tsio.py 24.5 KB
Newer Older
1
from datetime import datetime
2
import logging
3
4
5

import pandas as pd

6
7
from sqlalchemy import Table, Column, Integer, ForeignKey, Index
from sqlalchemy.sql.elements import NONE_NAME
8
from sqlalchemy.engine.base import Engine
Aurélien Campéas's avatar
Aurélien Campéas committed
9
from sqlalchemy.sql.expression import select, func, desc
Aurélien Campéas's avatar
Aurélien Campéas committed
10
from sqlalchemy.dialects.postgresql import TIMESTAMP
11

12
from tshistory.schema import tsschema
13
from tshistory.util import (
14
    closed_overlaps,
15
    num2float,
16
    subset,
17
    SeriesServices,
18
    start_end,
19
20
    tzaware_serie
)
21
from tshistory.snapshot import Snapshot, TABLES as SNAPTABLES
22

23
L = logging.getLogger('tshistory.tsio')
24
TABLES = {}
25
26


27
class TimeSerie(SeriesServices):
28
    namespace = 'tsh'
29
    schema = None
30
    metadatacache = None
31
32
33
34
35
36
37
38
    metakeys = {
        'tzaware',
        'index_type',
        'index_names',
        'index_dtype',
        'value_dtype',
        'value_type'
    }
39
    registry_map = None
40
    serie_tablename = None
41
    create_lock_id = None
42
43
44

    def __init__(self, namespace='tsh'):
        self.namespace = namespace
45
46
        self.schema = tsschema(namespace)
        self.schema.define()
47
        self.metadatacache = {}
48
        self.registry_map = {}
49
        self.serie_tablename = {}
50
        self.create_lock_id = sum(ord(c) for c in namespace)
51

52
53
54
55
56
57
    def _check_tx(self, cn):
        # safety belt to make sure important api points are tx-safe
        if isinstance(cn, Engine) or not cn.in_transaction():
            if not getattr(self, '_testing', False):
                raise TypeError('You must use a transaction object')

58
    def insert(self, cn, newts, seriename, author,
59
60
               metadata=None,
               _insertion_date=None):
61
        """Create a new revision of a given time series
62

63
        newts: pandas.Series with date index
64
        seriename: str unique identifier of the serie
65
        author: str free-form author name
66
        metadata: optional dict for changeset metadata
67
        """
68
        self._check_tx(cn)
69
70
71
72
        assert isinstance(newts, pd.Series), 'Not a pd.Series'
        assert isinstance(seriename, str), 'Name not a string'
        assert isinstance(author, str), 'Author not a string'
        assert metadata is None or isinstance(metadata, dict), 'Bad format for metadata'
73
74
        assert (_insertion_date is None or
                isinstance(_insertion_date, datetime)), 'Bad format for insertion date'
75
        assert not newts.index.duplicated().any(), 'There are some duplicates in the index'
76

77
        assert newts.index.notna().all(), 'The index contains NaT entries'
78
79
        if not newts.index.is_monotonic_increasing:
            newts = newts.sort_index()
80

81
        newts = num2float(newts)
82

83
        if not len(newts):
84
            return
85

86
        assert ('<M8[ns]' == newts.index.dtype or
87
                'datetime' in str(newts.index.dtype) and not
88
89
                isinstance(newts.index, pd.MultiIndex))

90
91
        newts.name = seriename
        table = self._get_ts_table(cn, seriename)
92

93
        if table is None:
94
            return self._create(cn, newts, seriename, author,
95
                                metadata, _insertion_date)
96

97
        return self._update(cn, table, newts, seriename, author,
98
                            metadata, _insertion_date)
99

100
    def get(self, cn, seriename, revision_date=None,
101
102
            from_value_date=None, to_value_date=None,
            _keep_nans=False):
103
104
105
106
107
        """Compute and return the serie of a given name

        revision_date: datetime filter to get previous versions of the
        serie

108
        """
109
        table = self._get_ts_table(cn, seriename)
110
111
        if table is None:
            return
112

113
        csetfilter = []
114
        if revision_date:
115
            csetfilter.append(lambda cset: cset.c.insertion_date <= revision_date)
116
        snap = Snapshot(cn, self, seriename)
117
        _, current = snap.find(csetfilter=csetfilter,
118
119
                               from_value_date=from_value_date,
                               to_value_date=to_value_date)
120

