mz_persist_client/
fetch.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
// Copyright Materialize, Inc. and contributors. All rights reserved.
//
// Use of this software is governed by the Business Source License
// included in the LICENSE file.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0.

//! Fetching batches of data from persist's backing store

use std::fmt::{self, Debug};
use std::marker::PhantomData;
use std::sync::Arc;
use std::time::Instant;

use anyhow::anyhow;
use arrow::array::{Array, AsArray, BooleanArray};
use arrow::compute::FilterBuilder;
use differential_dataflow::difference::Semigroup;
use differential_dataflow::lattice::Lattice;
use differential_dataflow::trace::Description;
use mz_dyncfg::{Config, ConfigSet, ConfigValHandle};
use mz_ore::bytes::SegmentedBytes;
use mz_ore::cast::CastFrom;
use mz_ore::{soft_panic_no_log, soft_panic_or_log};
use mz_persist::indexed::columnar::arrow::{realloc_any, realloc_array};
use mz_persist::indexed::columnar::{ColumnarRecords, ColumnarRecordsStructuredExt};
use mz_persist::indexed::encoding::{BlobTraceBatchPart, BlobTraceUpdates};
use mz_persist::location::{Blob, SeqNo};
use mz_persist::metrics::ColumnarMetrics;
use mz_persist_types::columnar::{ColumnDecoder, Schema2};
use mz_persist_types::stats::PartStats;
use mz_persist_types::{Codec, Codec64};
use mz_proto::{IntoRustIfSome, ProtoType, RustType, TryFromProtoError};
use serde::{Deserialize, Serialize};
use timely::progress::frontier::AntichainRef;
use timely::progress::{Antichain, Timestamp};
use timely::PartialOrder;
use tracing::{debug_span, trace_span, Instrument};

use crate::batch::{
    proto_fetch_batch_filter, ProtoFetchBatchFilter, ProtoFetchBatchFilterListen, ProtoLease,
    ProtoLeasedBatchPart,
};
use crate::cfg::PersistConfig;
use crate::error::InvalidUsage;
use crate::internal::encoding::{LazyInlineBatchPart, LazyPartStats, LazyProto, Schemas};
use crate::internal::machine::retry_external;
use crate::internal::metrics::{Metrics, MetricsPermits, ReadMetrics, ShardMetrics};
use crate::internal::paths::BlobKey;
use crate::internal::state::{BatchPart, HollowBatchPart};
use crate::project::ProjectionPushdown;
use crate::read::LeasedReaderId;
use crate::schema::{PartMigration, SchemaCache};
use crate::ShardId;

pub(crate) const FETCH_SEMAPHORE_COST_ADJUSTMENT: Config<f64> = Config::new(
    "persist_fetch_semaphore_cost_adjustment",
    // We use `encoded_size_bytes` as the number of permits, but the parsed size
    // is larger than the encoded one, so adjust it. This default value is from
    // eyeballing graphs in experiments that were run on tpch loadgen data.
    1.2,
    "\
    An adjustment multiplied by encoded_size_bytes to approximate an upper \
    bound on the size in lgalloc, which includes the decoded version.",
);

pub(crate) const FETCH_SEMAPHORE_PERMIT_ADJUSTMENT: Config<f64> = Config::new(
    "persist_fetch_semaphore_permit_adjustment",
    1.0,
    "\
    A limit on the number of outstanding persist bytes being fetched and \
    parsed, expressed as a multiplier of the process's memory limit. This data \
    all spills to lgalloc, so values > 1.0 are safe. Only applied to cc \
    replicas.",
);

pub(crate) const PART_DECODE_FORMAT: Config<&'static str> = Config::new(
    "persist_part_decode_format",
    PartDecodeFormat::default().as_str(),
    "\
    Format we'll use to decode a Persist Part, either 'row', \
    'row_with_validate', or 'arrow' (Materialize).",
);

#[derive(Debug, Clone)]
pub(crate) struct BatchFetcherConfig {
    pub(crate) part_decode_format: ConfigValHandle<String>,
}

impl BatchFetcherConfig {
    pub fn new(value: &PersistConfig) -> Self {
        BatchFetcherConfig {
            part_decode_format: PART_DECODE_FORMAT.handle(value),
        }
    }

    pub fn part_decode_format(&self) -> PartDecodeFormat {
        PartDecodeFormat::from_str(self.part_decode_format.get().as_str())
    }
}

/// Capable of fetching [`LeasedBatchPart`] while not holding any capabilities.
#[derive(Debug)]
pub struct BatchFetcher<K, V, T, D>
where
    T: Timestamp + Lattice + Codec64,
    // These are only here so we can use them in the auto-expiring `Drop` impl.
    K: Debug + Codec,
    V: Debug + Codec,
    D: Semigroup + Codec64 + Send + Sync,
{
    pub(crate) cfg: BatchFetcherConfig,
    pub(crate) blob: Arc<dyn Blob>,
    pub(crate) metrics: Arc<Metrics>,
    pub(crate) shard_metrics: Arc<ShardMetrics>,
    pub(crate) shard_id: ShardId,
    pub(crate) read_schemas: Schemas<K, V>,
    pub(crate) schema_cache: SchemaCache<K, V, T, D>,
    pub(crate) is_transient: bool,

    // Ensures that `BatchFetcher` is of the same type as the `ReadHandle` it's
    // derived from.
    pub(crate) _phantom: PhantomData<fn() -> (K, V, T, D)>,
}

impl<K, V, T, D> BatchFetcher<K, V, T, D>
where
    K: Debug + Codec,
    V: Debug + Codec,
    T: Timestamp + Lattice + Codec64,
    D: Semigroup + Codec64 + Send + Sync,
{
    /// Takes a [`SerdeLeasedBatchPart`] into a [`LeasedBatchPart`].
    pub fn leased_part_from_exchangeable(&self, x: SerdeLeasedBatchPart) -> LeasedBatchPart<T> {
        x.decode(Arc::clone(&self.metrics))
    }

