aboutsummaryrefslogblamecommitdiff
path: root/doc/talks/2024-02-03-fosdem/talk.tex
blob: 4c4dcdbc4697a5cf463e7fa84880dab60f2479af (plain) (tree)
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















































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































































                                                                                                                                                                                                                 
\nonstopmode
\documentclass[aspectratio=169]{beamer}
\usepackage[utf8]{inputenc}
% \usepackage[frenchb]{babel}
\usepackage{amsmath}
\usepackage{mathtools}
\usepackage{breqn}
\usepackage{multirow}
\usetheme{boxes}
\usepackage{graphicx}
\usepackage{import}
\usepackage{adjustbox}
%\useoutertheme[footline=authortitle,subsection=false]{miniframes}
%\useoutertheme[footline=authorinstitute,subsection=false]{miniframes}
\useoutertheme{infolines}
\setbeamertemplate{headline}{}

\beamertemplatenavigationsymbolsempty

\definecolor{TitleOrange}{RGB}{255,137,0}
\setbeamercolor{title}{fg=TitleOrange}
\setbeamercolor{frametitle}{fg=TitleOrange}

\definecolor{ListOrange}{RGB}{255,145,5}
\setbeamertemplate{itemize item}{\color{ListOrange}$\blacktriangleright$}

\definecolor{verygrey}{RGB}{70,70,70}
\setbeamercolor{normal text}{fg=verygrey}


\usepackage{tabu}
\usepackage{multicol}
\usepackage{vwcol}
\usepackage{stmaryrd}
\usepackage{graphicx}

\usepackage[normalem]{ulem}

\AtBeginSection[]{
  \begin{frame}
  \vfill
  \centering
  \begin{beamercolorbox}[sep=8pt,center,shadow=true,rounded=true]{title}
    \usebeamerfont{title}\insertsectionhead\par%
  \end{beamercolorbox}
  \vfill
  \end{frame}
}

\title{Garage, the low-tech storage platform for geo-distributed clusters}
\author{Alex Auvolat, Deuxfleurs}
\date{FOSDEM'24, 2024-02-03}

\begin{document}

\begin{frame}
	\centering
	\includegraphics[width=.3\linewidth]{../../sticker/Garage.png}
	\vspace{1em}

	{\large\bf Alex Auvolat, Deuxfleurs Association}
	\vspace{1em}

	\url{https://garagehq.deuxfleurs.fr/}

	Matrix channel: \texttt{\#garage:deuxfleurs.fr}
\end{frame}

\begin{frame}
	\frametitle{Who I am}
	\begin{columns}[t]
		\begin{column}{.2\textwidth}
			\centering
			\adjincludegraphics[width=.4\linewidth, valign=t]{../assets/alex.jpg}
		\end{column}
		\begin{column}{.6\textwidth}
			\textbf{Alex Auvolat}\\
			PhD; co-founder of Deuxfleurs
		\end{column}
		\begin{column}{.2\textwidth}
			~
		\end{column}
	\end{columns}
	\vspace{2em}

	\begin{columns}[t]
		\begin{column}{.2\textwidth}
			\centering
			\adjincludegraphics[width=.5\linewidth, valign=t]{../assets/deuxfleurs.pdf}
		\end{column}
		\begin{column}{.6\textwidth}
			\textbf{Deuxfleurs}\\
			A non-profit self-hosting collective,\\
			member of the CHATONS network
		\end{column}
		\begin{column}{.2\textwidth}
			\centering
			\adjincludegraphics[width=.7\linewidth, valign=t]{../assets/logo_chatons.png}
		\end{column}
	\end{columns}

