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
        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'
58
59
        assert (_insertion_date is None or
                isinstance(_insertion_date, datetime)), 'Bad format for insertion date'
60
        assert not newts.index.duplicated().any(), 'There are some duplicates in the index'
61
62
63

        if not newts.index.is_monotonic_increasing:
            newts = newts.sort_index()
64

65
        newts = num2float(newts)
66

67
        if not len(newts):
68
            return
69

70
        assert ('<M8[ns]' == newts.index.dtype or
71
                'datetime' in str(newts.index.dtype) and not
72
73
                isinstance(newts.index, pd.MultiIndex))

74
75
        newts.name = seriename
        table = self._get_ts_table(cn, seriename)
76

77
        if table is None:
78
            return self._create(cn, newts, seriename, author,
79
                                metadata, _insertion_date)
80

81
        return self._update(cn, table, newts, seriename, author,
82
                            metadata, _insertion_date)
83

84
    def get(self, cn, seriename, revision_date=None,
85
86
            from_value_date=None, to_value_date=None,
            _keep_nans=False):
87
88
89
90
91
        """Compute and return the serie of a given name

        revision_date: datetime filter to get previous versions of the
        serie

92
        """
93
        table = self._get_ts_table(cn, seriename)
94
95
        if table is None:
            return
96

97
        csetfilter = []
98
        if revision_date:
99
            csetfilter.append(lambda cset: cset.c.insertion_date <= revision_date)
100
        snap = Snapshot(cn, self, seriename)
101
        _, current = snap.find(csetfilter=csetfilter,
102
103
                               from_value_date=from_value_date,
                               to_value_date=to_value_date)
104

105
        if current is not None and not _keep_nans:
106
            current.name = seriename
107
            current = current[~current.isnull()]
108
        return current
109

110
    def metadata(self, cn, seriename):
111
        """Return metadata dict of timeserie."""
112
113
        if seriename in self.metadatacache:
            return self.metadatacache[seriename]
114
115
        reg = self.schema.registry
        sql = select([reg.c.metadata]).where(
116
            reg.c.seriename == seriename
117
118
        )
        meta = cn.execute(sql).scalar()
119
        self.metadatacache[seriename] = meta
120
121
        return meta

122
    def update_metadata(self, cn, seriename, metadata, internal=False):
123
        assert isinstance(metadata, dict)
124
        meta = self.metadata(cn, seriename)
125
126
127
128
129
        if not internal:
            assert set(meta.keys()).intersection(metadata.keys()) == set()
        meta.update(metadata)
        reg = self.schema.registry
        sql = reg.update().where(
130
            reg.c.seriename == seriename
131
132
        ).values(metadata=meta)
        self.metadatacache.pop(seriename)
133
134
        cn.execute(sql)

135
136
137
138
139
140
141
    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()

142
    def get_history(self, cn, seriename,
143
                    from_insertion_date=None,
144
145
                    to_insertion_date=None,
                    from_value_date=None,
146
                    to_value_date=None,
147
                    deltabefore=None,
148
149
                    deltaafter=None,
                    diffmode=False):
150
        table = self._get_ts_table(cn, seriename)
151
152
153
        if table is None:
            return

154
        cset = self.schema.changeset
155
156
157
158
159
160
161
162
163
164
165
166
        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)
167

168
169
170
171
172
173
174
175
176
        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()
177
        if not revs:
178
            return {}
179

180
181
182
183
184
185
        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))

186
        snapshot = Snapshot(cn, self, seriename)
187
        series = []
188
189
        if (deltabefore, deltaafter) != (None, None):
            for csid, idate in revs:
190
191
                from_date = None
                to_date = None
192
                if deltabefore is not None:
193
                    from_date = idate - deltabefore
194
                if deltaafter is not None:
195
                    to_date = idate + deltaafter
196
197
198
                series.append((
                    idate,
                    snapshot.find(csetfilter=[lambda cset: cset.c.id == csid],
199
200
                                  from_value_date=from_date,
                                  to_value_date=to_date)[1]
201
202
203
204
205
206
                ))
        else:
            series = snapshot.findall(revs,
                                      from_value_date,
                                      to_value_date
            )
207