    /// Trade in an exchange-able [LeasedBatchPart] for the data it represents.
    ///
    /// Note to check the `LeasedBatchPart` documentation for how to handle the
    /// returned value.
    pub async fn fetch_leased_part(
        &mut self,
        part: &LeasedBatchPart<T>,
    ) -> Result<FetchedBlob<K, V, T, D>, InvalidUsage<T>> {
        if &part.shard_id != &self.shard_id {
            let batch_shard = part.shard_id.clone();
            return Err(InvalidUsage::BatchNotFromThisShard {
                batch_shard,
                handle_shard: self.shard_id.clone(),
            });
        }

        let migration = PartMigration::new(
            part.part.schema_id(),
            self.read_schemas.clone(),
            &mut self.schema_cache,
        )
        .await
        .unwrap_or_else(|read_schemas| {
            panic!(
                "could not decode part {:?} with schema: {:?}",
                part.part.schema_id(),
                read_schemas
            )
        });

        let (buf, fetch_permit) = match &part.part {
            BatchPart::Hollow(x) => {
                let fetch_permit = self
                    .metrics
                    .semaphore
                    .acquire_fetch_permits(x.encoded_size_bytes)
                    .await;
                let read_metrics = if self.is_transient {
                    &self.metrics.read.unindexed
                } else {
                    &self.metrics.read.batch_fetcher
                };
                let buf = fetch_batch_part_blob(
                    &part.shard_id,
                    self.blob.as_ref(),
                    &self.metrics,
                    &self.shard_metrics,
                    read_metrics,
                    x,
                )
                .await
                .unwrap_or_else(|blob_key| {
                    // Ideally, readers should never encounter a missing blob. They place a seqno
                    // hold as they consume their snapshot/listen, preventing any blobs they need
                    // from being deleted by garbage collection, and all blob implementations are
                    // linearizable so there should be no possibility of stale reads.
                    //
                    // If we do have a bug and a reader does encounter a missing blob, the state
                    // cannot be recovered, and our best option is to panic and retry the whole
                    // process.
                    panic!("batch fetcher could not fetch batch part: {}", blob_key)
                });
                let buf = FetchedBlobBuf::Hollow {
                    buf,
                    part: x.clone(),
                };
                (buf, Some(Arc::new(fetch_permit)))
            }
            BatchPart::Inline {
                updates,
                ts_rewrite,
                ..
            } => {
                let buf = FetchedBlobBuf::Inline {
                    desc: part.desc.clone(),
                    updates: updates.clone(),
                    ts_rewrite: ts_rewrite.clone(),
                };
                (buf, None)
            }
        };
        let fetched_blob = FetchedBlob {
            metrics: Arc::clone(&self.metrics),
            read_metrics: self.metrics.read.batch_fetcher.clone(),
            buf,
            registered_desc: part.desc.clone(),
            migration,
            filter: part.filter.clone(),
            filter_pushdown_audit: part.filter_pushdown_audit,
            structured_part_audit: self.cfg.part_decode_format(),
            fetch_permit,
            _phantom: PhantomData,
        };
        Ok(fetched_blob)
    }
}

#[derive(Debug, Clone)]
pub(crate) enum FetchBatchFilter<T> {
    Snapshot {
        as_of: Antichain<T>,
    },
    Listen {
        as_of: Antichain<T>,
        lower: Antichain<T>,
    },
    Compaction {
        since: Antichain<T>,
    },
}

impl<T: Timestamp + Lattice> FetchBatchFilter<T> {
    pub(crate) fn filter_ts(&self, t: &mut T) -> bool {
        match self {
            FetchBatchFilter::Snapshot { as_of } => {
                // This time is covered by a listen
                if as_of.less_than(t) {
                    return false;
                }
                t.advance_by(as_of.borrow());
                true
            }
            FetchBatchFilter::Listen { as_of, lower } => {
                // This time is covered by a snapshot
                if !as_of.less_than(t) {
                    return false;
                }

                // Because of compaction, the next batch we get might also
                // contain updates we've already emitted. For example, we
                // emitted `[1, 2)` and then compaction combined that batch with
                // a `[2, 3)` batch into a new `[1, 3)` batch. If this happens,
                // we just need to filter out anything < the frontier. This
                // frontier was the upper of the last batch (and thus exclusive)
                // so for the == case, we still emit.
                if !lower.less_equal(t) {
                    return false;
                }
                true
            }
            FetchBatchFilter::Compaction { since } => {
                t.advance_by(since.borrow());
                true
            }
        }
    }
}

impl<T: Timestamp + Codec64> RustType<ProtoFetchBatchFilter> for FetchBatchFilter<T> {
    fn into_proto(&self) -> ProtoFetchBatchFilter {
        let kind = match self {
            FetchBatchFilter::Snapshot { as_of } => {
                proto_fetch_batch_filter::Kind::Snapshot(as_of.into_proto())
            }
            FetchBatchFilter::Listen { as_of, lower } => {
                proto_fetch_batch_filter::Kind::Listen(ProtoFetchBatchFilterListen {
                    as_of: Some(as_of.into_proto()),
                    lower: Some(lower.into_proto()),
                })
            }
            FetchBatchFilter::Compaction { .. } => unreachable!("not serialized"),
        };
        ProtoFetchBatchFilter { kind: Some(kind) }
    }

    fn from_proto(proto: ProtoFetchBatchFilter) -> Result<Self, TryFromProtoError> {
        let kind = proto
            .kind
            .ok_or_else(|| TryFromProtoError::missing_field("ProtoFetchBatchFilter::kind"))?;
        match kind {
            proto_fetch_batch_filter::Kind::Snapshot(as_of) => Ok(FetchBatchFilter::Snapshot {
                as_of: as_of.into_rust()?,
            }),
            proto_fetch_batch_filter::Kind::Listen(ProtoFetchBatchFilterListen {
                as_of,
                lower,
            }) => Ok(FetchBatchFilter::Listen {
                as_of: as_of.into_rust_if_some("ProtoFetchBatchFilterListen::as_of")?,
                lower: lower.into_rust_if_some("ProtoFetchBatchFilterListen::lower")?,
            }),
        }
    }
}

/// Trade in an exchange-able [LeasedBatchPart] for the data it represents.
///
/// Note to check the `LeasedBatchPart` documentation for how to handle the
/// returned value.
pub(crate) async fn fetch_leased_part<K, V, T, D>(
    cfg: &PersistConfig,
    part: &LeasedBatchPart<T>,
    blob: &dyn Blob,
    metrics: Arc<Metrics>,
    read_metrics: &ReadMetrics,
    shard_metrics: &ShardMetrics,
    reader_id: &LeasedReaderId,
    read_schemas: Schemas<K, V>,
    schema_cache: &mut SchemaCache<K, V, T, D>,
) -> FetchedPart<K, V, T, D>
where
    K: Debug + Codec,
    V: Debug + Codec,
    T: Timestamp + Lattice + Codec64,
    D: Semigroup + Codec64 + Send + Sync,
{
    let encoded_part = EncodedPart::fetch(
        &part.shard_id,
        blob,
        &metrics,
        shard_metrics,
        read_metrics,
        &part.desc,
        &part.part,
    )
    .await
    .unwrap_or_else(|blob_key| {
        // Ideally, readers should never encounter a missing blob. They place a seqno
        // hold as they consume their snapshot/listen, preventing any blobs they need
        // from being deleted by garbage collection, and all blob implementations are
        // linearizable so there should be no possibility of stale reads.
        //
        // If we do have a bug and a reader does encounter a missing blob, the state
        // cannot be recovered, and our best option is to panic and retry the whole
        // process.
        panic!("{} could not fetch batch part: {}", reader_id, blob_key)
    });
    let part_cfg = BatchFetcherConfig::new(cfg);
    let migration = PartMigration::new(part.part.schema_id(), read_schemas, schema_cache)
        .await
        .unwrap_or_else(|read_schemas| {
            panic!(
                "could not decode part {:?} with schema: {:?}",
                part.part.schema_id(),
                read_schemas
            )
        });
    FetchedPart::new(
        metrics,
        encoded_part,
        migration,
        part.filter.clone(),
        part.filter_pushdown_audit,
        part_cfg.part_decode_format(),
        part.part.stats(),
    )
}

