tsio.py 26.4 KB
Newer Older
1
from datetime import datetime
2
import logging
3
4
import hashlib
import uuid
5
from threading import Lock
6
7
8

import pandas as pd

9
10
from sqlalchemy import Table, Column, Integer, ForeignKey, Index
from sqlalchemy.sql.elements import NONE_NAME
11
from sqlalchemy.engine.base import Engine
Aurélien Campéas's avatar
Aurélien Campéas committed
12
from sqlalchemy.sql.expression import select, func, desc
Aurélien Campéas's avatar
Aurélien Campéas committed
13
from sqlalchemy.dialects.postgresql import TIMESTAMP
14

15
from tshistory.schema import tsschema
16
from tshistory.util import (
17
    closed_overlaps,
18
19
    num2float,
    SeriesServices,
20
    start_end,
21
22
    tzaware_serie
)
23
from tshistory.snapshot import Snapshot, TABLES as SNAPTABLES
24

25
L = logging.getLogger('tshistory.tsio')
26
TABLES = {}
27
28


29
30
31
32
33
34
35
36
37
38
39
def tx(func):
    def check_tx_and_call(self, cn, *a, **kw):
        # 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')

        return func(self, cn, *a, **kw)
    return check_tx_and_call


40
class TimeSerie(SeriesServices):
41
    namespace = 'tsh'
42
    schema = None
43
    metadatacache = None
44
45
46
47
48
49
50
51
    metakeys = {
        'tzaware',
        'index_type',
        'index_names',
        'index_dtype',
        'value_dtype',
        'value_type'
    }
52
    registry_map = None
53
    serie_tablename = None
54
    create_lock_id = None
55
    cachelock = Lock()
56
57
58

    def __init__(self, namespace='tsh'):
        self.namespace = namespace
59
60
        self.schema = tsschema(namespace)
        self.schema.define()
61
        self.metadatacache = {}
62
        self.registry_map = {}
63
        self.serie_tablename = {}
64
        self.create_lock_id = sum(ord(c) for c in namespace)
65

66
    @tx
67
    def insert(self, cn, newts, seriename, author,
68
69
               metadata=None,
               _insertion_date=None):
70
        """Create a new revision of a given time series
71

72
        newts: pandas.Series with date index
73
        seriename: str unique identifier of the serie
74
        author: str free-form author name
75
        metadata: optional dict for changeset metadata
76
        """
77
78
79
80
        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'
81
82
        assert (_insertion_date is None or
                isinstance(_insertion_date, datetime)), 'Bad format for insertion date'
83
        assert not newts.index.duplicated().any(), 'There are some duplicates in the index'
84

85
        assert newts.index.notna().all(), 'The index contains NaT entries'
86
87
        if not newts.index.is_monotonic_increasing:
            newts = newts.sort_index()
88

89
        newts = num2float(newts)
90

91
        if not len(newts):
92
            return
93

94
        assert ('<M8[ns]' == newts.index.dtype or
95
                'datetime' in str(newts.index.dtype) and not
96
97
                isinstance(newts.index, pd.MultiIndex))

98
99
        newts.name = seriename
        table = self._get_ts_table(cn, seriename)
100

101
        if table is None:
102
            return self._create(cn, newts, seriename, author,
103
                                metadata, _insertion_date)
104

105
        return self._update(cn, table, newts, seriename, author,
106
                            metadata, _insertion_date)
107

108
    def get(self, cn, seriename, revision_date=None,
109
110
            from_value_date=None, to_value_date=None,
            _keep_nans=False):
111
112
113
114
115
        """Compute and return the serie of a given name

        revision_date: datetime filter to get previous versions of the
        serie

116
        """
Aurélien Campéas's avatar
Aurélien Campéas committed
117
        if not self.exists(cn, seriename):
118
            return
119

120
        csetfilter = []
121
        if revision_date:
122
            csetfilter.append(lambda cset: cset.c.insertion_date <= revision_date)
123
        snap = Snapshot(cn, self, seriename)
124
        _, current = snap.find(csetfilter=csetfilter,
125
126
                               from_value_date=from_value_date,
                               to_value_date=to_value_date)
127

128
        if current is not None and not _keep_nans:
129
            current.name = seriename
130
            current = current[~current.isnull()]
131
        return current
132