\end{frame}

\begin{frame}
	\frametitle{Our objective at Deuxfleurs}
	
	\begin{center}
		\textbf{Promote self-hosting and small-scale hosting\\
			as an alternative to large cloud providers}
	\end{center}
	\vspace{2em}
	\visible<2->{
		Why is it hard?
	}
	\visible<3->{
		\vspace{2em}
		\begin{center}
			\textbf{\underline{Resilience}}\\
			{\footnotesize (we want good uptime/availability with low supervision)}
		\end{center}
	}
\end{frame}

\begin{frame}
	\frametitle{How to make a \underline{stable} system}

	Enterprise-grade systems typically employ:
	\vspace{1em}
	\begin{itemize}
		\item RAID
		\item Redundant power grid + UPS
		\item Redundant Internet connections
		\item Low-latency links
		\item ... 
	\end{itemize}
	\vspace{1em}
	$\to$ it's costly and only worth it at DC scale
\end{frame}

\begin{frame}
	\frametitle{How to make a \underline{resilient} system}

	\only<1,4-5>{
		Instead, we use:
		\vspace{1em}
		\begin{itemize}
			\item \textcolor<2->{gray}{Commodity hardware (e.g. old desktop PCs)}
				\vspace{.5em}
			\item<4-> \textcolor<5->{gray}{Commodity Internet (e.g. FTTB, FTTH) and power grid}
				\vspace{.5em}
			\item<5-> \textcolor<6->{gray}{\textbf{Geographical redundancy} (multi-site replication)}
		\end{itemize}
	}
	\only<2>{
		\begin{center}
			\includegraphics[width=.8\linewidth]{../assets/neptune.jpg}
		\end{center}
	}
	\only<3>{
		\begin{center}
			\includegraphics[width=.8\linewidth]{../assets/atuin.jpg}
		\end{center}
	}
	\only<6>{
		\begin{center}
			\includegraphics[width=.8\linewidth]{../assets/inframap_jdll2023.pdf}
		\end{center}
	}
\end{frame}

\begin{frame}
	\frametitle{Object storage: a crucial component}
		\begin{center}
			\includegraphics[height=6em]{../assets/Amazon-S3.jpg}
			\hspace{3em}
			\includegraphics[height=5em]{../assets/minio.png}
			\hspace{3em}
			\includegraphics[height=6em]{../../logo/garage_hires_crop.png}
		\end{center}
		\vspace{1em}
	S3: a de-facto standard, many compatible applications

	\vspace{1em}

	MinIO is self-hostable but not suited for geo-distributed deployments

	\vspace{1em}

	\textbf{Garage is a self-hosted drop-in replacement for the Amazon S3 object store}
\end{frame}

\begin{frame}
	\frametitle{CRDTs / weak consistency instead of consensus}
	Consensus can be implemented reasonably well in practice, so why avoid it?
	\vspace{1em}
	\begin{itemize}
		\item \textbf{Software complexity}
			\vspace{1em}
		\item \textbf{Performance issues:}
			\vspace{.5em}
			\begin{itemize}
				\item The leader is a \textbf{bottleneck} for all requests\\
					\vspace{.5em}
				\item \textbf{Sensitive to higher latency} between nodes
					\vspace{.5em}
				\item \textbf{Takes time to reconverge} when disrupted (e.g. node going down)
			\end{itemize}
	\end{itemize}
	\vspace{2em}
	$\to$ Garage uses only CRDTs internally (conflict-free replicated data types)
\end{frame}

\begin{frame}
	\frametitle{The data model of object storage}
	Object storage is basically a key-value store:
	\vspace{1em}

	{\footnotesize
		\begin{center}
		\begin{tabular}{|l|p{8cm}|}
			\hline
			\textbf{Key: file path + name} & \textbf{Value: file data + metadata} \\
			\hline
			\hline
			\texttt{index.html} &
				\texttt{Content-Type: text/html; charset=utf-8} \newline
				\texttt{Content-Length: 24929} \newline
				\texttt{<binary blob>} \\ 
			\hline
			\texttt{img/logo.svg} &
				\texttt{Content-Type: text/svg+xml} \newline
				\texttt{Content-Length: 13429} \newline
				\texttt{<binary blob>} \\ 
			\hline
			\texttt{download/index.html} &
				\texttt{Content-Type: text/html; charset=utf-8} \newline
				\texttt{Content-Length: 26563} \newline
				\texttt{<binary blob>} \\ 
			\hline
		\end{tabular}
		\end{center}
		}

