aboutsummaryrefslogtreecommitdiff
path: root/src/rpc/layout.rs
blob: 157656628f4d232fc24e7f930e64654b5ba5d15d (plain) (blame)
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
use std::cmp::Ordering;
use std::collections::HashMap;
use std::collections::HashSet;

use hex::ToHex;
use itertools::Itertools;

use serde::{Deserialize, Serialize};

use garage_util::crdt::{AutoCrdt, Crdt, Lww, LwwMap};
use garage_util::data::*;
use garage_util::error::*;

use crate::graph_algo::*;

use crate::ring::*;

use std::convert::TryInto;

const NB_PARTITIONS: usize = 1usize << PARTITION_BITS;

// The Message type will be used to collect information on the algorithm.
type Message = Vec<String>;

/// The layout of the cluster, i.e. the list of roles
/// which are assigned to each cluster node
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct ClusterLayout {
	pub version: u64,

	pub replication_factor: usize,

	/// This attribute is only used to retain the previously computed partition size,
	/// to know to what extent does it change with the layout update.
	pub partition_size: u32,
	/// Parameters used to compute the assignation currently given by
	/// ring_assignation_data
	pub parameters: LayoutParameters,

	pub roles: LwwMap<Uuid, NodeRoleV>,

	/// node_id_vec: a vector of node IDs with a role assigned
	/// in the system (this includes gateway nodes).
	/// The order here is different than the vec stored by `roles`, because:
	/// 1. non-gateway nodes are first so that they have lower numbers holding
	///     in u8 (the number of non-gateway nodes is at most 256).
	/// 2. nodes that don't have a role are excluded (but they need to
	///    stay in the CRDT as tombstones)
	pub node_id_vec: Vec<Uuid>,
	/// the assignation of data partitions to node, the values
	/// are indices in node_id_vec
	#[serde(with = "serde_bytes")]
	pub ring_assignation_data: Vec<CompactNodeType>,

	/// Parameters to be used in the next partition assignation computation.
	pub staging_parameters: Lww<LayoutParameters>,
	/// Role changes which are staged for the next version of the layout
	pub staging_roles: LwwMap<Uuid, NodeRoleV>,
	pub staging_hash: Hash,
}

/// This struct is used to set the parameters to be used in the assignation computation
/// algorithm. It is stored as a Crdt.
#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
pub struct LayoutParameters {
	pub zone_redundancy: usize,
}

impl AutoCrdt for LayoutParameters {
	const WARN_IF_DIFFERENT: bool = true;
}

#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
pub struct NodeRoleV(pub Option<NodeRole>);

impl AutoCrdt for NodeRoleV {
	const WARN_IF_DIFFERENT: bool = true;
}

/// The user-assigned roles of cluster nodes
#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
pub struct NodeRole {
	/// Datacenter at which this entry belong. This information is used to
	/// perform a better geodistribution
	pub zone: String,
	/// The capacity of the node
	/// If this is set to None, the node does not participate in storing data for the system
	/// and is only active as an API gateway to other nodes
	// TODO : change the capacity to u64 and use byte unit input/output
	pub capacity: Option<u32>,
	/// A set of tags to recognize the node
	pub tags: Vec<String>,
}

impl NodeRole {
	pub fn capacity_string(&self) -> String {
		match self.capacity {
			Some(c) => format!("{}", c),
			None => "gateway".to_string(),
		}
	}

	pub fn tags_string(&self) -> String {
		let mut tags = String::new();
		if self.tags.is_empty() {
			return tags;
		}
		tags.push_str(&self.tags[0].clone());
		for t in 1..self.tags.len() {
			tags.push(',');
			tags.push_str(&self.tags[t].clone());
		}
		tags
	}
}

// Implementation of the ClusterLayout methods unrelated to the assignation algorithm.
impl ClusterLayout {
	pub fn new(replication_factor: usize) -> Self {
		// We set the default zone redundancy to be equal to the replication factor,
		// i.e. as strict as possible.
		let parameters = LayoutParameters {
			zone_redundancy: replication_factor,
		};
		let staging_parameters = Lww::<LayoutParameters>::new(parameters.clone());

		let empty_lwwmap = LwwMap::new();

		let mut ret = ClusterLayout {
			version: 0,
			replication_factor,
			partition_size: 0,
			roles: LwwMap::new(),
			node_id_vec: Vec::new(),
			ring_assignation_data: Vec::new(),
			parameters,
			staging_parameters,
			staging_roles: empty_lwwmap,
			staging_hash: [0u8; 32].into(),
		};
		ret.staging_hash = ret.calculate_staging_hash();
		ret
	}

	fn calculate_staging_hash(&self) -> Hash {
		let hashed_tuple = (&self.staging_roles, &self.staging_parameters);
		blake2sum(&rmp_to_vec_all_named(&hashed_tuple).unwrap()[..])
	}

