tsio.py 22.9 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
    registry_map = None
32
    serie_tablename = None
33
    create_lock_id = None
34
35
36

    def __init__(self, namespace='tsh'):
        self.namespace = namespace
37
38
        self.schema = tsschema(namespace)
        self.schema.define()
39
        self.metadatacache = {}
40
        self.registry_map = {}
41
        self.serie_tablename = {}
42
        self.create_lock_id = sum(ord(c) for c in namespace)
43

44
    def insert(self, cn, newts, seriename, author,
45
46
               metadata=None,
               _insertion_date=None):
47
        """Create a new revision of a given time series
48

49
        newts: pandas.Series with date index
50
        seriename: str unique identifier of the serie
51
        author: str free-form author name
52
        metadata: optional dict for changeset metadata
53
        """
54
55
56
57
58
59
        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'
        assert _insertion_date is None or isinstance(_insertion_date, datetime), 'Bad format for insertion date'
        assert not newts.index.duplicated().any(), 'There are some duplicates in the index'
60
61
62

        if not newts.index.is_monotonic_increasing:
            newts = newts.sort_index()
63

64
        newts = num2float(newts)
65

66
        if not len(newts):
67
            return
68

69
        assert ('<M8[ns]' == newts.index.dtype or
70
                'datetime' in str(newts.index.dtype) and not
71
72
                isinstance(newts.index, pd.MultiIndex))

73
74
        newts.name = seriename
        table = self._get_ts_table(cn, seriename)
75

76
        if table is None:
77
            return self._create(cn, newts, seriename, author,
78
                                metadata, _insertion_date)
79

80
        return self._update(cn, table, newts, seriename, author,
81
                            metadata, _insertion_date)
82

83
    def get(self, cn, seriename, revision_date=None,
84
85
            from_value_date=None, to_value_date=None,
            _keep_nans=False):
86
87
88
89
90
        """Compute and return the serie of a given name

        revision_date: datetime filter to get previous versions of the
        serie

91
        """
92
        table = self._get_ts_table(cn, seriename)
93
94
        if table is None:
            return
95

96
        csetfilter = []
97
        if revision_date:
98
            csetfilter.append(lambda cset: cset.c.insertion_date <= revision_date)
99
        snap = Snapshot(cn, self, seriename)
100
        _, current = snap.find(csetfilter=csetfilter,
101
102
                               from_value_date=from_value_date,
                               to_value_date=to_value_date)
103

104
        if current is not None and not _keep_nans:
105
            current.name = seriename
106
            current = current[~current.isnull()]
107
        return current
108

109
    def metadata(self, cn, seriename):
110
        """Return metadata dict of timeserie."""
111
112
        if seriename in self.metadatacache:
            return self.metadatacache[seriename]
113
114
        reg = self.schema.registry
        sql = select([reg.c.metadata]).where(
115
            reg.c.seriename == seriename
116
117
        )
        meta = cn.execute(sql).scalar()
118
        self.metadatacache[seriename] = meta
119
120
        return meta

121
    def update_metadata(self, cn, seriename, metadata, internal=False):
122
        assert isinstance(metadata, dict)
123
        meta = self.metadata(cn, seriename)
124
125
126
127
128
        if not internal:
            assert set(meta.keys()).intersection(metadata.keys()) == set()
        meta.update(metadata)
        reg = self.schema.registry
        sql = reg.update().where(
129
            reg.c.seriename == seriename
130
131
        ).values(metadata=meta)
        self.metadatacache.pop(seriename)
132
133
        cn.execute(sql)

134
135
136
137
138
139
140
    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()

141
    def get_history(self, cn, seriename,
142
                    from_insertion_date=None,
143
144
                    to_insertion_date=None,
                    from_value_date=None,
145
                    to_value_date=None,
146
                    deltabefore=None,
147
148
                    deltaafter=None,
                    diffmode=False):
149
        table = self._get_ts_table(cn, seriename)
150
151
152
        if table is None:
            return

153
        cset = self.schema.changeset
154
155
156
157
158
159
160
161
162
163
164
165
        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)
166

167
168
169
170
171
172
173
174
175
        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()
176
        if not revs:
177
            return {}
178

179
180
181
182
183
184
        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))