pub(crate) async fn fetch_batch_part_blob<T>(
    shard_id: &ShardId,
    blob: &dyn Blob,
    metrics: &Metrics,
    shard_metrics: &ShardMetrics,
    read_metrics: &ReadMetrics,
    part: &HollowBatchPart<T>,
) -> Result<SegmentedBytes, BlobKey> {
    let now = Instant::now();
    let get_span = debug_span!("fetch_batch::get");
    let blob_key = part.key.complete(shard_id);
    let value = retry_external(&metrics.retries.external.fetch_batch_get, || async {
        shard_metrics.blob_gets.inc();
        blob.get(&blob_key).await
    })
    .instrument(get_span.clone())
    .await
    .ok_or(blob_key)?;

    drop(get_span);

    read_metrics.part_count.inc();
    read_metrics.part_bytes.inc_by(u64::cast_from(value.len()));
    read_metrics.seconds.inc_by(now.elapsed().as_secs_f64());

    Ok(value)
}

pub(crate) fn decode_batch_part_blob<T>(
    metrics: &Metrics,
    read_metrics: &ReadMetrics,
    registered_desc: Description<T>,
    part: &HollowBatchPart<T>,
    buf: &SegmentedBytes,
) -> EncodedPart<T>
where
    T: Timestamp + Lattice + Codec64,
{
    trace_span!("fetch_batch::decode").in_scope(|| {
        let parsed = metrics
            .codecs
            .batch
            .decode(|| BlobTraceBatchPart::decode(buf, &metrics.columnar))
            .map_err(|err| anyhow!("couldn't decode batch at key {}: {}", part.key, err))
            // We received a State that we couldn't decode. This could happen if
            // persist messes up backward/forward compatibility, if the durable
            // data was corrupted, or if operations messes up deployment. In any
            // case, fail loudly.
            .expect("internal error: invalid encoded state");
        read_metrics
            .part_goodbytes
            .inc_by(u64::cast_from(parsed.updates.records().goodbytes()));
        EncodedPart::from_hollow(read_metrics.clone(), registered_desc, part, parsed)
    })
}

pub(crate) async fn fetch_batch_part<T>(
    shard_id: &ShardId,
    blob: &dyn Blob,
    metrics: &Metrics,
    shard_metrics: &ShardMetrics,
    read_metrics: &ReadMetrics,
    registered_desc: &Description<T>,
    part: &HollowBatchPart<T>,
) -> Result<EncodedPart<T>, BlobKey>
where
    T: Timestamp + Lattice + Codec64,
{
    let buf =
        fetch_batch_part_blob(shard_id, blob, metrics, shard_metrics, read_metrics, part).await?;
    let part = decode_batch_part_blob(metrics, read_metrics, registered_desc.clone(), part, &buf);
    Ok(part)
}

/// This represents the lease of a seqno. It's generally paired with some external state,
/// like a hollow part: holding this lease indicates that we may still want to fetch that part,
/// and should hold back GC to keep it around.
///
/// Generally the state and lease are bundled together, as in [LeasedBatchPart]... but sometimes
/// it's necessary to handle them separately, so this struct is exposed as well. Handle with care.
#[derive(Clone, Debug, Default)]
pub(crate) struct Lease(Arc<()>);

impl Lease {
    /// Returns the number of live copies of this lease, including this one.
    pub fn count(&self) -> usize {
        Arc::strong_count(&self.0)
    }
}

/// A token representing one fetch-able batch part.
///
/// It is tradeable via `crate::fetch::fetch_batch` for the resulting data
/// stored in the part.
///
/// # Exchange
///
/// You can exchange `LeasedBatchPart`:
/// - If `leased_seqno.is_none()`
/// - By converting it to [`SerdeLeasedBatchPart`] through
///   `Self::into_exchangeable_part`. [`SerdeLeasedBatchPart`] is exchangeable,
///   including over the network.
///
/// n.b. `Self::into_exchangeable_part` is known to be equivalent to
/// `SerdeLeasedBatchPart::from(self)`, but we want the additional warning message to
/// be visible and sufficiently scary.
///
/// # Panics
/// `LeasedBatchPart` panics when dropped unless a very strict set of invariants are
/// held:
///
/// `LeasedBatchPart` may only be dropped if it:
/// - Does not have a leased `SeqNo (i.e. `self.leased_seqno.is_none()`)
///
/// In any other circumstance, dropping `LeasedBatchPart` panics.
#[derive(Debug)]
pub struct LeasedBatchPart<T> {
    pub(crate) metrics: Arc<Metrics>,
    pub(crate) shard_id: ShardId,
    pub(crate) reader_id: LeasedReaderId,
    pub(crate) filter: FetchBatchFilter<T>,
    pub(crate) desc: Description<T>,
    pub(crate) part: BatchPart<T>,
    /// The `SeqNo` from which this part originated; we track this value as
    /// to ensure the `SeqNo` isn't garbage collected while a
    /// read still depends on it.
    pub(crate) leased_seqno: SeqNo,
    /// The lease that prevents this part from being GCed. Code should ensure that this lease
    /// lives as long as the part is needed.
    pub(crate) lease: Option<Lease>,
    pub(crate) filter_pushdown_audit: bool,
}

impl<T> LeasedBatchPart<T>
where
    T: Timestamp + Codec64,
{
    /// Takes `self` into a [`SerdeLeasedBatchPart`], which allows `self` to be
    /// exchanged (potentially across the network).
    ///
    /// !!!WARNING!!!
    ///
    /// This method also returns the [Lease] associated with the given part, since
    /// that can't travel across process boundaries. The caller is responsible for
    /// ensuring that the lease is held for as long as the batch part may be in use:
    /// dropping it too early may cause a fetch to fail.
    pub(crate) fn into_exchangeable_part(mut self) -> (SerdeLeasedBatchPart, Option<Lease>) {
        let (proto, _metrics) = self.into_proto();
        // If `x` has a lease, we've effectively transferred it to `r`.
        let lease = self.lease.take();
        let part = SerdeLeasedBatchPart {
            encoded_size_bytes: self.part.encoded_size_bytes(),
            proto: LazyProto::from(&proto),
        };
        (part, lease)
    }

    /// The encoded size of this part in bytes
    pub fn encoded_size_bytes(&self) -> usize {
        self.part.encoded_size_bytes()
    }