	pub fn merge(&mut self, other: &ClusterLayout) -> bool {
		match other.version.cmp(&self.version) {
			Ordering::Greater => {
				*self = other.clone();
				true
			}
			Ordering::Equal => {
				self.staging_parameters.merge(&other.staging_parameters);
				self.staging_roles.merge(&other.staging_roles);

				let new_staging_hash = self.calculate_staging_hash();
				let changed = new_staging_hash != self.staging_hash;

				self.staging_hash = new_staging_hash;

				changed
			}
			Ordering::Less => false,
		}
	}

	pub fn apply_staged_changes(mut self, version: Option<u64>) -> Result<(Self, Message), Error> {
		match version {
			None => {
				let error = r#"
Please pass the new layout version number to ensure that you are writing the correct version of the cluster layout.
To know the correct value of the new layout version, invoke `garage layout show` and review the proposed changes.
				"#;
				return Err(Error::Message(error.into()));
			}
			Some(v) => {
				if v != self.version + 1 {
					return Err(Error::Message("Invalid new layout version".into()));
				}
			}
		}

		self.roles.merge(&self.staging_roles);
		self.roles.retain(|(_, _, v)| v.0.is_some());
		self.parameters = self.staging_parameters.get().clone();

		let msg = self.calculate_partition_assignation()?;

		self.staging_roles.clear();
		self.staging_hash = self.calculate_staging_hash();

		self.version += 1;

		Ok((self, msg))
	}

	pub fn revert_staged_changes(mut self, version: Option<u64>) -> Result<Self, Error> {
		match version {
			None => {
				let error = r#"
Please pass the new layout version number to ensure that you are writing the correct version of the cluster layout.
To know the correct value of the new layout version, invoke `garage layout show` and review the proposed changes.
				"#;
				return Err(Error::Message(error.into()));
			}
			Some(v) => {
				if v != self.version + 1 {
					return Err(Error::Message("Invalid new layout version".into()));
				}
			}
		}

		self.staging_roles.clear();
		self.staging_hash = self.calculate_staging_hash();
		self.staging_parameters.update(self.parameters.clone());

		self.version += 1;

		Ok(self)
	}

	/// Returns a list of IDs of nodes that currently have
	/// a role in the cluster
	pub fn node_ids(&self) -> &[Uuid] {
		&self.node_id_vec[..]
	}

	pub fn num_nodes(&self) -> usize {
		self.node_id_vec.len()
	}

	/// Returns the role of a node in the layout
	pub fn node_role(&self, node: &Uuid) -> Option<&NodeRole> {
		match self.roles.get(node) {
			Some(NodeRoleV(Some(v))) => Some(v),
			_ => None,
		}
	}

	/// Returns the uuids of the non_gateway nodes in self.node_id_vec.
	pub fn nongateway_nodes(&self) -> Vec<Uuid> {
		let mut result = Vec::<Uuid>::new();
		for uuid in self.node_id_vec.iter() {
			match self.node_role(uuid) {
				Some(role) if role.capacity != None => result.push(*uuid),
				_ => (),
			}
		}
		result
	}

	/// Given a node uuids, this function returns the label of its zone
	pub fn get_node_zone(&self, uuid: &Uuid) -> Result<String, Error> {
		match self.node_role(uuid) {
			Some(role) => Ok(role.zone.clone()),
			_ => Err(Error::Message(
				"The Uuid does not correspond to a node present in the cluster.".into(),
			)),
		}
	}

	/// Given a node uuids, this function returns its capacity or fails if it does not have any
	pub fn get_node_capacity(&self, uuid: &Uuid) -> Result<u32, Error> {
		match self.node_role(uuid) {
			Some(NodeRole {
				capacity: Some(cap),
				zone: _,
				tags: _,
			}) => Ok(*cap),
			_ => Err(Error::Message(
				"The Uuid does not correspond to a node present in the \
                    cluster or this node does not have a positive capacity."
					.into(),
			)),
		}
	}

	/// Returns the number of partitions associated to this node in the ring
	pub fn get_node_usage(&self, uuid: &Uuid) -> Result<usize, Error> {
		for (i, id) in self.node_id_vec.iter().enumerate() {
			if id == uuid {
				let mut count = 0;
				for nod in self.ring_assignation_data.iter() {
					if i as u8 == *nod {
						count += 1
					}
				}
				return Ok(count);
			}
		}
		Err(Error::Message(
			"The Uuid does not correspond to a node present in the \
                    cluster or this node does not have a positive capacity."
				.into(),
		))
	}

	/// Returns the sum of capacities of non gateway nodes in the cluster
	pub fn get_total_capacity(&self) -> Result<u32, Error> {
		let mut total_capacity = 0;
		for uuid in self.nongateway_nodes().iter() {
			total_capacity += self.get_node_capacity(uuid)?;
		}
		Ok(total_capacity)
	}