208
209
        if diffmode:
            diffs = []
Aurélien Campéas's avatar
Aurélien Campéas committed
210
            for (_revdate_a, serie_a), (revdate_b, serie_b) in zip(series, series[1:]):
211
212
213
214
215
216
217
218
                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

219
220
221
222
        return {
            idate: serie
            for idate, serie in series
        }
223

224
225
226
    def get_delta(self, cn, seriename, delta,
                  from_value_date=None,
                  to_value_date=None):
227
228
229
230
        """ compute a series whose value dates are bounded to be
        `delta` time after the insertion dates and where we
        keep the most recent ones
        """
231

232
        histo = self.get_history(
233
234
235
            cn, seriename, deltabefore=-delta,
            from_value_date=from_value_date,
            to_value_date=to_value_date
236
        )
237
238
        if histo is None:
            return None
239

240
241
242
243
244
245
246
247
        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

248
        ts = subset(pd.Series(vvmap).sort_index(), from_value_date, to_value_date)
249
250
        ts.name = seriename
        return ts
251

252
253
    def exists(self, cn, seriename):
        return self._get_ts_table(cn, seriename) is not None
254

255
    def latest_insertion_date(self, cn, seriename):
256
        cset = self.schema.changeset
257
        tstable = self._get_ts_table(cn, seriename)
258
        sql = select([func.max(cset.c.insertion_date)]
259
        ).where(tstable.c.cset == cset.c.id)
260
        return cn.execute(sql).scalar()
261

262
263
264
265
    def last_id(self, cn, seriename):
        snapshot = Snapshot(cn, self, seriename)
        return snapshot.last_id()

266
267
268
    def changeset_at(self, cn, seriename, revdate, mode='strict'):
        assert mode in ('strict', 'before', 'after')
        cset = self.schema.changeset
269
        table = self._table_definition_for(cn, seriename)
270
271
        sql = select([table.c.cset]).where(
            table.c.cset == cset.c.id
272
273
274
275
276
277
278
279
280
        )
        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()

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
    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()

315
316
317
318
319
320
321
322
323
    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
324
            metadata = self.changeset_metadata(cn, log['rev']) or {}
325
326
327
328
329
330
            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(
331
                cset_serie.c.cset == log['rev']
332
            ).where(
333
                cset_serie.c.serie == self._name_to_regid(cn, seriename)
334
335
336
337
            )
            cn.execute(sql)

        # wipe the diffs
338
        table = self._table_definition_for(cn, seriename)
339
        cn.execute(table.delete().where(table.c.cset >= csid))
340

341
    def info(self, cn):
342
343
        """Gather global statistics on the current tshistory repository
        """
344
        sql = 'select count(*) from "{}".registry'.format(self.namespace)
345
        stats = {'series count': cn.execute(sql).scalar()}
346
        sql = 'select max(id) from "{}".changeset'.format(self.namespace)
347
        stats['changeset count'] = cn.execute(sql).scalar()
348
349
350
        sql = 'select distinct seriename from "{}".registry order by seriename'.format(
            self.namespace
        )
351
        stats['serie names'] = [row for row, in cn.execute(sql).fetchall()]
352
353
        return stats

354
    def log(self, cn, limit=0, names=None, authors=None,
355
            stripped=False,
356
357
            fromrev=None, torev=None,
            fromdate=None, todate=None):
358
359
360
361
        """Build a structure showing the history of all the series in the db,
        per changeset, in chronological order.
        """
        log = []
362
363
364
365
366
        cset, cset_series, reg = (
            self.schema.changeset,
            self.schema.changeset_series,
            self.schema.registry
        )
367