133
    def metadata(self, cn, seriename):
134
        """Return metadata dict of timeserie."""
135
136
        if seriename in self.metadatacache:
            return self.metadatacache[seriename]
137
138
        reg = self.schema.registry
        sql = select([reg.c.metadata]).where(
139
            reg.c.seriename == seriename
140
141
        )
        meta = cn.execute(sql).scalar()
142
        self.metadatacache[seriename] = meta
143
144
        return meta

145
    @tx
146
    def update_metadata(self, cn, seriename, metadata):
147
        assert isinstance(metadata, dict)
148
        assert not set(metadata.keys()) & self.metakeys
149
        meta = self.metadata(cn, seriename)
150
151
152
        # remove al but internal stuff
        newmeta = {key: meta[key] for key in self.metakeys}
        newmeta.update(metadata)
153
154
        reg = self.schema.registry
        sql = reg.update().where(
155
            reg.c.seriename == seriename
156
        ).values(metadata=newmeta)
157
        self.metadatacache.pop(seriename)
158
159
        cn.execute(sql)

160
161
162
163
164
165
166
    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()

167
168
169
    def type(self, cn, name):
        return 'primary'

170
    @tx
171
    def get_history(self, cn, seriename,
172
                    from_insertion_date=None,
173
174
                    to_insertion_date=None,
                    from_value_date=None,
175
                    to_value_date=None,
176
                    deltabefore=None,
177
                    deltaafter=None,
178
179
                    diffmode=False,
                    _keep_nans=False):
180
        table = self._get_ts_table(cn, seriename)
181
182
183
        if table is None:
            return

184
        cset = self.schema.changeset
185
186
187
188
189
190
191
192
193
194
195
196
        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)
197

198
199
200
201
202
203
204
205
206
        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()
207
        if not revs:
208
            return {}
209

210
211
212
213
214
215
        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))

216
        snapshot = Snapshot(cn, self, seriename)
217
        series = []
218
219
        if (deltabefore, deltaafter) != (None, None):
            for csid, idate in revs:
220
221
                from_date = None
                to_date = None
222
                if deltabefore is not None:
223
                    from_date = idate - deltabefore
224
                if deltaafter is not None:
225
                    to_date = idate + deltaafter
226
227
228
                series.append((
                    idate,
                    snapshot.find(csetfilter=[lambda cset: cset.c.id == csid],
229
230
                                  from_value_date=from_date,
                                  to_value_date=to_date)[1]
231
232
233
234
235
236
                ))
        else:
            series = snapshot.findall(revs,
                                      from_value_date,
                                      to_value_date
            )
237

238
239
        if diffmode:
            diffs = []
Aurélien Campéas's avatar
Aurélien Campéas committed
240
            for (_revdate_a, serie_a), (revdate_b, serie_b) in zip(series, series[1:]):
241
242
243
244
245
246
247
                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
248
249
        else:
            series = [
250
                (idate, ts if _keep_nans else ts.dropna() )
251
252
                 for idate, ts in series
            ]
253

254
        return {
255
            pd.Timestamp(idate).astimezone('UTC'): serie
256
257
            for idate, serie in series
        }
258

259
    @tx
260
261
262
    def get_delta(self, cn, seriename, delta,
                  from_value_date=None,
                  to_value_date=None):
263
264
265
266
        """ compute a series whose value dates are bounded to be
        `delta` time after the insertion dates and where we
        keep the most recent ones
        """
267
        histo = self.get_history(
268
269
            cn, seriename, deltabefore=-delta,
            from_value_date=from_value_date,
270
271
            to_value_date=to_value_date,
            _keep_nans=True
272
        )
273
274
        if histo is None:
            return None
275

276
277
278
279
280
281
282
283
        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

284
        ts = pd.Series(vvmap).sort_index().loc[from_value_date:to_value_date]
285
        ts.name = seriename
286
        return ts.dropna()
287

288
289
    def exists(self, cn, seriename):
        return self._get_ts_table(cn, seriename) is not None
290

291
    def latest_insertion_date(self, cn, seriename):
292
        cset = self.schema.changeset
293
        tstable = self._get_ts_table(cn, seriename)
294
295
296
297
298
299
300
301
        sql = select(
            [func.max(cset.c.insertion_date)]
        ).where(
            tstable.c.cset == cset.c.id
        )
        return pd.Timestamp(
            cn.execute(sql).scalar()
        ).astimezone('UTC')
302