121
        if current is not None and not _keep_nans:
122
            current.name = seriename
123
            current = current[~current.isnull()]
124
        return current
125

126
    def metadata(self, cn, seriename):
127
        """Return metadata dict of timeserie."""
128
129
        if seriename in self.metadatacache:
            return self.metadatacache[seriename]
130
131
        reg = self.schema.registry
        sql = select([reg.c.metadata]).where(
132
            reg.c.seriename == seriename
133
134
        )
        meta = cn.execute(sql).scalar()
135
        self.metadatacache[seriename] = meta
136
137
        return meta

138
    def update_metadata(self, cn, seriename, metadata):
139
        self._check_tx(cn)
140
        assert isinstance(metadata, dict)
141
        assert not set(metadata.keys()) & self.metakeys
142
        meta = self.metadata(cn, seriename)
143
144
145
        # remove al but internal stuff
        newmeta = {key: meta[key] for key in self.metakeys}
        newmeta.update(metadata)
146
147
        reg = self.schema.registry
        sql = reg.update().where(
148
            reg.c.seriename == seriename
149
        ).values(metadata=newmeta)
150
        self.metadatacache.pop(seriename)
151
152
        cn.execute(sql)

153
154
155
156
157
158
159
    def changeset_metadata(self, cn, csid):
        sql = 'select metadata from "{ns}".changeset where id = {id}'.format(
            ns=self.namespace,
            id=csid
        )
        return cn.execute(sql).scalar()

160
    def get_history(self, cn, seriename,
161
                    from_insertion_date=None,
162
163
                    to_insertion_date=None,
                    from_value_date=None,
164
                    to_value_date=None,
165
                    deltabefore=None,
166
167
                    deltaafter=None,
                    diffmode=False):
168
        table = self._get_ts_table(cn, seriename)
169
170
171
        if table is None:
            return

172
        cset = self.schema.changeset
173
174
175
176
177
178
179
180
181
182
183
184
        revsql = select(
            [cset.c.id, cset.c.insertion_date]
        ).order_by(
            cset.c.id
        ).where(
            table.c.cset == cset.c.id
        )

        if from_insertion_date:
            revsql = revsql.where(cset.c.insertion_date >= from_insertion_date)
        if to_insertion_date:
            revsql = revsql.where(cset.c.insertion_date <= to_insertion_date)
185

186
187
188
189
190
191
192
193
194
        if from_value_date or to_value_date:
            revsql = revsql.where(
                closed_overlaps(from_value_date, to_value_date)
            )

        revs = cn.execute(
            revsql,
            {'fromdate': from_value_date, 'todate': to_value_date}
        ).fetchall()
195
        if not revs:
196
            return {}
197

198
199
200
201
202
203
        if diffmode:
            # compute the previous serie value
            first_csid = revs[0][0]
            previous_csid = self._previous_cset(cn, seriename, first_csid)
            revs.insert(0, (previous_csid, None))

204
        snapshot = Snapshot(cn, self, seriename)
205
        series = []
206
207
        if (deltabefore, deltaafter) != (None, None):
            for csid, idate in revs:
208
209
                from_date = None
                to_date = None
210
                if deltabefore is not None:
211
                    from_date = idate - deltabefore
212
                if deltaafter is not None:
213
                    to_date = idate + deltaafter
214
215
216
                series.append((
                    idate,
                    snapshot.find(csetfilter=[lambda cset: cset.c.id == csid],
217
218
                                  from_value_date=from_date,
                                  to_value_date=to_date)[1]
219
220
221
222
223
224
                ))
        else:
            series = snapshot.findall(revs,
                                      from_value_date,
                                      to_value_date
            )
225

226
227
        if diffmode:
            diffs = []
Aurélien Campéas's avatar
Aurélien Campéas committed
228
            for (_revdate_a, serie_a), (revdate_b, serie_b) in zip(series, series[1:]):
229
230
231
232
233
234
235
                if serie_a is None:
                    # when we scan the entirety of the history: there exists no "previous" serie
                    # we therefore consider the first serie as a diff to the "null" serie
                    diffs.append((revdate_b, serie_b))
                else:
                    diffs.append((revdate_b, self.diff(serie_a, serie_b)))
            series = diffs
236
237
238
239
240
        else:
            series = [
                (idate, ts.dropna())
                 for idate, ts in series
            ]
241