    /// The filter has indicated we don't need this part, we can verify the
    /// ongoing end-to-end correctness of corner cases via "audit". This means
    /// we fetch the part like normal and if the MFP keeps anything from it,
    /// then something has gone horribly wrong.
    pub fn request_filter_pushdown_audit(&mut self) {
        self.filter_pushdown_audit = true;
    }

    /// Returns the pushdown stats for this part.
    pub fn stats(&self) -> Option<PartStats> {
        self.part.stats().map(|x| x.decode())
    }

    /// Apply any relevant projection pushdown optimizations.
    ///
    /// NB: Until we implement full projection pushdown, this doesn't guarantee
    /// any projection.
    pub fn maybe_optimize(&mut self, cfg: &ConfigSet, project: &ProjectionPushdown) {
        let as_of = match &self.filter {
            FetchBatchFilter::Snapshot { as_of } => as_of,
            FetchBatchFilter::Listen { .. } | FetchBatchFilter::Compaction { .. } => return,
        };
        let faked_part = project.try_optimize_ignored_data_fetch(
            cfg,
            &self.metrics,
            as_of,
            &self.desc,
            &self.part,
        );
        if let Some(faked_part) = faked_part {
            self.part = faked_part;
        }
    }
}

impl<T> Drop for LeasedBatchPart<T> {
    /// For details, see [`LeasedBatchPart`].
    fn drop(&mut self) {
        self.metrics.lease.dropped_part.inc()
    }
}

/// A [Blob] object that has been fetched, but not at all decoded.
///
/// In contrast to [FetchedPart], this representation hasn't yet done parquet
/// decoding.
#[derive(Debug)]
pub struct FetchedBlob<K: Codec, V: Codec, T, D> {
    metrics: Arc<Metrics>,
    read_metrics: ReadMetrics,
    buf: FetchedBlobBuf<T>,
    registered_desc: Description<T>,
    migration: PartMigration<K, V>,
    filter: FetchBatchFilter<T>,
    filter_pushdown_audit: bool,
    structured_part_audit: PartDecodeFormat,
    fetch_permit: Option<Arc<MetricsPermits>>,
    _phantom: PhantomData<fn() -> D>,
}

#[derive(Debug, Clone)]
enum FetchedBlobBuf<T> {
    Hollow {
        buf: SegmentedBytes,
        part: HollowBatchPart<T>,
    },
    Inline {
        desc: Description<T>,
        updates: LazyInlineBatchPart,
        ts_rewrite: Option<Antichain<T>>,
    },
}

impl<K: Codec, V: Codec, T: Clone, D> Clone for FetchedBlob<K, V, T, D> {
    fn clone(&self) -> Self {
        Self {
            metrics: Arc::clone(&self.metrics),
            read_metrics: self.read_metrics.clone(),
            buf: self.buf.clone(),
            registered_desc: self.registered_desc.clone(),
            migration: self.migration.clone(),
            filter: self.filter.clone(),
            filter_pushdown_audit: self.filter_pushdown_audit.clone(),
            fetch_permit: self.fetch_permit.clone(),
            structured_part_audit: self.structured_part_audit.clone(),
            _phantom: self._phantom.clone(),
        }
    }
}

/// [FetchedPart] but with an accompanying permit from the fetch mem/disk
/// semaphore.
pub struct ShardSourcePart<K: Codec, V: Codec, T, D> {
    /// The underlying [FetchedPart].
    pub part: FetchedPart<K, V, T, D>,
    fetch_permit: Option<Arc<MetricsPermits>>,
}

impl<K: Codec, V: Codec, T: Clone, D> Clone for ShardSourcePart<K, V, T, D> {
    fn clone(&self) -> Self {
        Self {
            part: self.part.clone(),
            fetch_permit: self.fetch_permit.clone(),
        }
    }
}

impl<K, V, T: Debug, D: Debug> Debug for ShardSourcePart<K, V, T, D>
where
    K: Codec + Debug,
    <K as Codec>::Storage: Debug,
    V: Codec + Debug,
    <V as Codec>::Storage: Debug,
{
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        let ShardSourcePart { part, fetch_permit } = self;
        f.debug_struct("ShardSourcePart")
            .field("part", part)
            .field("fetch_permit", fetch_permit)
            .finish()
    }
}

impl<K: Codec, V: Codec, T: Timestamp + Lattice + Codec64, D> FetchedBlob<K, V, T, D> {
    /// Partially decodes this blob into a [FetchedPart].
    pub fn parse(&self) -> ShardSourcePart<K, V, T, D> {
        let (part, stats) = match &self.buf {
            FetchedBlobBuf::Hollow { buf, part } => {
                let parsed = decode_batch_part_blob(
                    &self.metrics,
                    &self.read_metrics,
                    self.registered_desc.clone(),
                    part,
                    buf,
                );
                (parsed, part.stats.as_ref())
            }
            FetchedBlobBuf::Inline {
                desc,
                updates,
                ts_rewrite,
            } => {
                let parsed = EncodedPart::from_inline(
                    &self.metrics,
                    self.read_metrics.clone(),
                    desc.clone(),
                    updates,
                    ts_rewrite.as_ref(),
                );
                (parsed, None)
            }
        };
        let part = FetchedPart::new(
            Arc::clone(&self.metrics),
            part,
            self.migration.clone(),
            self.filter.clone(),
            self.filter_pushdown_audit,
            self.structured_part_audit,
            stats,
        );
        ShardSourcePart {
            part,
            fetch_permit: self.fetch_permit.clone(),
        }
    }

    /// Decodes and returns the pushdown stats for this part, if known.
    pub fn stats(&self) -> Option<PartStats> {
        match &self.buf {
            FetchedBlobBuf::Hollow { part, .. } => part.stats.as_ref().map(|x| x.decode()),
            FetchedBlobBuf::Inline { .. } => None,
        }
    }
}

/// A [Blob] object that has been fetched, but not yet fully decoded.
///
/// In contrast to [FetchedBlob], this representation has already done parquet
/// decoding.
#[derive(Debug)]
pub struct FetchedPart<K: Codec, V: Codec, T, D> {
    metrics: Arc<Metrics>,
    ts_filter: FetchBatchFilter<T>,
    part: EncodedPart<T>,
    // If migration is Either, then the columnar one will have already been
    // applied here.
    structured_part: (
        Option<Arc<<K::Schema as Schema2<K>>::Decoder>>,
        Option<Arc<<V::Schema as Schema2<V>>::Decoder>>,
    ),
    part_decode_format: PartDecodeFormat,
    migration: PartMigration<K, V>,
    filter_pushdown_audit: Option<LazyPartStats>,
    part_cursor: Cursor,
    key_storage: Option<K::Storage>,
    val_storage: Option<V::Storage>,