303
304
    def insertion_dates(self, cn, seriename,
                        fromdate=None, todate=None):
305
306
307
308
309
310
311
312
313
        cset = self.schema.changeset
        tstable = self._get_ts_table(cn, seriename)
        sql = select(
            [cset.c.insertion_date]
        ).where(
            tstable.c.cset == cset.c.id
        ).order_by(
            cset.c.id
        )
314
315
316
317
318
319

        if fromdate:
            sql = sql.where(cset.c.insertion_date >= fromdate)
        if todate:
            sql = sql.where(cset.c.insertion_date <= todate)

320
321
322
323
324
        return [
            pd.Timestamp(idate).astimezone('UTC')
            for idate, in cn.execute(sql).fetchall()
        ]

325
326
327
328
    def last_id(self, cn, seriename):
        snapshot = Snapshot(cn, self, seriename)
        return snapshot.last_id()

329
330
331
    def changeset_at(self, cn, seriename, revdate, mode='strict'):
        assert mode in ('strict', 'before', 'after')
        cset = self.schema.changeset
332
        table = self._table_definition_for(cn, seriename)
333
334
        sql = select([table.c.cset]).where(
            table.c.cset == cset.c.id
335
336
337
338
339
340
341
342
343
        )
        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()

344
    @tx
345
    def delete(self, cn, seriename):
346
347
348
        if not self.exists(cn, seriename):
            print('not deleting unknown series', seriename, self.namespace)
            return
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
        # 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()
379
        print('deleted', seriename, self.namespace)
380

381
    @tx
382
383
384
385
386
387
388
389
390
    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
391
            metadata = self.changeset_metadata(cn, log['rev']) or {}
392
393
394
395
396
397
            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(
398
                cset_serie.c.cset == log['rev']
399
            ).where(
400
                cset_serie.c.serie == self._name_to_regid(cn, seriename)
401
402
403
404
            )
            cn.execute(sql)

        # wipe the diffs
405
        table = self._table_definition_for(cn, seriename)
406
        cn.execute(table.delete().where(table.c.cset >= csid))
407

408
    def info(self, cn):
409
410
        """Gather global statistics on the current tshistory repository
        """
411
        sql = 'select count(*) from "{}".registry'.format(self.namespace)
412
        stats = {'series count': cn.execute(sql).scalar()}
413
        sql = 'select max(id) from "{}".changeset'.format(self.namespace)
414
        stats['changeset count'] = cn.execute(sql).scalar()
415
416
417
        sql = 'select distinct seriename from "{}".registry order by seriename'.format(
            self.namespace
        )
418
        stats['serie names'] = [row for row, in cn.execute(sql).fetchall()]
419
420
        return stats

421
    def log(self, cn, limit=0, names=None, authors=None,
422
            stripped=False,
423
424
            fromrev=None, torev=None,
            fromdate=None, todate=None):
425
426
427
428
        """Build a structure showing the history of all the series in the db,
        per changeset, in chronological order.
        """
        log = []
429
430
431
432
433
        cset, cset_series, reg = (
            self.schema.changeset,
            self.schema.changeset_series,
            self.schema.registry
        )
434

435
        sql = select([cset.c.id, cset.c.author, cset.c.insertion_date, cset.c.metadata]
436
        ).distinct().order_by(desc(cset.c.id))
437
438
439

        if limit:
            sql = sql.limit(limit)
440
        if names:
441
            sql = sql.where(reg.c.seriename.in_(names))
442
443
        if authors:
            sql = sql.where(cset.c.author.in_(authors))
444
445
446
447
        if fromrev:
            sql = sql.where(cset.c.id >= fromrev)
        if torev:
            sql = sql.where(cset.c.id <= torev)
448
449
450
451
        if fromdate:
            sql = sql.where(cset.c.insertion_date >= fromdate)
        if todate:
            sql = sql.where(cset.c.insertion_date <= todate)
452
453
454
455
        if stripped:
            # outerjoin to show dead things
            sql = sql.select_from(cset.outerjoin(cset_series))
        else:
456
            sql = sql.where(cset.c.id == cset_series.c.cset
457
            ).where(cset_series.c.serie == reg.c.id)
458

459
        rset = cn.execute(sql)
460
        for csetid, author, revdate, meta in rset.fetchall():
461
462
            log.append({'rev': csetid, 'author': author,
                        'date': pd.Timestamp(revdate, tz='utc'),
463
                        'meta': meta or {},
464
                        'names': self._changeset_series(cn, csetid)})
465