	/// Check a cluster layout for internal consistency
	/// (assignation, roles, parameters, partition size)
	/// returns true if consistent, false if error
	pub fn check(&self) -> bool {
		// Check that the hash of the staging data is correct
		let staging_hash = self.calculate_staging_hash();
		if staging_hash != self.staging_hash {
			return false;
		}

		// Check that node_id_vec contains the correct list of nodes
		let mut expected_nodes = self
			.roles
			.items()
			.iter()
			.filter(|(_, _, v)| v.0.is_some())
			.map(|(id, _, _)| *id)
			.collect::<Vec<_>>();
		expected_nodes.sort();
		let mut node_id_vec = self.node_id_vec.clone();
		node_id_vec.sort();
		if expected_nodes != node_id_vec {
			return false;
		}

		// Check that the assignation data has the correct length
		if self.ring_assignation_data.len() != (1 << PARTITION_BITS) * self.replication_factor {
			return false;
		}

		// Check that the assigned nodes are correct identifiers
		// of nodes that are assigned a role
		// and that role is not the role of a gateway nodes
		for x in self.ring_assignation_data.iter() {
			if *x as usize >= self.node_id_vec.len() {
				return false;
			}
			let node = self.node_id_vec[*x as usize];
			match self.roles.get(&node) {
				Some(NodeRoleV(Some(x))) if x.capacity.is_some() => (),
				_ => return false,
			}
		}

		// Check that every partition is associated to distinct nodes
		let rf = self.replication_factor;
		for p in 0..(1 << PARTITION_BITS) {
			let mut nodes_of_p = self.ring_assignation_data[rf * p..rf * (p + 1)].to_vec();
			nodes_of_p.sort();
			if nodes_of_p.iter().unique().count() != rf {
				return false;
			}
			// Check that every partition is spread over at least zone_redundancy zones.
			let mut zones_of_p = nodes_of_p
				.iter()
				.map(|n| {
					self.get_node_zone(&self.node_id_vec[*n as usize])
						.expect("Zone not found.")
				})
				.collect::<Vec<_>>();
			zones_of_p.sort();
			let redundancy = self.parameters.zone_redundancy;
			if zones_of_p.iter().unique().count() < redundancy {
				return false;
			}
		}

		// Check that the nodes capacities is consistent with the stored partitions
		let mut node_usage = vec![0; MAX_NODE_NUMBER];
		for n in self.ring_assignation_data.iter() {
			node_usage[*n as usize] += 1;
		}
		for (n, usage) in node_usage.iter().enumerate() {
			if *usage > 0 {
				let uuid = self.node_id_vec[n];
				if usage * self.partition_size > self.get_node_capacity(&uuid).unwrap() {
					return false;
				}
			}
		}

		// Check that the partition size stored is the one computed by the asignation
		// algorithm.
		let cl2 = self.clone();
		let (_, zone_to_id) = cl2.generate_nongateway_zone_ids().unwrap();
		match cl2.compute_optimal_partition_size(&zone_to_id) {
			Ok(s) if s != self.partition_size => return false,
			Err(_) => return false,
			_ => (),
		}

		true
	}
}

// Implementation of the ClusterLayout methods related to the assignation algorithm.
impl ClusterLayout {
	/// This function calculates a new partition-to-node assignation.
	/// The computed assignation respects the node replication factor
	/// and the zone redundancy parameter It maximizes the capacity of a
	/// partition (assuming all partitions have the same size).
	/// Among such optimal assignation, it minimizes the distance to
	/// the former assignation (if any) to minimize the amount of
	/// data to be moved.
	/// Staged role changes must be merged with nodes roles before calling this function,
	/// hence it must only be called from apply_staged_changes() and hence is not public.
	fn calculate_partition_assignation(&mut self) -> Result<Message, Error> {
		// We update the node ids, since the node role list might have changed with the
		// changes in the layout. We retrieve the old_assignation reframed with new ids
		let old_assignation_opt = self.update_node_id_vec()?;

		let mut msg = Message::new();
		msg.push("==== COMPUTATION OF A NEW PARTITION ASSIGNATION ====".into());
		msg.push("".into());
		msg.push(format!(
			"Partitions are \
        replicated {} times on at least {} distinct zones.",
			self.replication_factor, self.parameters.zone_redundancy
		));

		// We generate for once numerical ids for the zones of non gateway nodes,
		// to use them as indices in the flow graphs.
		let (id_to_zone, zone_to_id) = self.generate_nongateway_zone_ids()?;

		let nb_nongateway_nodes = self.nongateway_nodes().len();
		if nb_nongateway_nodes < self.replication_factor {
			return Err(Error::Message(format!(
				"The number of nodes with positive \
            capacity ({}) is smaller than the replication factor ({}).",
				nb_nongateway_nodes, self.replication_factor
			)));
		}
		if id_to_zone.len() < self.parameters.zone_redundancy {
			return Err(Error::Message(format!(
				"The number of zones with non-gateway \
            nodes ({}) is smaller than the redundancy parameter ({})",
				id_to_zone.len(),
				self.parameters.zone_redundancy
			)));
		}