185
        snapshot = Snapshot(cn, self, seriename)
186
        series = []
187
188
        if (deltabefore, deltaafter) != (None, None):
            for csid, idate in revs:
189
190
                from_date = None
                to_date = None
191
                if deltabefore is not None:
192
                    from_date = idate - deltabefore
193
                if deltaafter is not None:
194
                    to_date = idate + deltaafter
195
196
197
                series.append((
                    idate,
                    snapshot.find(csetfilter=[lambda cset: cset.c.id == csid],
198
199
                                  from_value_date=from_date,
                                  to_value_date=to_date)[1]
200
201
202
203
204
205
                ))
        else:
            series = snapshot.findall(revs,
                                      from_value_date,
                                      to_value_date
            )
206

207
208
        if diffmode:
            diffs = []
Aurélien Campéas's avatar
Aurélien Campéas committed
209
            for (_revdate_a, serie_a), (revdate_b, serie_b) in zip(series, series[1:]):
210
211
212
213
214
215
216
217
                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

218
219
220
221
        return {
            idate: serie
            for idate, serie in series
        }
222

223
224
225
    def get_delta(self, cn, seriename, delta,
                  from_value_date=None,
                  to_value_date=None):
226
227
228
229
        """ compute a series whose value dates are bounded to be
        `delta` time after the insertion dates and where we
        keep the most recent ones
        """
230

231
        histo = self.get_history(
232
233
234
            cn, seriename, deltabefore=-delta,
            from_value_date=from_value_date,
            to_value_date=to_value_date
235
        )
236
237
        if histo is None:
            return None
238

239
240
241
242
243
244
245
246
        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

247
        ts = subset(pd.Series(vvmap).sort_index(), from_value_date, to_value_date)
248
249
        ts.name = seriename
        return ts
250

251
252
    def exists(self, cn, seriename):
        return self._get_ts_table(cn, seriename) is not None
253

254
    def latest_insertion_date(self, cn, seriename):
255
        cset = self.schema.changeset
256
        tstable = self._get_ts_table(cn, seriename)
257
        sql = select([func.max(cset.c.insertion_date)]
258
        ).where(tstable.c.cset == cset.c.id)
259
        return cn.execute(sql).scalar()
260

261
262
263
264
    def last_id(self, cn, seriename):
        snapshot = Snapshot(cn, self, seriename)
        return snapshot.last_id()

265
266
267
    def changeset_at(self, cn, seriename, revdate, mode='strict'):
        assert mode in ('strict', 'before', 'after')
        cset = self.schema.changeset
268
        table = self._table_definition_for(cn, seriename)
269
270
        sql = select([table.c.cset]).where(
            table.c.cset == cset.c.id
271
272
273
274
275
276
277
278
279
        )
        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()

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
    def delete(self, cn, seriename):
        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()

314
315
316
317
318
319
320
321
322
    def strip(self, cn, seriename, csid):
        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
323
            metadata = self.changeset_metadata(cn, log['rev']) or {}
324
325
326
327
328
329
            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(
330
                cset_serie.c.cset == log['rev']
331
            ).where(
332
                cset_serie.c.serie == self._name_to_regid(cn, seriename)
333
334
335
336
            )
            cn.execute(sql)

        # wipe the diffs
337
        table = self._table_definition_for(cn, seriename)
338
        cn.execute(table.delete().where(table.c.cset >= csid))
339

340
    def info(self, cn):
341
342
        """Gather global statistics on the current tshistory repository
        """
343
        sql = 'select count(*) from "{}".registry'.format(self.namespace)
344
        stats = {'series count': cn.execute(sql).scalar()}
345
        sql = 'select max(id) from "{}".changeset'.format(self.namespace)
346
        stats['changeset count'] = cn.execute(sql).scalar()
347
348
349
        sql = 'select distinct seriename from "{}".registry order by seriename'.format(
            self.namespace
        )
350
        stats['serie names'] = [row for row, in cn.execute(sql).fetchall()]
351
352
        return stats

353
    def log(self, cn, limit=0, names=None, authors=None,
354
            stripped=False,
355
356
            fromrev=None, torev=None,
            fromdate=None, todate=None):
357
358
359
360
        """Build a structure showing the history of all the series in the db,
        per changeset, in chronological order.
        """
        log = []
361
362
363
364
365
        cset, cset_series, reg = (
            self.schema.changeset,
            self.schema.changeset_series,
            self.schema.registry
        )
366