242
243
244
245
        return {
            idate: serie
            for idate, serie in series
        }
246

247
248
249
    def get_delta(self, cn, seriename, delta,
                  from_value_date=None,
                  to_value_date=None):
250
251
252
253
        """ compute a series whose value dates are bounded to be
        `delta` time after the insertion dates and where we
        keep the most recent ones
        """
254
        histo = self.get_history(
255
256
257
            cn, seriename, deltabefore=-delta,
            from_value_date=from_value_date,
            to_value_date=to_value_date
258
        )
259
260
        if histo is None:
            return None
261

262
263
264
265
266
267
268
269
        vimap = {}
        vvmap = {}
        for idate, series in histo.items():
            for vdate, value in series.iteritems():
                if vdate not in vimap or vimap[vdate] < idate:
                    vimap[vdate] = idate
                    vvmap[vdate] = value

270
        ts = subset(pd.Series(vvmap).sort_index(), from_value_date, to_value_date)
271
272
        ts.name = seriename
        return ts
273

274
275
    def exists(self, cn, seriename):
        return self._get_ts_table(cn, seriename) is not None
276

277
    def latest_insertion_date(self, cn, seriename):
278
        cset = self.schema.changeset
279
        tstable = self._get_ts_table(cn, seriename)
280
        sql = select([func.max(cset.c.insertion_date)]
281
        ).where(tstable.c.cset == cset.c.id)
282
        return cn.execute(sql).scalar()
283

284
285
286
287
    def last_id(self, cn, seriename):
        snapshot = Snapshot(cn, self, seriename)
        return snapshot.last_id()

288
289
290
    def changeset_at(self, cn, seriename, revdate, mode='strict'):
        assert mode in ('strict', 'before', 'after')
        cset = self.schema.changeset
291
        table = self._table_definition_for(cn, seriename)
292
293
        sql = select([table.c.cset]).where(
            table.c.cset == cset.c.id
294
295
296
297
298
299
300
301
302
        )
        if mode == 'strict':
            sql = sql.where(cset.c.insertion_date == revdate)
        elif mode == 'before':
            sql = sql.where(cset.c.insertion_date <= revdate)
        else:
            sql = sql.where(cset.c.insertion_date >= revdate)
        return cn.execute(sql).scalar()

303
    def delete(self, cn, seriename):
304
        self._check_tx(cn)
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
        assert not isinstance(cn, Engine), 'use a transaction object'
        assert self.exists(cn, seriename)
        # changeset will keep ghost entries
        # we cleanup changeset series, then registry
        # then we drop the two remaining tables
        # cn *must* be a transaction scope
        rid, tablename = cn.execute(
            'select id, table_name from "{}".registry '
            'where seriename = %(seriename)s'.format(self.namespace),
            seriename=seriename
        ).fetchone()
        # drop series tables
        cn.execute(
            'drop table "{}.timeserie"."{}" cascade'.format(self.namespace, tablename)
        )
        cn.execute(
            'drop table "{}.snapshot"."{}" cascade'.format(self.namespace, tablename)
        )
        # cleanup changesets table
        cn.execute('with csets as ('
                   ' select cset from "{ns}".changeset_series '
                   ' where serie = %(rid)s'
                   ') '
                   'delete from "{ns}".changeset as cset using csets '
                   'where cset.id = csets.cset'.format(ns=self.namespace),
                   rid=rid
        )
        cn.execute('delete from "{}".registry '
                   'where id = %(rid)s'.format(self.namespace),
                   rid=rid)
        # -> this will transitively cleanup state changeset_series entries
        self._resetcaches()

338
    def strip(self, cn, seriename, csid):
339
        self._check_tx(cn)
340
341
342
343
344
345
346
347
        logs = self.log(cn, fromrev=csid, names=(seriename,))
        assert logs