		// We compute the optimal partition size
		// Capacities should be given in a unit so that partition size is at least 100.
		// In this case, integer rounding plays a marginal role in the percentages of
		// optimality.
		let partition_size = self.compute_optimal_partition_size(&zone_to_id)?;

		if old_assignation_opt != None {
			msg.push(format!(
				"Optimal size of a partition: {} (was {} in the previous layout).",
				partition_size, self.partition_size
			));
		} else {
			msg.push(format!(
				"Given the replication and redundancy constraints, the \
                optimal size of a partition is {}.",
				partition_size
			));
		}
		// We write the partition size.
		self.partition_size = partition_size;

		if partition_size < 100 {
			msg.push(
				"WARNING: The partition size is low (< 100), you might consider to \
            provide the nodes capacities in a smaller unit (e.g. Mb instead of Gb)."
					.into(),
			);
		}

		// We compute a first flow/assignation that is heuristically close to the previous
		// assignation
		let mut gflow = self.compute_candidate_assignation(&zone_to_id, &old_assignation_opt)?;
		if let Some(assoc) = &old_assignation_opt {
			// We minimize the distance to the previous assignation.
			self.minimize_rebalance_load(&mut gflow, &zone_to_id, assoc)?;
		}

		// We display statistics of the computation
		msg.extend(self.output_stat(&gflow, &old_assignation_opt, &zone_to_id, &id_to_zone)?);
		msg.push("".to_string());

		// We update the layout structure
		self.update_ring_from_flow(id_to_zone.len(), &gflow)?;

		if !self.check() {
			return Err(Error::Message(
				"Critical error: The computed layout happens to be incorrect".into(),
			));
		}

		Ok(msg)
	}

	/// The LwwMap of node roles might have changed. This function updates the node_id_vec
	/// and returns the assignation given by ring, with the new indices of the nodes, and
	/// None if the node is not present anymore.
	/// We work with the assumption that only this function and calculate_new_assignation
	/// do modify assignation_ring and node_id_vec.
	fn update_node_id_vec(&mut self) -> Result<Option<Vec<Vec<usize>>>, Error> {
		// (1) We compute the new node list
		// Non gateway nodes should be coded on 8bits, hence they must be first in the list
		// We build the new node ids
		let mut new_non_gateway_nodes: Vec<Uuid> = self
			.roles
			.items()
			.iter()
			.filter(|(_, _, v)| matches!(&v.0, Some(r) if r.capacity != None))
			.map(|(k, _, _)| *k)
			.collect();

		if new_non_gateway_nodes.len() > MAX_NODE_NUMBER {
			return Err(Error::Message(format!(
				"There are more than {} non-gateway nodes in the new \
                            layout. This is not allowed.",
				MAX_NODE_NUMBER
			)));
		}

		let mut new_gateway_nodes: Vec<Uuid> = self
			.roles
			.items()
			.iter()
			.filter(|(_, _, v)| matches!(v, NodeRoleV(Some(r)) if r.capacity == None))
			.map(|(k, _, _)| *k)
			.collect();

		let mut new_node_id_vec = Vec::<Uuid>::new();
		new_node_id_vec.append(&mut new_non_gateway_nodes);
		new_node_id_vec.append(&mut new_gateway_nodes);

		let old_node_id_vec = self.node_id_vec.clone();
		self.node_id_vec = new_node_id_vec.clone();

		// (2) We retrieve the old association
		// We rewrite the old association with the new indices. We only consider partition
		// to node assignations where the node is still in use.
		let mut old_assignation = vec![Vec::<usize>::new(); NB_PARTITIONS];

		if self.ring_assignation_data.is_empty() {
			// This is a new association
			return Ok(None);
		}
		if self.ring_assignation_data.len() != NB_PARTITIONS * self.replication_factor {
			return Err(Error::Message(
				"The old assignation does not have a size corresponding to \
                the old replication factor or the number of partitions."
					.into(),
			));
		}

		// We build a translation table between the uuid and new ids
		let mut uuid_to_new_id = HashMap::<Uuid, usize>::new();

		// We add the indices of only the new non-gateway nodes that can be used in the
		// association ring
		for (i, uuid) in new_node_id_vec.iter().enumerate() {
			uuid_to_new_id.insert(*uuid, i);
		}

		let rf = self.replication_factor;
		for (p, old_assign_p) in old_assignation.iter_mut().enumerate() {
			for old_id in &self.ring_assignation_data[p * rf..(p + 1) * rf] {
				let uuid = old_node_id_vec[*old_id as usize];
				if uuid_to_new_id.contains_key(&uuid) {
					old_assign_p.push(uuid_to_new_id[&uuid]);
				}
			}
		}

		// We write the ring
		self.ring_assignation_data = Vec::<CompactNodeType>::new();