	\vspace{1em}
	Simple interface, compatible with many existing applications

	\vspace{1em}
	Maps well to CRDT data types
\end{frame}

\begin{frame}
	\frametitle{Performance gains in practice}
	\begin{center}
		\includegraphics[width=.8\linewidth]{../assets/endpoint_latency_0.7_0.8_minio.png}
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Timeline}
	\begin{center}
		\includegraphics[width=.8\linewidth]{../assets/timeline-22-24.pdf}
	\end{center}
\end{frame}

% ====================== v0.7.0 ===============================

\begin{frame}
	\frametitle{April 2022 - Garage v0.7.0}
	Focus on \underline{observability and ecosystem integration}
	\vspace{2em}
	\begin{itemize}
		\item \textbf{Monitoring:} metrics and traces, using OpenTelemetry
			\vspace{1em}
		\item Alternative replication modes with 1 or 2 copies,\\
			modes with weaker consistency
			\vspace{1em}
		\item Kubernetes integration
			\vspace{1em}
		\item Admin API (v0.7.2)
			\vspace{1em}
		\item Experimental K2V API (v0.7.2)
	\end{itemize}
\end{frame}

\begin{frame}
	\frametitle{Metrics (Prometheus + Grafana)}
	\begin{center}
		\includegraphics[width=.9\linewidth]{../assets/grafana_dashboard.png}
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Traces (Jaeger)}
	\begin{center}
		\includegraphics[width=.8\linewidth]{../assets/jaeger_listobjects.png}
	\end{center}
\end{frame}

% ====================== v0.8.0 ===============================

\begin{frame}
	\frametitle{November 2022 - Garage v0.8.0}
	Focus on \underline{performance}
	\vspace{2em}
	\begin{itemize}
		\item \textbf{Alternative metadata DB engines} (LMDB, Sqlite)
			\vspace{1em}
		\item \textbf{Performance improvements:} block streaming, various optimizations...
			\vspace{1em}
		\item Bucket quotas (max size, max \#objects)
			\vspace{1em}
		\item Quality of life improvements, observability, etc.
	\end{itemize}
\end{frame}

\begin{frame}
	\frametitle{About metadata DB engines}
	Issues with Sled:
	\vspace{2em}
	\begin{itemize}
		\item Huge files on disk
			\vspace{.5em}
		\item Unpredictable performance, especially on HDD
			\vspace{.5em}
		\item API limitations
			\vspace{.5em}
		\item Not actively maintained
	\end{itemize}
	\vspace{2em}
	LMDB: very stable, good performance, reasonably small files on disk

	\vspace{1em}
	Sled will be removed in Garage v1.0
\end{frame}

\begin{frame}
	\frametitle{DB engine performance comparison}
	\begin{center}
		\includegraphics[width=.6\linewidth]{../assets/db_engine.png}
	\end{center}
	NB: Sqlite was slow due to synchronous journaling mode, now configurable
\end{frame}

\begin{frame}
	\frametitle{Block streaming}
	\begin{center}
		\only<1>{\includegraphics[width=.8\linewidth]{../assets/schema-streaming-1.png}}
		\only<2>{\includegraphics[width=.8\linewidth]{../assets/schema-streaming-2.png}}
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{TTFB benchmark}
	\begin{center}
		\includegraphics[width=.8\linewidth]{../assets/ttfb.png}
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Throughput benchmark}
	\begin{center}
		\includegraphics[width=.7\linewidth]{../assets/io-0.7-0.8-minio.png}
	\end{center}
\end{frame}

% ====================== v0.9.0 ===============================

\begin{frame}
	\frametitle{October 2023 - Garage v0.9.0}
	Focus on \underline{streamlining \& usability}
	\vspace{2em}
	\begin{itemize}
		\item Support multiple HDDs per node
			\vspace{1em}
		\item S3 compatibility: 
			\vspace{1em}
			\begin{itemize}
				\item support basic lifecycle configurations
					\vspace{.5em}
				\item allow for multipart upload part retries
			\end{itemize}
			\vspace{1em}
		\item LMDB by default, deprecation of Sled
			\vspace{1em}
		\item New layout computation algorithm
	\end{itemize}
\end{frame}


\begin{frame}
	\frametitle{Layout computation}
	\begin{center}
		\includegraphics[width=\linewidth]{../assets/location-aware.png}
	\end{center}
	\vspace{2em}
	Garage replicates data on different zones when possible
\end{frame}