        # put stripping info in the metadata
        cset = self.schema.changeset
        cset_serie = self.schema.changeset_series
        for log in logs:
            # update changeset.metadata
348
            metadata = self.changeset_metadata(cn, log['rev']) or {}
349
350
351
352
353
354
            metadata['tshistory.info'] = 'got stripped from {}'.format(csid)
            sql = cset.update().where(cset.c.id == log['rev']
            ).values(metadata=metadata)
            cn.execute(sql)
            # delete changset_serie item
            sql = cset_serie.delete().where(
355
                cset_serie.c.cset == log['rev']
356
            ).where(
357
                cset_serie.c.serie == self._name_to_regid(cn, seriename)
358
359
360
361
            )
            cn.execute(sql)

        # wipe the diffs
362
        table = self._table_definition_for(cn, seriename)
363
        cn.execute(table.delete().where(table.c.cset >= csid))
364

365
    def info(self, cn):
366
367
        """Gather global statistics on the current tshistory repository
        """
368
        sql = 'select count(*) from "{}".registry'.format(self.namespace)
369
        stats = {'series count': cn.execute(sql).scalar()}
370
        sql = 'select max(id) from "{}".changeset'.format(self.namespace)
371
        stats['changeset count'] = cn.execute(sql).scalar()
372
373
374
        sql = 'select distinct seriename from "{}".registry order by seriename'.format(
            self.namespace
        )
375
        stats['serie names'] = [row for row, in cn.execute(sql).fetchall()]
376
377
        return stats

378
    def log(self, cn, limit=0, names=None, authors=None,
379
            stripped=False,
380
381
            fromrev=None, torev=None,
            fromdate=None, todate=None):
382
383
384
385
        """Build a structure showing the history of all the series in the db,
        per changeset, in chronological order.
        """
        log = []
386
387
388
389
390
        cset, cset_series, reg = (
            self.schema.changeset,
            self.schema.changeset_series,
            self.schema.registry
        )
391

392
        sql = select([cset.c.id, cset.c.author, cset.c.insertion_date, cset.c.metadata]
393
        ).distinct().order_by(desc(cset.c.id))
394
395
396

        if limit:
            sql = sql.limit(limit)
397
        if names:
398
            sql = sql.where(reg.c.seriename.in_(names))
399
400
        if authors:
            sql = sql.where(cset.c.author.in_(authors))
401
402
403
404
        if fromrev:
            sql = sql.where(cset.c.id >= fromrev)
        if torev:
            sql = sql.where(cset.c.id <= torev)
405
406
407
408
        if fromdate:
            sql = sql.where(cset.c.insertion_date >= fromdate)
        if todate:
            sql = sql.where(cset.c.insertion_date <= todate)
409
410
411
412
        if stripped:
            # outerjoin to show dead things
            sql = sql.select_from(cset.outerjoin(cset_series))
        else:
413
            sql = sql.where(cset.c.id == cset_series.c.cset
414
            ).where(cset_series.c.serie == reg.c.id)
415

416
        rset = cn.execute(sql)
417
        for csetid, author, revdate, meta in rset.fetchall():
418
419
            log.append({'rev': csetid, 'author': author,
                        'date': pd.Timestamp(revdate, tz='utc'),
420
                        'meta': meta or {},
421
                        'names': self._changeset_series(cn, csetid)})
422

423
        log.sort(key=lambda rev: rev['rev'])
424
425
        return log

426
    def interval(self, cn, seriename, notz=False):
427
428
429
430
431
432
433
434
        tablename = self._serie_to_tablename(cn, seriename)
        if tablename is None:
            raise ValueError(f'no such serie: {seriename}')
        sql = (f'select start, "end" '
               f'from "{self.namespace}.timeserie"."{tablename}" '
               f'order by cset desc limit 1')
        res = cn.execute(sql).fetchone()
        start, end = res.start, res.end
435
        if self.metadata(cn, seriename).get('tzaware') and not notz:
436
437
438
            start, end = pd.Timestamp(start, tz='UTC'), pd.Timestamp(end, tz='UTC')
        return pd.Interval(left=start, right=end, closed='both')

439
440
    # /API
    # Helpers
441

442
443
    # creation / update

444
    def _create(self, cn, newts, seriename, author,
445
                metadata=None, insertion_date=None):
446
        start, end = start_end(newts, notz=False)
447
448
449
450
        if start is None:
            assert end is None
            # this is just full of nans
            return None
451
452
453
454
455
        # chop off unwanted nans
        newts = newts.loc[start:end]
        if len(newts) == 0:
            return None