    _phantom: PhantomData<fn() -> D>,
}

impl<K: Codec, V: Codec, T: Clone, D> Clone for FetchedPart<K, V, T, D> {
    fn clone(&self) -> Self {
        Self {
            metrics: Arc::clone(&self.metrics),
            ts_filter: self.ts_filter.clone(),
            part: self.part.clone(),
            structured_part: self.structured_part.clone(),
            part_decode_format: self.part_decode_format,
            migration: self.migration.clone(),
            filter_pushdown_audit: self.filter_pushdown_audit.clone(),
            part_cursor: self.part_cursor.clone(),
            key_storage: None,
            val_storage: None,
            _phantom: self._phantom.clone(),
        }
    }
}

impl<K: Codec, V: Codec, T: Timestamp + Lattice + Codec64, D> FetchedPart<K, V, T, D> {
    fn new(
        metrics: Arc<Metrics>,
        part: EncodedPart<T>,
        migration: PartMigration<K, V>,
        ts_filter: FetchBatchFilter<T>,
        filter_pushdown_audit: bool,
        part_decode_format: PartDecodeFormat,
        stats: Option<&LazyPartStats>,
    ) -> Self {
        let part_len = u64::cast_from(part.part.updates.records().len());
        match &migration {
            PartMigration::SameSchema { .. } => metrics.schema.migration_count_same.inc(),
            PartMigration::Codec { .. } => {
                metrics.schema.migration_count_codec.inc();
                metrics.schema.migration_len_legacy_codec.inc_by(part_len);
            }
            PartMigration::Either { .. } => {
                metrics.schema.migration_count_either.inc();
                match part_decode_format {
                    PartDecodeFormat::Row {
                        validate_structured: false,
                    } => metrics.schema.migration_len_either_codec.inc_by(part_len),
                    PartDecodeFormat::Row {
                        validate_structured: true,
                    } => {
                        metrics.schema.migration_len_either_codec.inc_by(part_len);
                        metrics.schema.migration_len_either_arrow.inc_by(part_len);
                    }
                    PartDecodeFormat::Arrow => {
                        metrics.schema.migration_len_either_arrow.inc_by(part_len)
                    }
                }
            }
        }

        let filter_pushdown_audit = if filter_pushdown_audit {
            stats.cloned()
        } else {
            None
        };

        // TODO(parkmycar): We should probably refactor this since these columns are duplicated
        // (via a smart pointer) in EncodedPart.
        //
        // For structured columnar data we need to downcast from `dyn Array`s to concrete types.
        // Downcasting is relatively expensive so we want to do this once, which is why we do it
        // when creating a FetchedPart.
        let should_downcast = match part_decode_format {
            PartDecodeFormat::Row {
                validate_structured,
            } => validate_structured,
            PartDecodeFormat::Arrow => true,
        };
        let structured_part = match (&part.part.updates, should_downcast) {
            // Only downcast and create decoders if we have structured data AND
            // (an audit of the data is requested OR we'd like to decode
            // directly from the structured data).
            (BlobTraceUpdates::Both(_codec, structured), true) => {
                fn decode<C: Codec>(
                    name: &str,
                    schema: &C::Schema,
                    array: &Arc<dyn Array>,
                ) -> Option<Arc<<C::Schema as Schema2<C>>::Decoder>> {
                    match Schema2::decoder_any(schema, array) {
                        Ok(x) => Some(Arc::new(x)),
                        Err(err) => {
                            tracing::error!(?err, "failed to create {} decoder", name);
                            None
                        }
                    }
                }
                match &migration {
                    PartMigration::SameSchema { both } => (
                        decode::<K>("key", &*both.key, &structured.key),
                        decode::<V>("val", &*both.val, &structured.val),
                    ),
                    PartMigration::Codec { .. } => (None, None),
                    PartMigration::Either {
                        _write,
                        read,
                        key_migration,
                        val_migration,
                    } => {
                        let start = Instant::now();
                        let key = key_migration.migrate(Arc::clone(&structured.key));
                        let val = val_migration.migrate(Arc::clone(&structured.val));
                        metrics
                            .schema
                            .migration_migrate_seconds
                            .inc_by(start.elapsed().as_secs_f64());

                        (
                            decode::<K>("key", &*read.key, &key),
                            decode::<V>("val", &*read.val, &val),
                        )
                    }
                }
            }
            _ => (None, None),
        };

        FetchedPart {
            metrics,
            ts_filter,
            part,
            structured_part,
            part_decode_format,
            migration,
            filter_pushdown_audit,
            part_cursor: Cursor::default(),
            key_storage: None,
            val_storage: None,
            _phantom: PhantomData,
        }
    }

    /// Returns Some if this part was only fetched as part of a filter pushdown
    /// audit. See [LeasedBatchPart::request_filter_pushdown_audit].
    ///
    /// If set, the value in the Option is for debugging and should be included
    /// in any error messages.
    pub fn is_filter_pushdown_audit(&self) -> Option<impl std::fmt::Debug> {
        self.filter_pushdown_audit.clone()
    }
}

/// A [Blob] object that has been fetched, but has no associated decoding
/// logic.
#[derive(Debug, Clone)]
pub(crate) struct EncodedPart<T> {
    metrics: ReadMetrics,
    registered_desc: Description<T>,
    part: Arc<BlobTraceBatchPart<T>>,
    needs_truncation: bool,
    ts_rewrite: Option<Antichain<T>>,
}

impl<K, V, T, D> FetchedPart<K, V, T, D>
where
    K: Debug + Codec,
    V: Debug + Codec,
    T: Timestamp + Lattice + Codec64,
    D: Semigroup + Codec64 + Send + Sync,
{
    /// [Self::next] but optionally providing a `K` and `V` for alloc reuse.
    ///
    /// When `result_override` is specified, return it instead of decoding data.
    /// This is used when we know the decoded result will be ignored.
    pub fn next_with_storage(
        &mut self,
        key: &mut Option<K>,
        val: &mut Option<V>,
        result_override: Option<(K, V)>,
    ) -> Option<((Result<K, String>, Result<V, String>), T, D)> {
        while let Some(((k, v, mut t, d), idx)) = self.part_cursor.pop(&self.part) {
            if !self.ts_filter.filter_ts(&mut t) {
                continue;
            }

            let mut d = D::decode(d);

            // If `filter_ts` advances our timestamp, we may end up with the same K, V, T in successive
            // records. If so, opportunistically consolidate those out.
            while let Some((k_next, v_next, mut t_next, d_next)) = self.part_cursor.peek(&self.part)
            {
                if (k, v) != (k_next, v_next) {
                    break;
                }

                if !self.ts_filter.filter_ts(&mut t_next) {
                    break;
                }
                if t != t_next {
                    break;
                }

                // All equal... consolidate!
                self.part_cursor.idx += 1;
                d.plus_equals(&D::decode(d_next));
            }