367
        sql = select([cset.c.id, cset.c.author, cset.c.insertion_date, cset.c.metadata]
368
        ).distinct().order_by(desc(cset.c.id))
369
370
371

        if limit:
            sql = sql.limit(limit)
372
        if names:
373
            sql = sql.where(reg.c.seriename.in_(names))
374
375
        if authors:
            sql = sql.where(cset.c.author.in_(authors))
376
377
378
379
        if fromrev:
            sql = sql.where(cset.c.id >= fromrev)
        if torev:
            sql = sql.where(cset.c.id <= torev)
380
381
382
383
        if fromdate:
            sql = sql.where(cset.c.insertion_date >= fromdate)
        if todate:
            sql = sql.where(cset.c.insertion_date <= todate)
384
385
386
387
        if stripped:
            # outerjoin to show dead things
            sql = sql.select_from(cset.outerjoin(cset_series))
        else:
388
            sql = sql.where(cset.c.id == cset_series.c.cset
389
            ).where(cset_series.c.serie == reg.c.id)
390

391
        rset = cn.execute(sql)
392
        for csetid, author, revdate, meta in rset.fetchall():
393
394
            log.append({'rev': csetid, 'author': author,
                        'date': pd.Timestamp(revdate, tz='utc'),
395
                        'meta': meta or {},
396
                        'names': self._changeset_series(cn, csetid)})
397

398
        log.sort(key=lambda rev: rev['rev'])
399
400
        return log

401
402
403
404
405
406
407
408
409
410
411
412
413
    def interval(self, cn, seriename):
        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
        if self.metadata(cn, seriename).get('tzaware'):
            start, end = pd.Timestamp(start, tz='UTC'), pd.Timestamp(end, tz='UTC')
        return pd.Interval(left=start, right=end, closed='both')

414
415
    # /API
    # Helpers
416

417
418
    # creation / update

419
    def _create(self, cn, newts, seriename, author,
420
                metadata=None, insertion_date=None):
421
422
423
        # initial insertion
        if len(newts) == 0:
            return None
424
425
426
427
428
        # 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)
        )
429
        self._register_serie(cn, seriename, newts)
430
        snapshot = Snapshot(cn, self, seriename)
431
        csid = self._newchangeset(cn, author, insertion_date, metadata)
432
        head = snapshot.create(newts)
433
        start, end = start_end(newts)
434
435
        value = {
            'cset': csid,
436
437
438
            'snapshot': head,
            'start': start,
            'end': end
439
        }
Aurélien Campéas's avatar
Aurélien Campéas committed
440
        table = self._make_ts_table(cn, seriename)
441
        cn.execute(table.insert().values(value))
442
        self._finalize_insertion(cn, csid, seriename)
443
        L.info('first insertion of %s (size=%s) by %s',
444
               seriename, len(newts), author)
445
446
        return newts

447
    def _update(self, cn, table, newts, seriename, author,
448
                metadata=None, insertion_date=None):
449
450
        self._validate(cn, newts, seriename)
        snapshot = Snapshot(cn, self, seriename)
451
452
453
        diff = self.diff(snapshot.last(newts.index.min(),
                                       newts.index.max()),
                         newts)
454
455
        if not len(diff):
            L.info('no difference in %s by %s (for ts of size %s)',
456
                   seriename, author, len(newts))
457
458
            return

459
        csid = self._newchangeset(cn, author, insertion_date, metadata)
460
461
462
463
        tsstart, tsend = start_end(newts)
        ival = self.interval(cn, seriename)
        start = min(tsstart, ival.left.replace(tzinfo=None))
        end = max(tsend, ival.right.replace(tzinfo=None))
464
465
466
        head = snapshot.update(diff)
        value = {
            'cset': csid,
467
468
469
            'snapshot': head,
            'start': start,
            'end': end
470
471
        }
        cn.execute(table.insert().values(value))
472
        self._finalize_insertion(cn, csid, seriename)
473
474

        L.info('inserted diff (size=%s) for ts %s by %s',
475
               len(diff), seriename, author)
476
477
        return diff

478
    # serie table handling
479

480
481
482
483
    def _serie_to_tablename(self, cn, seriename):
        tablename = self.serie_tablename.get(seriename)
        if tablename is not None:
            return tablename
484