368
        sql = select([cset.c.id, cset.c.author, cset.c.insertion_date, cset.c.metadata]
369
        ).distinct().order_by(desc(cset.c.id))
370
371
372

        if limit:
            sql = sql.limit(limit)
373
        if names:
374
            sql = sql.where(reg.c.seriename.in_(names))
375
376
        if authors:
            sql = sql.where(cset.c.author.in_(authors))
377
378
379
380
        if fromrev:
            sql = sql.where(cset.c.id >= fromrev)
        if torev:
            sql = sql.where(cset.c.id <= torev)
381
382
383
384
        if fromdate:
            sql = sql.where(cset.c.insertion_date >= fromdate)
        if todate:
            sql = sql.where(cset.c.insertion_date <= todate)
385
386
387
388
        if stripped:
            # outerjoin to show dead things
            sql = sql.select_from(cset.outerjoin(cset_series))
        else:
389
            sql = sql.where(cset.c.id == cset_series.c.cset
390
            ).where(cset_series.c.serie == reg.c.id)
391

392
        rset = cn.execute(sql)
393
        for csetid, author, revdate, meta in rset.fetchall():
394
395
            log.append({'rev': csetid, 'author': author,
                        'date': pd.Timestamp(revdate, tz='utc'),
396
                        'meta': meta or {},
397
                        'names': self._changeset_series(cn, csetid)})
398

399
        log.sort(key=lambda rev: rev['rev'])
400
401
        return log

402
403
404
405
406
407
408
409
410
411
412
413
414
    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')

415
416
    # /API
    # Helpers
417

418
419
    # creation / update

420
    def _create(self, cn, newts, seriename, author,
421
                metadata=None, insertion_date=None):
422
423
424
        # initial insertion
        if len(newts) == 0:
            return None
425
426
427
428
429
        # 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)
        )
430
        self._register_serie(cn, seriename, newts)
431
        snapshot = Snapshot(cn, self, seriename)
432
        csid = self._newchangeset(cn, author, insertion_date, metadata)
433
        head = snapshot.create(newts)
434
        start, end = start_end(newts)
435
436
        value = {
            'cset': csid,
437
438
439
            'snapshot': head,
            'start': start,
            'end': end
440
        }
Aurélien Campéas's avatar
Aurélien Campéas committed
441
        table = self._make_ts_table(cn, seriename)
442
        cn.execute(table.insert().values(value))
443
        self._finalize_insertion(cn, csid, seriename)
444
        L.info('first insertion of %s (size=%s) by %s',
445
               seriename, len(newts), author)
446
447
        return newts

448
    def _update(self, cn, table, newts, seriename, author,
449
                metadata=None, insertion_date=None):
450
451
        self._validate(cn, newts, seriename)
        snapshot = Snapshot(cn, self, seriename)
452
453
454
        diff = self.diff(snapshot.last(newts.index.min(),
                                       newts.index.max()),
                         newts)
455
456
        if not len(diff):
            L.info('no difference in %s by %s (for ts of size %s)',
457
                   seriename, author, len(newts))
458
459
            return

460
        csid = self._newchangeset(cn, author, insertion_date, metadata)
461
462
463
464
        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))
465
466
467
        head = snapshot.update(diff)
        value = {
            'cset': csid,
468
469
470
            'snapshot': head,
            'start': start,
            'end': end
471
472
        }
        cn.execute(table.insert().values(value))
473
        self._finalize_insertion(cn, csid, seriename)
474
475

        L.info('inserted diff (size=%s) for ts %s by %s',
476
               len(diff), seriename, author)
477
478
        return diff

479
    # serie table handling
480

481
482
483
484
    def _serie_to_tablename(self, cn, seriename):
        tablename = self.serie_tablename.get(seriename)
        if tablename is not None:
            return tablename
485

