomaly_detection)
        return;

    int rc = sqlite3_close_v2(db);
    if (unlikely(rc != SQLITE_OK))
        error_report("Error %d while closing the SQLite database, %s", rc, sqlite3_errstr(rc));
}

void ml_start_threads() {
    if (!Cfg.enable_anomaly_detection)
        return;

    // start detection & training threads
    Cfg.detection_stop = false;
    Cfg.training_stop = false;

    char tag[NETDATA_THREAD_TAG_MAX + 1];

    snprintfz(tag, NETDATA_THREAD_TAG_MAX, "%s", "PREDICT");
    netdata_thread_create(&Cfg.detection_thread, tag, NETDATA_THREAD_OPTION_JOINABLE, ml_detect_main, NULL);

    for (size_t idx = 0; idx != Cfg.num_training_threads; idx++) {
        ml_training_thread_t *training_thread = &Cfg.training_threads[idx];
        snprintfz(tag, NETDATA_THREAD_TAG_MAX, "TRAIN[%zu]", training_thread->id);
        netdata_thread_create(&training_thread->nd_thread, tag, NETDATA_THREAD_OPTION_JOINABLE, ml_train_main, training_thread);
    }
}

void ml_stop_threads()
{
    if (!Cfg.enable_anomaly_detection)
        return;

    Cfg.detection_stop = true;
    Cfg.training_stop = true;

    if (!Cfg.detection_thread)
        return;

    netdata_thread_cancel(Cfg.detection_thread);
    netdata_thread_join(Cfg.detection_thread, NULL);

    // signal the training queue of each thread
    for (size_t idx = 0; idx != Cfg.num_training_threads; idx++) {
        ml_training_thread_t *training_thread = &Cfg.training_threads[idx];

        ml_queue_signal(training_thread->training_queue);
    }

    // cancel training threads
    for (size_t idx = 0; idx != Cfg.num_training_threads; idx++) {
        ml_training_thread_t *training_thread = &Cfg.training_threads[idx];

        netdata_thread_cancel(training_thread->nd_thread);
    }

    // join training threads
    for (size_t idx = 0; idx != Cfg.num_training_threads; idx++) {
        ml_training_thread_t *training_thread = &Cfg.training_threads[idx];

        netdata_thread_join(training_thread->nd_thread, NULL);
    }

    // clear training thread data
    for (size_t idx = 0; idx != Cfg.num_training_threads; idx++) {
        ml_training_thread_t *training_thread = &Cfg.training_threads[idx];

        delete[] training_thread->training_cns;
        delete[] training_thread->scratch_training_cns;
        ml_queue_destroy(training_thread->training_queue);
        netdata_mutex_destroy(&training_thread->nd_mutex);
    }
}