466
        log.sort(key=lambda rev: rev['rev'])
467
468
        return log

469
    def interval(self, cn, seriename, notz=False):
470
471
472
473
474
475
476
477
        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
478
        if self.metadata(cn, seriename).get('tzaware') and not notz:
479
480
481
            start, end = pd.Timestamp(start, tz='UTC'), pd.Timestamp(end, tz='UTC')
        return pd.Interval(left=start, right=end, closed='both')

482
483
    # /API
    # Helpers
484

485
486
    # creation / update

487
    def _create(self, cn, newts, seriename, author,
488
                metadata=None, insertion_date=None):
489
        start, end = start_end(newts, notz=False)
490
491
492
493
        if start is None:
            assert end is None
            # this is just full of nans
            return None
494
495
496
497
498
        # chop off unwanted nans
        newts = newts.loc[start:end]
        if len(newts) == 0:
            return None

499
500
501
502
503
        # 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)
        )
504
        self._register_serie(cn, seriename, newts)
505
        snapshot = Snapshot(cn, self, seriename)
506
        csid = self._newchangeset(cn, author, insertion_date, metadata)
507
        head = snapshot.create(newts)
508
        start, end = start_end(newts)
509
510
        value = {
            'cset': csid,
511
512
513
            'snapshot': head,
            'start': start,
            'end': end
514
        }
Aurélien Campéas's avatar
Aurélien Campéas committed
515
        table = self._make_ts_table(cn, seriename)
516
        cn.execute(table.insert().values(value))
517
        self._finalize_insertion(cn, csid, seriename)
518
        L.info('first insertion of %s (size=%s) by %s',
519
               seriename, len(newts), author)
520
521
        return newts

522
    def _update(self, cn, table, newts, seriename, author,
523
                metadata=None, insertion_date=None):
524
525
        self._validate(cn, newts, seriename)
        snapshot = Snapshot(cn, self, seriename)
526
527
528
        diff = self.diff(snapshot.last(newts.index.min(),
                                       newts.index.max()),
                         newts)
529
530
        if not len(diff):
            L.info('no difference in %s by %s (for ts of size %s)',
531
                   seriename, author, len(newts))
532
533
            return

534
        # compute series start/end stamps
535
        tsstart, tsend = start_end(newts)
536
537
538
        ival = self.interval(cn, seriename, notz=True)
        start = min(tsstart or ival.left, ival.left)
        end = max(tsend or ival.right, ival.right)
539
540
541
542
543
544
545
546
547
548
549
550
551
552

        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)
553
554
555
        head = snapshot.update(diff)
        value = {
            'cset': csid,
556
557
558
            'snapshot': head,
            'start': start,
            'end': end
559
560
        }
        cn.execute(table.insert().values(value))
561
        self._finalize_insertion(cn, csid, seriename)
562
563

        L.info('inserted diff (size=%s) for ts %s by %s',
564
               len(diff), seriename, author)
565
566
        return diff

567
    # serie table handling
568

569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
    def _make_tablename(self, cn, seriename):
        """ compute the unqualified (no namespace) table name
        from a serie name, to allow arbitrary serie names
        """
        # postgresql table names are limited to 63 chars.
        if len(seriename) > 63:
            seriename = hashlib.sha1(seriename.encode('utf-8')).hexdigest()

        # collision detection (collision can happen after a rename)
        if cn.execute(f'select table_name '
                      f'from "{self.namespace}".registry '
                      f'where table_name = %(seriename)s',
                      seriename=seriename).scalar():
            return str(uuid.uuid4())

        return seriename

586
587
588
589
    def _serie_to_tablename(self, cn, seriename):
        tablename = self.serie_tablename.get(seriename)
        if tablename is not None:
            return tablename
590