		Ok(Some(old_assignation))
	}

	/// This function generates ids for the zone of the nodes appearing in
	/// self.node_id_vec.
	fn generate_nongateway_zone_ids(&self) -> Result<(Vec<String>, HashMap<String, usize>), Error> {
		let mut id_to_zone = Vec::<String>::new();
		let mut zone_to_id = HashMap::<String, usize>::new();

		for uuid in self.nongateway_nodes().iter() {
			if self.roles.get(uuid) == None {
				return Err(Error::Message(
					"The uuid was not found in the node roles (this should \
                    not happen, it might be a critical error)."
						.into(),
				));
			}
			if let Some(r) = self.node_role(uuid) {
				if !zone_to_id.contains_key(&r.zone) && r.capacity != None {
					zone_to_id.insert(r.zone.clone(), id_to_zone.len());
					id_to_zone.push(r.zone.clone());
				}
			}
		}
		Ok((id_to_zone, zone_to_id))
	}

	/// This function computes by dichotomy the largest realizable partition size, given
	/// the layout roles and parameters.
	fn compute_optimal_partition_size(
		&self,
		zone_to_id: &HashMap<String, usize>,
	) -> Result<u32, Error> {
		let empty_set = HashSet::<(usize, usize)>::new();
		let mut g = self.generate_flow_graph(1, zone_to_id, &empty_set)?;
		g.compute_maximal_flow()?;
		if g.get_flow_value()?
			< (NB_PARTITIONS * self.replication_factor)
				.try_into()
				.unwrap()
		{
			return Err(Error::Message(
				"The storage capacity of he cluster is to small. It is \
                       impossible to store partitions of size 1."
					.into(),
			));
		}

		let mut s_down = 1;
		let mut s_up = self.get_total_capacity()?;
		while s_down + 1 < s_up {
			g = self.generate_flow_graph((s_down + s_up) / 2, zone_to_id, &empty_set)?;
			g.compute_maximal_flow()?;
			if g.get_flow_value()?
				< (NB_PARTITIONS * self.replication_factor)
					.try_into()
					.unwrap()
			{
				s_up = (s_down + s_up) / 2;
			} else {
				s_down = (s_down + s_up) / 2;
			}
		}

		Ok(s_down)
	}

	fn generate_graph_vertices(nb_zones: usize, nb_nodes: usize) -> Vec<Vertex> {
		let mut vertices = vec![Vertex::Source, Vertex::Sink];
		for p in 0..NB_PARTITIONS {
			vertices.push(Vertex::Pup(p));
			vertices.push(Vertex::Pdown(p));
			for z in 0..nb_zones {
				vertices.push(Vertex::PZ(p, z));
			}
		}
		for n in 0..nb_nodes {
			vertices.push(Vertex::N(n));
		}
		vertices
	}

	/// Generates the graph to compute the maximal flow corresponding to the optimal
	/// partition assignation.
	/// exclude_assoc is the set of (partition, node) association that we are forbidden
	/// to use (hence we do not add the corresponding edge to the graph). This parameter
	/// is used to compute a first flow that uses only edges appearing in the previous
	/// assignation. This produces a solution that heuristically should be close to the
	/// previous one.
	fn generate_flow_graph(
		&self,
		partition_size: u32,
		zone_to_id: &HashMap<String, usize>,
		exclude_assoc: &HashSet<(usize, usize)>,
	) -> Result<Graph<FlowEdge>, Error> {
		let vertices =
			ClusterLayout::generate_graph_vertices(zone_to_id.len(), self.nongateway_nodes().len());
		let mut g = Graph::<FlowEdge>::new(&vertices);
		let nb_zones = zone_to_id.len();
		let redundancy = self.parameters.zone_redundancy;
		for p in 0..NB_PARTITIONS {
			g.add_edge(Vertex::Source, Vertex::Pup(p), redundancy as u32)?;
			g.add_edge(
				Vertex::Source,
				Vertex::Pdown(p),
				(self.replication_factor - redundancy) as u32,
			)?;
			for z in 0..nb_zones {
				g.add_edge(Vertex::Pup(p), Vertex::PZ(p, z), 1)?;
				g.add_edge(
					Vertex::Pdown(p),
					Vertex::PZ(p, z),
					self.replication_factor as u32,
				)?;
			}
		}
		for n in 0..self.nongateway_nodes().len() {
			let node_capacity = self.get_node_capacity(&self.node_id_vec[n])?;
			let node_zone = zone_to_id[&self.get_node_zone(&self.node_id_vec[n])?];
			g.add_edge(Vertex::N(n), Vertex::Sink, node_capacity / partition_size)?;
			for p in 0..NB_PARTITIONS {
				if !exclude_assoc.contains(&(p, n)) {
					g.add_edge(Vertex::PZ(p, node_zone), Vertex::N(n), 1)?;
				}
			}
		}
		Ok(g)
	}