            // If multiple updates consolidate out entirely, drop the record.
            if d.is_zero() {
                continue;
            }

            if let Some((key, val)) = result_override {
                return Some(((Ok(key), Ok(val)), t, d));
            }

            // TODO: Putting this here relies on the Codec data still being
            // populated (i.e. for the consolidate optimization above).
            // Eventually we'll have to rewrite this path to work entirely
            // without Codec data, but in the meantime, putting in here allows
            // us to see the performance impact of decoding from arrow instead
            // of Codec.
            //
            // Plus, it'll likely be easier to port all the logic here to work
            // solely on arrow data once we finish migrating things like the
            // ConsolidatingIter.
            if let ((Some(keys), Some(vals)), PartDecodeFormat::Arrow) =
                (&self.structured_part, self.part_decode_format)
            {
                let (k, v) = self.decode_structured(idx, keys, vals, key, val);
                return Some(((k, v), t, d));
            }

            let (k, v) = Self::decode_codec(
                &self.metrics,
                self.migration.codec_read(),
                k,
                v,
                key,
                val,
                &mut self.key_storage,
                &mut self.val_storage,
            );
            // Note: We only provide structured columns, if they were originally written, and a
            // dyncfg was specified to run validation.
            if let (Some(keys), Some(vals)) = &self.structured_part {
                let (k_s, v_s) = self.decode_structured(idx, keys, vals, &mut None, &mut None);

                // Purposefully do not trace to prevent blowing up Sentry.
                let is_valid = self
                    .metrics
                    .columnar
                    .arrow()
                    .key()
                    .report_valid(|| k_s == k);
                if !is_valid {
                    soft_panic_no_log!("structured key did not match, {k_s:?} != {k:?}");
                }
                // Purposefully do not trace to prevent blowing up Sentry.
                let is_valid = self
                    .metrics
                    .columnar
                    .arrow()
                    .val()
                    .report_valid(|| v_s == v);
                if !is_valid {
                    soft_panic_no_log!("structured val did not match, {v_s:?} != {v:?}");
                }
            }

            return Some(((k, v), t, d));
        }
        None
    }

    fn decode_codec(
        metrics: &Metrics,
        read_schemas: &Schemas<K, V>,
        key_buf: &[u8],
        val_buf: &[u8],
        key: &mut Option<K>,
        val: &mut Option<V>,
        key_storage: &mut Option<K::Storage>,
        val_storage: &mut Option<V::Storage>,
    ) -> (Result<K, String>, Result<V, String>) {
        let k = metrics.codecs.key.decode(|| match key.take() {
            Some(mut key) => {
                match K::decode_from(&mut key, key_buf, key_storage, &read_schemas.key) {
                    Ok(()) => Ok(key),
                    Err(err) => Err(err),
                }
            }
            None => K::decode(key_buf, &read_schemas.key),
        });
        let v = metrics.codecs.val.decode(|| match val.take() {
            Some(mut val) => {
                match V::decode_from(&mut val, val_buf, val_storage, &read_schemas.val) {
                    Ok(()) => Ok(val),
                    Err(err) => Err(err),
                }
            }
            None => V::decode(val_buf, &read_schemas.val),
        });
        (k, v)
    }

    fn decode_structured(
        &self,
        idx: usize,
        keys: &<K::Schema as Schema2<K>>::Decoder,
        vals: &<V::Schema as Schema2<V>>::Decoder,
        key: &mut Option<K>,
        val: &mut Option<V>,
    ) -> (Result<K, String>, Result<V, String>) {
        let key = self.metrics.columnar.arrow().key().measure_decoding(|| {
            let mut key = key.take().unwrap_or_default();
            keys.decode(idx, &mut key);
            key
        });
        let val = self.metrics.columnar.arrow().val().measure_decoding(|| {
            let mut val = val.take().unwrap_or_default();
            vals.decode(idx, &mut val);
            val
        });
        (Ok(key), Ok(val))
    }
}

impl<K, V, T, D> Iterator for FetchedPart<K, V, T, D>
where
    K: Debug + Codec,
    V: Debug + Codec,
    T: Timestamp + Lattice + Codec64,
    D: Semigroup + Codec64 + Send + Sync,
{
    type Item = ((Result<K, String>, Result<V, String>), T, D);

    fn next(&mut self) -> Option<Self::Item> {
        self.next_with_storage(&mut None, &mut None, None)
    }

    fn size_hint(&self) -> (usize, Option<usize>) {
        // We don't know in advance how restrictive the filter will be.
        let max_len = self.part.part.updates.records().len();
        (0, Some(max_len))
    }
}

impl<T> EncodedPart<T>
where
    T: Timestamp + Lattice + Codec64,
{
    pub async fn fetch(
        shard_id: &ShardId,
        blob: &dyn Blob,
        metrics: &Metrics,
        shard_metrics: &ShardMetrics,
        read_metrics: &ReadMetrics,
        registered_desc: &Description<T>,
        part: &BatchPart<T>,
    ) -> Result<Self, BlobKey> {
        match part {
            BatchPart::Hollow(x) => {
                fetch_batch_part(
                    shard_id,
                    blob,
                    metrics,
                    shard_metrics,
                    read_metrics,
                    registered_desc,
                    x,
                )
                .await
            }
            BatchPart::Inline {
                updates,
                ts_rewrite,
                ..
            } => Ok(EncodedPart::from_inline(
                metrics,
                read_metrics.clone(),
                registered_desc.clone(),
                updates,
                ts_rewrite.as_ref(),
            )),
        }
    }

    pub(crate) fn from_inline(
        metrics: &Metrics,
        read_metrics: ReadMetrics,
        desc: Description<T>,
        x: &LazyInlineBatchPart,
        ts_rewrite: Option<&Antichain<T>>,
    ) -> Self {
        let parsed = x.decode(&metrics.columnar).expect("valid inline part");
        Self::new(read_metrics, desc, "inline", ts_rewrite, parsed)
    }

    pub(crate) fn from_hollow(
        metrics: ReadMetrics,
        registered_desc: Description<T>,
        part: &HollowBatchPart<T>,
        parsed: BlobTraceBatchPart<T>,
    ) -> Self {
        Self::new(
            metrics,
            registered_desc,
            &part.key.0,
            part.ts_rewrite.as_ref(),
            parsed,
        )
    }