456
457
458
459
460
        # at creation time we take an exclusive lock to avoid
        # a deadlock on created tables against the changeset-series fk
        cn.execute(
            'select pg_advisory_xact_lock({})'.format(self.create_lock_id)
        )
461
        self._register_serie(cn, seriename, newts)
462
        snapshot = Snapshot(cn, self, seriename)
463
        csid = self._newchangeset(cn, author, insertion_date, metadata)
464
        head = snapshot.create(newts)
465
        start, end = start_end(newts)
466
467
        value = {
            'cset': csid,
468
469
470
            'snapshot': head,
            'start': start,
            'end': end
471
        }
Aurélien Campéas's avatar
Aurélien Campéas committed
472
        table = self._make_ts_table(cn, seriename)
473
        cn.execute(table.insert().values(value))
474
        self._finalize_insertion(cn, csid, seriename)
475
        L.info('first insertion of %s (size=%s) by %s',
476
               seriename, len(newts), author)
477
478
        return newts

479
    def _update(self, cn, table, newts, seriename, author,
480
                metadata=None, insertion_date=None):
481
482
        self._validate(cn, newts, seriename)
        snapshot = Snapshot(cn, self, seriename)
483
484
485
        diff = self.diff(snapshot.last(newts.index.min(),
                                       newts.index.max()),
                         newts)
486
487
        if not len(diff):
            L.info('no difference in %s by %s (for ts of size %s)',
488
                   seriename, author, len(newts))
489
490
            return

491
        # compute series start/end stamps
492
        tsstart, tsend = start_end(newts)
493
494
495
        ival = self.interval(cn, seriename, notz=True)
        start = min(tsstart or ival.left, ival.left)
        end = max(tsend or ival.right, ival.right)
496
497
498
499
500
501
502
503
504
505
506
507
508
509

        if pd.isnull(diff[0]) or pd.isnull(diff[-1]):
            # we *might* be shrinking, let's look at the full series
            # and yes, shrinkers have a slow path
            last = snapshot.last()
            patched = self.patch(last, diff).dropna()
            if not len(patched):
                raise ValueError('complete erasure of a series is forbidden')
            if pd.isnull(diff[0]):
                start = patched.index[0]
            if pd.isnull(diff[-1]):
                end = patched.index[-1]

        csid = self._newchangeset(cn, author, insertion_date, metadata)
510
511
512
        head = snapshot.update(diff)
        value = {
            'cset': csid,
513
514
515
            'snapshot': head,
            'start': start,
            'end': end
516
517
        }
        cn.execute(table.insert().values(value))
518
        self._finalize_insertion(cn, csid, seriename)
519
520

        L.info('inserted diff (size=%s) for ts %s by %s',
521
               len(diff), seriename, author)
522
523
        return diff

524
    # serie table handling
525

526
527
528
529
    def _serie_to_tablename(self, cn, seriename):
        tablename = self.serie_tablename.get(seriename)
        if tablename is not None:
            return tablename
530

531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
        reg = self.schema.registry
        sql = select([reg.c.table_name]).where(reg.c.seriename == seriename)
        tablename = cn.execute(sql).scalar()
        if tablename is None:
            # creation time
            return
        self.serie_tablename[seriename] = tablename
        return tablename

    def _table_definition_for(self, cn, seriename):
        tablename = self._serie_to_tablename(cn, seriename)
        if tablename is None:
            # creation time
            tablename = self._make_tablename(seriename)
        fq_tablename = '{}.timeserie.{}'.format(self.namespace, tablename)
        table = TABLES.get(fq_tablename)
547
        if table is None:
548
549
            TABLES[fq_tablename] = table = Table(
                tablename, self.schema.meta,
550
551
552
                Column('id', Integer, primary_key=True),
                Column('cset', Integer,
                       ForeignKey('{}.changeset.id'.format(self.namespace)),
553
                       nullable=False),
554
555
                Column('start', TIMESTAMP, nullable=False),
                Column('end', TIMESTAMP, nullable=False),
556
557
558
                Column('snapshot', Integer,
                       ForeignKey('{}.snapshot.{}.id'.format(
                           self.namespace,
559
560
561
                           tablename))),
                Index(NONE_NAME, 'cset'),
                Index(NONE_NAME, 'snapshot'),
562
563
                Index(NONE_NAME, 'start'),
                Index(NONE_NAME, 'end'),
564
                schema='{}.timeserie'.format(self.namespace),
565
                keep_existing=True
566
567
            )
        return table
568