	/// This function computes a first optimal assignation (in the form of a flow graph).
	fn compute_candidate_assignation(
		&self,
		zone_to_id: &HashMap<String, usize>,
		prev_assign_opt: &Option<Vec<Vec<usize>>>,
	) -> Result<Graph<FlowEdge>, Error> {
		// We list the (partition,node) associations that are not used in the
		// previous assignation
		let mut exclude_edge = HashSet::<(usize, usize)>::new();
		if let Some(prev_assign) = prev_assign_opt {
			let nb_nodes = self.nongateway_nodes().len();
			for (p, prev_assign_p) in prev_assign.iter().enumerate() {
				for n in 0..nb_nodes {
					exclude_edge.insert((p, n));
				}
				for n in prev_assign_p.iter() {
					exclude_edge.remove(&(p, *n));
				}
			}
		}

		// We compute the best flow using only the edges used in the previous assignation
		let mut g = self.generate_flow_graph(self.partition_size, zone_to_id, &exclude_edge)?;
		g.compute_maximal_flow()?;

		// We add the excluded edges and compute the maximal flow with the full graph.
		// The algorithm is such that it will start with the flow that we just computed
		// and find ameliorating paths from that.
		for (p, n) in exclude_edge.iter() {
			let node_zone = zone_to_id[&self.get_node_zone(&self.node_id_vec[*n])?];
			g.add_edge(Vertex::PZ(*p, node_zone), Vertex::N(*n), 1)?;
		}
		g.compute_maximal_flow()?;
		Ok(g)
	}

	/// This function updates the flow graph gflow to minimize the distance between
	/// its corresponding assignation and the previous one
	fn minimize_rebalance_load(
		&self,
		gflow: &mut Graph<FlowEdge>,
		zone_to_id: &HashMap<String, usize>,
		prev_assign: &[Vec<usize>],
	) -> Result<(), Error> {
		// We define a cost function on the edges (pairs of vertices) corresponding
		// to the distance between the two assignations.
		let mut cost = CostFunction::new();
		for (p, assoc_p) in prev_assign.iter().enumerate() {
			for n in assoc_p.iter() {
				let node_zone = zone_to_id[&self.get_node_zone(&self.node_id_vec[*n])?];
				cost.insert((Vertex::PZ(p, node_zone), Vertex::N(*n)), -1);
			}
		}

		// We compute the maximal length of a simple path in gflow. It is used in the
		// Bellman-Ford algorithm in optimize_flow_with_cost to set the number
		// of iterations.
		let nb_nodes = self.nongateway_nodes().len();
		let path_length = 4 * nb_nodes;
		gflow.optimize_flow_with_cost(&cost, path_length)?;

		Ok(())
	}

	/// This function updates the assignation ring from the flow graph.
	fn update_ring_from_flow(
		&mut self,
		nb_zones: usize,
		gflow: &Graph<FlowEdge>,
	) -> Result<(), Error> {
		self.ring_assignation_data = Vec::<CompactNodeType>::new();
		for p in 0..NB_PARTITIONS {
			for z in 0..nb_zones {
				let assoc_vertex = gflow.get_positive_flow_from(Vertex::PZ(p, z))?;
				for vertex in assoc_vertex.iter() {
					if let Vertex::N(n) = vertex {
						self.ring_assignation_data.push((*n).try_into().unwrap());
					}
				}
			}
		}

		if self.ring_assignation_data.len() != NB_PARTITIONS * self.replication_factor {
			return Err(Error::Message(
				"Critical Error : the association ring we produced does not \
                       have the right size."
					.into(),
			));
		}
		Ok(())
	}

	/// This function returns a message summing up the partition repartition of the new
	/// layout, and other statistics of the partition assignation computation.
	fn output_stat(
		&self,
		gflow: &Graph<FlowEdge>,
		prev_assign_opt: &Option<Vec<Vec<usize>>>,
		zone_to_id: &HashMap<String, usize>,
		id_to_zone: &[String],
	) -> Result<Message, Error> {
		let mut msg = Message::new();

		let used_cap = self.partition_size * NB_PARTITIONS as u32 * self.replication_factor as u32;
		let total_cap = self.get_total_capacity()?;
		let percent_cap = 100.0 * (used_cap as f32) / (total_cap as f32);
		msg.push("".into());
		msg.push(format!(
			"Usable capacity / Total cluster capacity: {} / {} ({:.1} %)",
			used_cap, total_cap, percent_cap
		));
		msg.push("".into());
		msg.push(
			"If the percentage is to low, it might be that the \
        replication/redundancy constraints force the use of nodes/zones with small \
        storage capacities. \
        You might want to rebalance the storage capacities or relax the constraints. \
        See the detailed statistics below and look for saturated nodes/zones."
				.into(),
		);
		msg.push(format!(
			"Recall that because of the replication factor, the actual available \
                         storage capacity is {} / {} = {}.",
			used_cap,
			self.replication_factor,
			used_cap / self.replication_factor as u32
		));