    pub(crate) fn new(
        metrics: ReadMetrics,
        registered_desc: Description<T>,
        printable_name: &str,
        ts_rewrite: Option<&Antichain<T>>,
        parsed: BlobTraceBatchPart<T>,
    ) -> Self {
        // There are two types of batches in persist:
        // - Batches written by a persist user (either directly or indirectly
        //   via BatchBuilder). These always have a since of the minimum
        //   timestamp and may be registered in persist state with a tighter set
        //   of bounds than are inline in the batch (truncation). To read one of
        //   these batches, all data physically in the batch but outside of the
        //   truncated bounds must be ignored. Not every user batch is
        //   truncated.
        // - Batches written by compaction. These always have an inline desc
        //   that exactly matches the one they are registered with. The since
        //   can be anything.
        let inline_desc = &parsed.desc;
        let needs_truncation = inline_desc.lower() != registered_desc.lower()
            || inline_desc.upper() != registered_desc.upper();
        if needs_truncation {
            assert!(
                PartialOrder::less_equal(inline_desc.lower(), registered_desc.lower()),
                "key={} inline={:?} registered={:?}",
                printable_name,
                inline_desc,
                registered_desc
            );
            if ts_rewrite.is_none() {
                // The ts rewrite feature allows us to advance the registered
                // upper of a batch that's already been staged (the inline
                // upper), so if it's been used, then there's no useful
                // invariant that we can assert here.
                assert!(
                    PartialOrder::less_equal(registered_desc.upper(), inline_desc.upper()),
                    "key={} inline={:?} registered={:?}",
                    printable_name,
                    inline_desc,
                    registered_desc
                );
            }
            // As mentioned above, batches that needs truncation will always have a
            // since of the minimum timestamp. Technically we could truncate any
            // batch where the since is less_than the output_desc's lower, but we're
            // strict here so we don't get any surprises.
            assert_eq!(
                inline_desc.since(),
                &Antichain::from_elem(T::minimum()),
                "key={} inline={:?} registered={:?}",
                printable_name,
                inline_desc,
                registered_desc
            );
        } else {
            assert_eq!(
                inline_desc, &registered_desc,
                "key={} inline={:?} registered={:?}",
                printable_name, inline_desc, registered_desc
            );
        }

        EncodedPart {
            metrics,
            registered_desc,
            part: Arc::new(parsed),
            needs_truncation,
            ts_rewrite: ts_rewrite.cloned(),
        }
    }

    pub(crate) fn maybe_unconsolidated(&self) -> bool {
        // At time of writing, only user parts may be unconsolidated, and they are always
        // written with a since of [T::minimum()].
        self.part.desc.since().borrow() == AntichainRef::new(&[T::minimum()])
    }

    /// Returns the updates with all truncation / timestamp rewriting applied.
    pub(crate) fn normalize(&self, metrics: &ColumnarMetrics) -> BlobTraceUpdates {
        let updates = self.part.updates.clone();
        if !self.needs_truncation && self.ts_rewrite.is_none() {
            return updates;
        }

        let (records, ext) = match updates {
            BlobTraceUpdates::Row(r) => (r, None),
            BlobTraceUpdates::Both(r, e) => (r, Some(e)),
        };

        let records = match self.ts_rewrite.as_ref() {
            Some(rewrite) => {
                let timestamps = records.timestamps().clone();
                let rewrite = |i: i64| {
                    let mut t = T::decode(i.to_le_bytes());
                    t.advance_by(rewrite.borrow());
                    i64::from_le_bytes(T::encode(&t))
                };
                let timestamps = arrow::compute::unary_mut(timestamps, rewrite)
                    .unwrap_or_else(|i| arrow::compute::unary(&i, rewrite));
                ColumnarRecords::new(
                    records.keys().clone(),
                    records.vals().clone(),
                    realloc_array(&timestamps, metrics),
                    records.diffs().clone(),
                )
            }
            None => records,
        };
        let (records, ext) = if self.needs_truncation {
            let filter = BooleanArray::from_unary(records.timestamps(), |i| {
                let t = T::decode(i.to_le_bytes());
                let truncate_t = {
                    !self.registered_desc.lower().less_equal(&t)
                        || self.registered_desc.upper().less_equal(&t)
                };
                !truncate_t
            });
            if filter.false_count() == 0 {
                // If we're not filtering anything in practice, skip filtering and reallocating.
                (records, ext)
            } else {
                let filter = FilterBuilder::new(&filter).optimize().build();
                let do_filter = |array: &dyn Array| filter.filter(array).expect("valid filter len");
                let keys = realloc_array(do_filter(records.keys()).as_binary(), metrics);
                let values = realloc_array(do_filter(records.vals()).as_binary(), metrics);
                let timestamps =
                    realloc_array(do_filter(records.timestamps()).as_primitive(), metrics);
                let diffs = realloc_array(do_filter(records.diffs()).as_primitive(), metrics);
                let records = ColumnarRecords::new(keys, values, timestamps, diffs);
                let ext = ext.map(|ext| ColumnarRecordsStructuredExt {
                    key: realloc_any(do_filter(&ext.key), metrics),
                    val: realloc_any(do_filter(&ext.val), metrics),
                });
                (records, ext)
            }
        } else {
            (records, ext)
        };

        match ext {
            Some(ext) => BlobTraceUpdates::Both(records, ext),
            None => BlobTraceUpdates::Row(records),
        }
    }
}

/// A pointer into a particular encoded part, with methods for fetching an update and
/// scanning forward to the next. It is an error to use the same cursor for distinct
/// parts.
///
/// We avoid implementing copy to make it hard to accidentally duplicate a cursor. However,
/// clone is very cheap.
#[derive(Debug, Clone, Default)]
pub(crate) struct Cursor {
    idx: usize,
}

impl Cursor {
    /// Get the tuple at the specified pair of indices. If there is no such tuple,
    /// either because we are out of range or because this tuple has been filtered out,
    /// this returns `None`.
    pub fn get<'a, T: Timestamp + Lattice + Codec64>(
        &self,
        encoded: &'a EncodedPart<T>,
    ) -> Option<(&'a [u8], &'a [u8], T, [u8; 8])> {
        let part = encoded.part.updates.records();
        let ((k, v), t, d) = part.get(self.idx)?;

        let mut t = T::decode(t);
        // We assert on the write side that at most one of rewrite or
        // truncation is used, so it shouldn't matter which is run first.
        //
        // That said, my (Dan's) intuition here is that rewrite goes first,
        // though I don't particularly have a justification for it.
        if let Some(ts_rewrite) = encoded.ts_rewrite.as_ref() {
            t.advance_by(ts_rewrite.borrow());
            encoded.metrics.ts_rewrite.inc();
        }

        // This filtering is really subtle, see the comment above for
        // what's going on here.
        let truncated_t = encoded.needs_truncation && {
            !encoded.registered_desc.lower().less_equal(&t)
                || encoded.registered_desc.upper().less_equal(&t)
        };
        if truncated_t {
            return None;
        }
        Some((k, v, t, d))
    }

    /// A cursor points to a particular update in the backing part data.
    /// If the update it points to is not valid, advance it to the next valid update
    /// if there is one, and return the pointed-to data.
    pub fn peek<'a, T: Timestamp + Lattice + Codec64>(
        &mut self,
        part: &'a EncodedPart<T>,
    ) -> Option<(&'a [u8], &'a [u8], T, [u8; 8])> {
        while !self.is_exhausted(part) {
            let current = self.get(part);
            if current.is_some() {
                return current;
            }
            self.advance(part);
        }
        None
    }