485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
        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)
501
        if table is None:
502
503
            TABLES[fq_tablename] = table = Table(
                tablename, self.schema.meta,
504
505
506
                Column('id', Integer, primary_key=True),
                Column('cset', Integer,
                       ForeignKey('{}.changeset.id'.format(self.namespace)),
507
                       nullable=False),
508
509
                Column('start', TIMESTAMP, nullable=False),
                Column('end', TIMESTAMP, nullable=False),
510
511
512
                Column('snapshot', Integer,
                       ForeignKey('{}.snapshot.{}.id'.format(
                           self.namespace,
513
514
515
                           tablename))),
                Index(NONE_NAME, 'cset'),
                Index(NONE_NAME, 'snapshot'),
516
517
                Index(NONE_NAME, 'start'),
                Index(NONE_NAME, 'end'),
518
                schema='{}.timeserie'.format(self.namespace),
519
                keep_existing=True
520
521
            )
        return table
522

Aurélien Campéas's avatar
Aurélien Campéas committed
523
    def _make_ts_table(self, cn, seriename):
524
        table = self._table_definition_for(cn, seriename)
525
        table.create(cn)
526
527
528
        return table

    def _register_serie(self, cn, seriename, ts):
529
530
        index = ts.index
        inames = [name for name in index.names if name]
531
        sql = self.schema.registry.insert().values(
532
533
            seriename=seriename,
            table_name=self._make_tablename(seriename),
534
535
536
537
            metadata={
                'tzaware': tzaware_serie(ts),
                'index_type': index.dtype.name,
                'index_names': inames,
538
539
                'index_dtype': index.dtype.str,
                'value_dtype': ts.dtypes.str,
540
                'value_type': ts.dtypes.name
541
            }
542
        )
543
544
        regid = cn.execute(sql).inserted_primary_key[0]
        self.registry_map[seriename] = regid
545

546
    def _get_ts_table(self, cn, seriename):
547
548
549
        tablename = self._serie_to_tablename(cn, seriename)
        if tablename:
            return self._table_definition_for(cn, tablename)
550

551
552
    # changeset handling

553
    def _newchangeset(self, cn, author, insertion_date=None, metadata=None):
554
        table = self.schema.changeset
555
556
557
        if insertion_date is not None:
            assert insertion_date.tzinfo is not None
        idate = pd.Timestamp(insertion_date or datetime.utcnow(), tz='UTC')
558
559
        sql = table.insert().values(
            author=author,
560
            metadata=metadata,
561
            insertion_date=idate)
562
        return cn.execute(sql).inserted_primary_key[0]
563

564
    def _changeset_series(self, cn, csid):
565
        cset_serie = self.schema.changeset_series
566
567
        reg = self.schema.registry
        sql = select(
568
            [reg.c.seriename]
569
570
        ).where(cset_serie.c.cset == csid
        ).where(cset_serie.c.serie == reg.c.id)
571

572
        return [
573
            row.seriename
574
575
            for row in cn.execute(sql).fetchall()
        ]
576

577
578
579
580
581
582
583
    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()

584
585
    # insertion handling

586
    def _validate(self, cn, ts, seriename):
587
588
        if ts.isnull().all():
            # ts erasure
589
            return
590
        tstype = ts.dtype
591
        meta = self.metadata(cn, seriename)
592
        if tstype != meta['value_type']:
593
            m = 'Type error when inserting {}, new type is {}, type in base is {}'.format(
594
                seriename, tstype, meta['value_type'])
595
            raise Exception(m)
596
        if ts.index.dtype.name != meta['index_type']:
597
            raise Exception('Incompatible index types')
598

599
600
    def _name_to_regid(self, cn, seriename):
        regid = self.registry_map.get(seriename)
601
602
603
604
        if regid is not None:
            return regid

        registry = self.schema.registry
605
        sql = select([registry.c.id]).where(registry.c.seriename == seriename)
606
        regid = self.registry_map[seriename] = cn.execute(sql).scalar()
607
608
        return regid

609
    def _finalize_insertion(self, cn, csid, seriename):
610
611
612
        table = self.schema.changeset_series
        sql = table.insert().values(
            cset=csid,
613
            serie=self._name_to_regid(cn, seriename)
614
615
        )
        cn.execute(sql)
616

617
618
619
    def _resetcaches(self):
        TABLES.clear()
        SNAPTABLES.clear()
620
621
622
        self.metadatacache.clear()
        self.registry_map.clear()
        self.serie_tablename.clear()