		// We define and fill in the following tables
		let storing_nodes = self.nongateway_nodes();
		let mut new_partitions = vec![0; storing_nodes.len()];
		let mut stored_partitions = vec![0; storing_nodes.len()];

		let mut new_partitions_zone = vec![0; id_to_zone.len()];
		let mut stored_partitions_zone = vec![0; id_to_zone.len()];

		for p in 0..NB_PARTITIONS {
			for z in 0..id_to_zone.len() {
				let pz_nodes = gflow.get_positive_flow_from(Vertex::PZ(p, z))?;
				if !pz_nodes.is_empty() {
					stored_partitions_zone[z] += 1;
					if let Some(prev_assign) = prev_assign_opt {
						let mut old_zones_of_p = Vec::<usize>::new();
						for n in prev_assign[p].iter() {
							old_zones_of_p
								.push(zone_to_id[&self.get_node_zone(&self.node_id_vec[*n])?]);
						}
						if !old_zones_of_p.contains(&z) {
							new_partitions_zone[z] += 1;
						}
					}
				}
				for vert in pz_nodes.iter() {
					if let Vertex::N(n) = *vert {
						stored_partitions[n] += 1;
						if let Some(prev_assign) = prev_assign_opt {
							if !prev_assign[p].contains(&n) {
								new_partitions[n] += 1;
							}
						}
					}
				}
			}
		}

		if *prev_assign_opt == None {
			new_partitions = stored_partitions.clone();
			new_partitions_zone = stored_partitions_zone.clone();
		}

		// We display the statistics

		msg.push("".into());
		if *prev_assign_opt != None {
			let total_new_partitions: usize = new_partitions.iter().sum();
			msg.push(format!(
				"A total of {} new copies of partitions need to be \
                             transferred.",
				total_new_partitions
			));
		}
		msg.push("".into());
		msg.push("==== DETAILED STATISTICS BY ZONES AND NODES ====".into());

		for z in 0..id_to_zone.len() {
			let mut nodes_of_z = Vec::<usize>::new();
			for n in 0..storing_nodes.len() {
				if self.get_node_zone(&self.node_id_vec[n])? == id_to_zone[z] {
					nodes_of_z.push(n);
				}
			}
			let replicated_partitions: usize =
				nodes_of_z.iter().map(|n| stored_partitions[*n]).sum();
			msg.push("".into());

			msg.push(format!(
				"Zone {}: {} distinct partitions stored ({} new, \
                {} partition copies) ",
				id_to_zone[z],
				stored_partitions_zone[z],
				new_partitions_zone[z],
				replicated_partitions
			));

			let available_cap_z: u32 = self.partition_size * replicated_partitions as u32;
			let mut total_cap_z = 0;
			for n in nodes_of_z.iter() {
				total_cap_z += self.get_node_capacity(&self.node_id_vec[*n])?;
			}
			let percent_cap_z = 100.0 * (available_cap_z as f32) / (total_cap_z as f32);
			msg.push(format!(
				"  Usable capacity / Total capacity: {}/{} ({:.1}%).",
				available_cap_z, total_cap_z, percent_cap_z
			));

			for n in nodes_of_z.iter() {
				let available_cap_n = stored_partitions[*n] as u32 * self.partition_size;
				let total_cap_n = self.get_node_capacity(&self.node_id_vec[*n])?;
				let tags_n = (self
					.node_role(&self.node_id_vec[*n])
					.ok_or("Node not found."))?
				.tags_string();
				msg.push(format!(
					"  Node {}: {} partitions ({} new) ; \
                                 usable/total capacity: {} / {} ({:.1}%) ; tags:{}",
					&self.node_id_vec[*n].to_vec()[0..2]
						.to_vec()
						.encode_hex::<String>(),
					stored_partitions[*n],
					new_partitions[*n],
					available_cap_n,
					total_cap_n,
					(available_cap_n as f32) / (total_cap_n as f32) * 100.0,
					tags_n
				));
			}
		}

		Ok(msg)
	}
}

// ====================================================================================

#[cfg(test)]
mod tests {
	use super::{Error, *};
	use std::cmp::min;

	// This function checks that the partition size S computed is at least better than the
	// one given by a very naive algorithm. To do so, we try to run the naive algorithm
	// assuming a partion size of S+1. If we succed, it means that the optimal assignation
	// was not optimal. The naive algorithm is the following :
	// - we compute the max number of partitions associated to every node, capped at the
	// partition number. It gives the number of tokens of every node.
	// - every zone has a number of tokens equal to the sum of the tokens of its nodes.
	// - we cycle over the partitions and associate zone tokens while respecting the
	// zone redundancy constraint.
	// NOTE: the naive algorithm is not optimal. Counter example:
	// take nb_partition = 3  ; replication_factor = 5; redundancy = 4;
	// number of tokens by zone : (A, 4), (B,1), (C,4), (D, 4), (E, 2)
	// With these parameters, the naive algo fails, whereas there is a solution:
	// (A,A,C,D,E) , (A,B,C,D,D) (A,C,C,D,E)
	fn check_against_naive(cl: &ClusterLayout) -> Result<bool, Error> {
		let over_size = cl.partition_size + 1;
		let mut zone_token = HashMap::<String, usize>::new();

		let (zones, zone_to_id) = cl.generate_nongateway_zone_ids()?;

		if zones.is_empty() {
			return Ok(false);
		}

		for z in zones.iter() {
			zone_token.insert(z.clone(), 0);
		}
		for uuid in cl.nongateway_nodes().iter() {
			let z = cl.get_node_zone(uuid)?;
			let c = cl.get_node_capacity(uuid)?;
			zone_token.insert(
				z.clone(),
				zone_token[&z] + min(NB_PARTITIONS, (c / over_size) as usize),
			);
		}