    /// Similar to peek, but advance the cursor just past the end of the most recent update.
    /// Returns the update and the `(part_idx, idx)` that is was popped at.
    pub fn pop<'a, T: Timestamp + Lattice + Codec64>(
        &mut self,
        part: &'a EncodedPart<T>,
    ) -> Option<((&'a [u8], &'a [u8], T, [u8; 8]), usize)> {
        while !self.is_exhausted(part) {
            let current = self.get(part);
            let popped_idx = self.idx;
            self.advance(part);
            if current.is_some() {
                return current.map(|p| (p, popped_idx));
            }
        }
        None
    }

    /// Returns true if the cursor is past the end of the part data.
    pub fn is_exhausted<T: Timestamp + Codec64>(&self, part: &EncodedPart<T>) -> bool {
        self.idx >= part.part.updates.records().len()
    }

    /// Advance the cursor just past the end of the most recent update, if there is one.
    pub fn advance<T: Timestamp + Codec64>(&mut self, part: &EncodedPart<T>) {
        if !self.is_exhausted(part) {
            self.idx += 1;
        }
    }
}

/// This represents the serde encoding for [`LeasedBatchPart`]. We expose the struct
/// itself (unlike other encodable structs) to attempt to provide stricter drop
/// semantics on `LeasedBatchPart`, i.e. `SerdeLeasedBatchPart` is exchangeable
/// (including over the network), where `LeasedBatchPart` is not.
///
/// For more details see documentation and comments on:
/// - [`LeasedBatchPart`]
/// - `From<SerdeLeasedBatchPart>` for `LeasedBatchPart<T>`
#[derive(Debug, Serialize, Deserialize, Clone)]
pub struct SerdeLeasedBatchPart {
    // Duplicated with the one serialized in the proto for use in backpressure.
    encoded_size_bytes: usize,
    // We wrap this in a LazyProto because it guarantees that we use the proto
    // encoding for the serde impls.
    proto: LazyProto<ProtoLeasedBatchPart>,
}

impl SerdeLeasedBatchPart {
    /// Returns the encoded size of the given part.
    pub fn encoded_size_bytes(&self) -> usize {
        self.encoded_size_bytes
    }

    pub(crate) fn decode<T: Timestamp + Codec64>(
        &self,
        metrics: Arc<Metrics>,
    ) -> LeasedBatchPart<T> {
        let proto = self.proto.decode().expect("valid leased batch part");
        (proto, metrics)
            .into_rust()
            .expect("valid leased batch part")
    }
}

// TODO: The way we're smuggling the metrics through here is a bit odd. Perhaps
// we could refactor `LeasedBatchPart` into some proto-able struct plus the
// metrics for the Drop bit?
impl<T: Timestamp + Codec64> RustType<(ProtoLeasedBatchPart, Arc<Metrics>)> for LeasedBatchPart<T> {
    fn into_proto(&self) -> (ProtoLeasedBatchPart, Arc<Metrics>) {
        let proto = ProtoLeasedBatchPart {
            shard_id: self.shard_id.into_proto(),
            filter: Some(self.filter.into_proto()),
            desc: Some(self.desc.into_proto()),
            part: Some(self.part.into_proto()),
            lease: Some(ProtoLease {
                reader_id: self.reader_id.into_proto(),
                seqno: Some(self.leased_seqno.into_proto()),
            }),
            filter_pushdown_audit: self.filter_pushdown_audit,
        };
        (proto, Arc::clone(&self.metrics))
    }

    fn from_proto(proto: (ProtoLeasedBatchPart, Arc<Metrics>)) -> Result<Self, TryFromProtoError> {
        let (proto, metrics) = proto;
        let lease = proto
            .lease
            .ok_or_else(|| TryFromProtoError::missing_field("ProtoLeasedBatchPart::lease"))?;
        Ok(LeasedBatchPart {
            metrics,
            shard_id: proto.shard_id.into_rust()?,
            filter: proto
                .filter
                .into_rust_if_some("ProtoLeasedBatchPart::filter")?,
            desc: proto.desc.into_rust_if_some("ProtoLeasedBatchPart::desc")?,
            part: proto.part.into_rust_if_some("ProtoLeasedBatchPart::part")?,
            reader_id: lease.reader_id.into_rust()?,
            leased_seqno: lease.seqno.into_rust_if_some("ProtoLease::seqno")?,
            lease: None,
            filter_pushdown_audit: proto.filter_pushdown_audit,
        })
    }
}

/// Format we'll use when decoding a [`Part2`].
///
/// [`Part2`]: mz_persist_types::part::Part2
#[derive(Debug, Copy, Clone)]
pub enum PartDecodeFormat {
    /// Decode from opaque `Codec` data.
    Row {
        /// Will also decode the structured data, and validate it matches.
        validate_structured: bool,
    },
    /// Decode from arrow data
    Arrow,
}

impl PartDecodeFormat {
    /// Returns a default value for [`PartDecodeFormat`].
    pub const fn default() -> Self {
        // IMPORTANT: By default we will not decode or validate our structured format until it's
        // more stable.
        PartDecodeFormat::Row {
            validate_structured: false,
        }
    }

    /// Parses a [`PartDecodeFormat`] from the provided string, falling back to the default if the
    /// provided value is unrecognized.
    pub fn from_str(s: &str) -> Self {
        match s {
            "row" => PartDecodeFormat::Row {
                validate_structured: false,
            },
            "row_with_validate" => PartDecodeFormat::Row {
                validate_structured: true,
            },
            "arrow" => PartDecodeFormat::Arrow,
            x => {
                let default = PartDecodeFormat::default();
                soft_panic_or_log!("Invalid part decode format: '{x}', falling back to {default}");
                default
            }
        }
    }

    /// Returns a string representation of [`PartDecodeFormat`].
    pub const fn as_str(&self) -> &'static str {
        match self {
            PartDecodeFormat::Row {
                validate_structured: false,
            } => "row",
            PartDecodeFormat::Row {
                validate_structured: true,
            } => "row_with_validate",
            PartDecodeFormat::Arrow => "arrow",
        }
    }
}

impl fmt::Display for PartDecodeFormat {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.write_str(self.as_str())
    }
}

#[mz_ore::test]
fn client_exchange_data() {
    // The whole point of SerdeLeasedBatchPart is that it can be exchanged
    // between timely workers, including over the network. Enforce then that it
    // implements ExchangeData.
    fn is_exchange_data<T: timely::ExchangeData>() {}
    is_exchange_data::<SerdeLeasedBatchPart>();
    is_exchange_data::<SerdeLeasedBatchPart>();
}