Aurélien Campéas's avatar
Aurélien Campéas committed
569
    def _make_ts_table(self, cn, seriename):
570
        table = self._table_definition_for(cn, seriename)
571
        table.create(cn)
572
573
574
        return table

    def _register_serie(self, cn, seriename, ts):
575
576
        index = ts.index
        inames = [name for name in index.names if name]
577
        sql = self.schema.registry.insert().values(
578
579
            seriename=seriename,
            table_name=self._make_tablename(seriename),
580
581
582
583
            metadata={
                'tzaware': tzaware_serie(ts),
                'index_type': index.dtype.name,
                'index_names': inames,
584
585
                'index_dtype': index.dtype.str,
                'value_dtype': ts.dtypes.str,
586
                'value_type': ts.dtypes.name
587
            }
588
        )
589
590
        regid = cn.execute(sql).inserted_primary_key[0]
        self.registry_map[seriename] = regid
591

592
    def _get_ts_table(self, cn, seriename):
593
594
595
        tablename = self._serie_to_tablename(cn, seriename)
        if tablename:
            return self._table_definition_for(cn, tablename)
596

597
598
    # changeset handling

599
    def _newchangeset(self, cn, author, insertion_date=None, metadata=None):
600
        table = self.schema.changeset
601
602
603
        if insertion_date is not None:
            assert insertion_date.tzinfo is not None
        idate = pd.Timestamp(insertion_date or datetime.utcnow(), tz='UTC')
604
605
        sql = table.insert().values(
            author=author,
606
            metadata=metadata,
607
            insertion_date=idate)
608
        return cn.execute(sql).inserted_primary_key[0]
609

610
    def _changeset_series(self, cn, csid):
611
        cset_serie = self.schema.changeset_series
612
613
        reg = self.schema.registry
        sql = select(
614
            [reg.c.seriename]
615
616
        ).where(cset_serie.c.cset == csid
        ).where(cset_serie.c.serie == reg.c.id)
617

618
        return [
619
            row.seriename
620
621
            for row in cn.execute(sql).fetchall()
        ]
622

623
624
625
626
627
628
629
    def _previous_cset(self, cn, seriename, csid):
        tablename = self._serie_to_tablename(cn, seriename)
        sql = ('select cset from "{}.timeserie"."{}" '
               'where cset < %(csid)s '
               'order by cset desc limit 1').format(self.namespace, tablename)
        return cn.execute(sql, csid=csid).scalar()

630
631
    # insertion handling

632
    def _validate(self, cn, ts, seriename):
633
634
        if ts.isnull().all():
            # ts erasure
635
            return
636
        tstype = ts.dtype
637
        meta = self.metadata(cn, seriename)
638
        if tstype != meta['value_type']:
639
            m = 'Type error when inserting {}, new type is {}, type in base is {}'.format(
640
                seriename, tstype, meta['value_type'])
641
            raise Exception(m)
642
        if ts.index.dtype.name != meta['index_type']:
643
            raise Exception('Incompatible index types')
644

645
646
    def _name_to_regid(self, cn, seriename):
        regid = self.registry_map.get(seriename)
647
648
649
650
        if regid is not None:
            return regid

        registry = self.schema.registry
651
        sql = select([registry.c.id]).where(registry.c.seriename == seriename)
652
        regid = self.registry_map[seriename] = cn.execute(sql).scalar()
653
654
        return regid

655
    def _finalize_insertion(self, cn, csid, seriename):
656
657
658
        table = self.schema.changeset_series
        sql = table.insert().values(
            cset=csid,
659
            serie=self._name_to_regid(cn, seriename)
660
661
        )
        cn.execute(sql)
662

663
664
665
    def _resetcaches(self):
        TABLES.clear()
        SNAPTABLES.clear()
666
667
668
        self.metadatacache.clear()
        self.registry_map.clear()
        self.serie_tablename.clear()