		// For every partition, we count the number of zone already associated and
		// the name of the last zone associated

		let mut id_zone_token = vec![0; zones.len()];
		for (z, t) in zone_token.iter() {
			id_zone_token[zone_to_id[z]] = *t;
		}

		let mut nb_token = vec![0; NB_PARTITIONS];
		let mut last_zone = vec![zones.len(); NB_PARTITIONS];

		let mut curr_zone = 0;

		let redundancy = cl.parameters.zone_redundancy;

		for replic in 0..cl.replication_factor {
			for p in 0..NB_PARTITIONS {
				while id_zone_token[curr_zone] == 0
					|| (last_zone[p] == curr_zone
						&& redundancy - nb_token[p] <= cl.replication_factor - replic)
				{
					curr_zone += 1;
					if curr_zone >= zones.len() {
						return Ok(true);
					}
				}
				id_zone_token[curr_zone] -= 1;
				if last_zone[p] != curr_zone {
					nb_token[p] += 1;
					last_zone[p] = curr_zone;
				}
			}
		}

		return Ok(false);
	}

	fn show_msg(msg: &Message) {
		for s in msg.iter() {
			println!("{}", s);
		}
	}

	fn update_layout(
		cl: &mut ClusterLayout,
		node_id_vec: &Vec<u8>,
		node_capacity_vec: &Vec<u32>,
		node_zone_vec: &Vec<String>,
		zone_redundancy: usize,
	) {
		for i in 0..node_id_vec.len() {
			if let Some(x) = FixedBytes32::try_from(&[i as u8; 32]) {
				cl.node_id_vec.push(x);
			}

			let update = cl.staging_roles.update_mutator(
				cl.node_id_vec[i],
				NodeRoleV(Some(NodeRole {
					zone: (node_zone_vec[i].to_string()),
					capacity: (Some(node_capacity_vec[i])),
					tags: (vec![]),
				})),
			);
			cl.staging_roles.merge(&update);
		}
		cl.staging_hash = cl.calculate_staging_hash();
		cl.staging_parameters
			.update(LayoutParameters { zone_redundancy });
	}

	#[test]
	fn test_assignation() {
		let mut node_id_vec = vec![1, 2, 3];
		let mut node_capacity_vec = vec![4000, 1000, 2000];
		let mut node_zone_vec = vec!["A", "B", "C"]
			.into_iter()
			.map(|x| x.to_string())
			.collect();

		let mut cl = ClusterLayout::new(3);
		update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec, 3);
		let v = cl.version;
		let (mut cl, msg) = cl.apply_staged_changes(Some(v + 1)).unwrap();
		show_msg(&msg);
		assert!(cl.check());
		assert!(matches!(check_against_naive(&cl), Ok(true)));

		node_id_vec = vec![1, 2, 3, 4, 5, 6, 7, 8, 9];
		node_capacity_vec = vec![4000, 1000, 1000, 3000, 1000, 1000, 2000, 10000, 2000];
		node_zone_vec = vec!["A", "B", "C", "C", "C", "B", "G", "H", "I"]
			.into_iter()
			.map(|x| x.to_string())
			.collect();
		update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec, 2);
		let v = cl.version;
		let (mut cl, msg) = cl.apply_staged_changes(Some(v + 1)).unwrap();
		show_msg(&msg);
		assert!(cl.check());
		assert!(matches!(check_against_naive(&cl), Ok(true)));

		node_capacity_vec = vec![4000, 1000, 2000, 7000, 1000, 1000, 2000, 10000, 2000];
		update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec, 3);
		let v = cl.version;
		let (mut cl, msg) = cl.apply_staged_changes(Some(v + 1)).unwrap();
		show_msg(&msg);
		assert!(cl.check());
		assert!(matches!(check_against_naive(&cl), Ok(true)));

		node_capacity_vec = vec![
			4000000, 4000000, 2000000, 7000000, 1000000, 9000000, 2000000, 10000, 2000000,
		];
		update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec, 1);
		let v = cl.version;
		let (cl, msg) = cl.apply_staged_changes(Some(v + 1)).unwrap();
		show_msg(&msg);
		assert!(cl.check());
		assert!(matches!(check_against_naive(&cl), Ok(true)));
	}
}