591
592
593
594
595
596
597
598
599
600
601
602
603
        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
Aurélien Campéas's avatar
Aurélien Campéas committed
604
            tablename = self._make_tablename(cn, seriename)
605
606
        fq_tablename = '{}.timeserie.{}'.format(self.namespace, tablename)
        table = TABLES.get(fq_tablename)
607
        if table is None:
608
609
            TABLES[fq_tablename] = table = Table(
                tablename, self.schema.meta,
610
611
612
                Column('id', Integer, primary_key=True),
                Column('cset', Integer,
                       ForeignKey('{}.changeset.id'.format(self.namespace)),
613
                       nullable=False),
614
615
                Column('start', TIMESTAMP, nullable=False),
                Column('end', TIMESTAMP, nullable=False),
616
617
618
                Column('snapshot', Integer,
                       ForeignKey('{}.snapshot.{}.id'.format(
                           self.namespace,
619
620
621
                           tablename))),
                Index(NONE_NAME, 'cset'),
                Index(NONE_NAME, 'snapshot'),
622
623
                Index(NONE_NAME, 'start'),
                Index(NONE_NAME, 'end'),
624
                schema='{}.timeserie'.format(self.namespace),
625
                keep_existing=True
626
627
            )
        return table
628

Aurélien Campéas's avatar
Aurélien Campéas committed
629
    def _make_ts_table(self, cn, seriename):
630
        table = self._table_definition_for(cn, seriename)
631
        table.create(cn)
632
633
634
        return table

    def _register_serie(self, cn, seriename, ts):
635
636
        index = ts.index
        inames = [name for name in index.names if name]
637
        sql = self.schema.registry.insert().values(
638
            seriename=seriename,
Aurélien Campéas's avatar
Aurélien Campéas committed
639
            table_name=self._make_tablename(cn, seriename),
640
641
642
643
            metadata={
                'tzaware': tzaware_serie(ts),
                'index_type': index.dtype.name,
                'index_names': inames,
644
645
                'index_dtype': index.dtype.str,
                'value_dtype': ts.dtypes.str,
646
                'value_type': ts.dtypes.name
647
            }
648
        )
649
650
        regid = cn.execute(sql).inserted_primary_key[0]
        self.registry_map[seriename] = regid
651

652
    def _get_ts_table(self, cn, seriename):
653
654
        tablename = self._serie_to_tablename(cn, seriename)
        if tablename:
Aurélien Campéas's avatar
Aurélien Campéas committed
655
            return self._table_definition_for(cn, seriename)
656

657
658
    # changeset handling

659
    def _newchangeset(self, cn, author, insertion_date=None, metadata=None):
660
        table = self.schema.changeset
661
662
663
        if insertion_date is not None:
            assert insertion_date.tzinfo is not None
        idate = pd.Timestamp(insertion_date or datetime.utcnow(), tz='UTC')
664
665
        sql = table.insert().values(
            author=author,
666
            metadata=metadata,
667
            insertion_date=idate)
668
        return cn.execute(sql).inserted_primary_key[0]
669

670
    def _changeset_series(self, cn, csid):
671
        cset_serie = self.schema.changeset_series
672
673
        reg = self.schema.registry
        sql = select(
674
            [reg.c.seriename]
675
676
        ).where(cset_serie.c.cset == csid
        ).where(cset_serie.c.serie == reg.c.id)
677

678
        return [
679
            row.seriename
680
681
            for row in cn.execute(sql).fetchall()
        ]
682

683
684
685
686
687
688
689
    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()

690
691
    # insertion handling

692
    def _validate(self, cn, ts, seriename):
693
694
        if ts.isnull().all():
            # ts erasure
695
            return
696
        tstype = ts.dtype
697
        meta = self.metadata(cn, seriename)
698
        if tstype != meta['value_type']:
699
            m = 'Type error when inserting {}, new type is {}, type in base is {}'.format(
700
                seriename, tstype, meta['value_type'])
701
            raise Exception(m)
702
        if ts.index.dtype.name != meta['index_type']:
703
            raise Exception('Incompatible index types')
704

705
706
    def _name_to_regid(self, cn, seriename):
        regid = self.registry_map.get(seriename)
707
708
709
710
        if regid is not None:
            return regid

        registry = self.schema.registry
711
        sql = select([registry.c.id]).where(registry.c.seriename == seriename)
712
        regid = self.registry_map[seriename] = cn.execute(sql).scalar()
713
714
        return regid

715
    def _finalize_insertion(self, cn, csid, seriename):
716
717
718
        table = self.schema.changeset_series
        sql = table.insert().values(
            cset=csid,
719
            serie=self._name_to_regid(cn, seriename)
720
721
        )
        cn.execute(sql)
722

723
    def _resetcaches(self):
724
725
726
727
728
729
        with self.cachelock:
            TABLES.clear()
            SNAPTABLES.clear()
            self.metadatacache.clear()
            self.registry_map.clear()
            self.serie_tablename.clear()