486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
        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)
502
        if table is None:
503
504
            TABLES[fq_tablename] = table = Table(
                tablename, self.schema.meta,
505
506
507
                Column('id', Integer, primary_key=True),
                Column('cset', Integer,
                       ForeignKey('{}.changeset.id'.format(self.namespace)),
508
                       nullable=False),
509
510
                Column('start', TIMESTAMP, nullable=False),
                Column('end', TIMESTAMP, nullable=False),
511
512
513
                Column('snapshot', Integer,
                       ForeignKey('{}.snapshot.{}.id'.format(
                           self.namespace,
514
515
516
                           tablename))),
                Index(NONE_NAME, 'cset'),
                Index(NONE_NAME, 'snapshot'),
517
518
                Index(NONE_NAME, 'start'),
                Index(NONE_NAME, 'end'),
519
                schema='{}.timeserie'.format(self.namespace),
520
                keep_existing=True
521
522
            )
        return table
523

Aurélien Campéas's avatar
Aurélien Campéas committed
524
    def _make_ts_table(self, cn, seriename):
525
        table = self._table_definition_for(cn, seriename)
526
        table.create(cn)
527
528
529
        return table

    def _register_serie(self, cn, seriename, ts):
530
531
        index = ts.index
        inames = [name for name in index.names if name]
532
        sql = self.schema.registry.insert().values(
533
534
            seriename=seriename,
            table_name=self._make_tablename(seriename),
535
536
537
538
            metadata={
                'tzaware': tzaware_serie(ts),
                'index_type': index.dtype.name,
                'index_names': inames,
539
540
                'index_dtype': index.dtype.str,
                'value_dtype': ts.dtypes.str,
541
                'value_type': ts.dtypes.name
542
            }
543
        )
544
545
        regid = cn.execute(sql).inserted_primary_key[0]
        self.registry_map[seriename] = regid
546

547
    def _get_ts_table(self, cn, seriename):
548
549
550
        tablename = self._serie_to_tablename(cn, seriename)
        if tablename:
            return self._table_definition_for(cn, tablename)
551

552
553
    # changeset handling

554
    def _newchangeset(self, cn, author, insertion_date=None, metadata=None):
555
        table = self.schema.changeset
556
557
558
        if insertion_date is not None:
            assert insertion_date.tzinfo is not None
        idate = pd.Timestamp(insertion_date or datetime.utcnow(), tz='UTC')
559
560
        sql = table.insert().values(
            author=author,
561
            metadata=metadata,
562
            insertion_date=idate)
563
        return cn.execute(sql).inserted_primary_key[0]
564

565
    def _changeset_series(self, cn, csid):
566
        cset_serie = self.schema.changeset_series
567
568
        reg = self.schema.registry
        sql = select(
569
            [reg.c.seriename]
570
571
        ).where(cset_serie.c.cset == csid
        ).where(cset_serie.c.serie == reg.c.id)
572

573
        return [
574
            row.seriename
575
576
            for row in cn.execute(sql).fetchall()
        ]
577

578
579
580
581
582
583
584
    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()

585
586
    # insertion handling

587
    def _validate(self, cn, ts, seriename):
588
589
        if ts.isnull().all():
            # ts erasure
590
            return
591
        tstype = ts.dtype
592
        meta = self.metadata(cn, seriename)
593
        if tstype != meta['value_type']:
594
            m = 'Type error when inserting {}, new type is {}, type in base is {}'.format(
595
                seriename, tstype, meta['value_type'])
596
            raise Exception(m)
597
        if ts.index.dtype.name != meta['index_type']:
598
            raise Exception('Incompatible index types')
599

600
601
    def _name_to_regid(self, cn, seriename):
        regid = self.registry_map.get(seriename)
602
603
604
605
        if regid is not None:
            return regid

        registry = self.schema.registry
606
        sql = select([registry.c.id]).where(registry.c.seriename == seriename)
607
        regid = self.registry_map[seriename] = cn.execute(sql).scalar()
608
609
        return regid

610
    def _finalize_insertion(self, cn, csid, seriename):
611
612
613
        table = self.schema.changeset_series
        sql = table.insert().values(
            cset=csid,
614
            serie=self._name_to_regid(cn, seriename)
615
616
        )
        cn.execute(sql)
617

618
619
620
    def _resetcaches(self):
        TABLES.clear()
        SNAPTABLES.clear()
621
622
623
        self.metadatacache.clear()
        self.registry_map.clear()
        self.serie_tablename.clear()