\begin{frame}
	\frametitle{Layout computation}
		\begin{center}
			\includegraphics[width=.8\linewidth]{../assets/map.png}
		\end{center}
\end{frame}

\begin{frame}
	\frametitle{Optimal layout computation}
		\begin{center}
			\includegraphics[width=.6\linewidth]{../assets/geodistrib_paper.png}
		\end{center}
\end{frame}

\begin{frame}
	\frametitle{What a "layout" is}
	\textbf{A layout is a precomputed index table}
	\vspace{2em}

		\begin{center}
		\begin{tabular}{|l|l|l|l|}
			\hline
			\textbf{Partition} & \textbf{Node 1} & \textbf{Node 2} & \textbf{Node 3} \\
			\hline
			\hline
			Partition 0 & Io (jupiter) & Drosera (atuin) & Courgette (neptune) \\
			\hline
			Partition 1 & Datura (atuin) & Courgette (neptune) & Io (jupiter) \\
			\hline
			Partition 2 & Io(jupiter) & Celeri (neptune) & Drosera (atuin) \\
			\hline
			\hspace{1em}$\vdots$ & \hspace{1em}$\vdots$ & \hspace{1em}$\vdots$ & \hspace{1em}$\vdots$ \\
			\hline
			Partition 255 & Concombre (neptune) & Io (jupiter) & Drosera (atuin) \\
			\hline
		\end{tabular}
		\end{center}

	\vspace{1em}
		The index table is built centrally using an optimal algorithm,\\
		then propagated to all nodes
\end{frame}

\begin{frame}
	\frametitle{The relationship between \emph{partition} and \emph{partition key}}
	\begin{center}
	\begin{tabular}{|l|l|l|l|}
		\hline
		\textbf{Partition key} & \textbf{Partition} & \textbf{Sort key} & \textbf{Value} \\
		\hline
		\hline
		\texttt{website} & Partition 12 & \texttt{index.html} & (file data) \\
		\hline
		\texttt{website} & Partition 12 & \texttt{img/logo.svg} & (file data) \\
		\hline
		\texttt{website} & Partition 12 &\texttt{download/index.html} & (file data) \\
		\hline
		\hline
		\texttt{backup} & Partition 42 & \texttt{borg/index.2822} & (file data) \\
		\hline
		\texttt{backup} & Partition 42 & \texttt{borg/data/2/2329} & (file data) \\
		\hline
		\texttt{backup} & Partition 42 & \texttt{borg/data/2/2680} & (file data) \\
		\hline
		\hline
		\texttt{private} & Partition 42 & \texttt{qq3a2nbe1qjq0ebbvo6ocsp6co} & (file data) \\
		\hline
	\end{tabular}
	\end{center}
	\vspace{1em}
	\textbf{To read or write an item:} hash partition key
	\\ \hspace{5cm} $\to$ determine partition number (first 8 bits)
	\\ \hspace{5cm} $\to$ find associated nodes
\end{frame}

% ====================== v0.10.0 ===============================

\begin{frame}
	\frametitle{October 2023 - Garage v0.10.0 beta}
	Focus on \underline{consistency}
	\vspace{2em}
	\begin{itemize}
		\item Fix consistency issues when reshuffling data
	\end{itemize}
\end{frame}




% --------------------------------------------------------------------
% =================================================================================
% =================================================================================
% --------------------------------------------------------------------
% --------------------------------------------------------------------
% --------------------------------------------------------------------
% =================================================================================
% =================================================================================
% --------------------------------------------------------------------




\begin{frame}
	\frametitle{Two big problems}
	\begin{enumerate}
		\item \textbf{How to place data on different nodes?}\\
			\vspace{1em}
			\underline{Constraints:} heterogeneous hardware\\
			\underline{Objective:} $n$ copies of everything, maximize usable capacity, maximize resilience\\
			\vspace{1em}
			$\to$ the Dynamo model + optimization algorithms
			\vspace{2em}
		\item<2-> \textbf{How to guarantee consistency?}\\
			\vspace{1em}
			\underline{Constraints:} slow network (geographical distance), node unavailability/crashes\\
			\underline{Objective:} maximize availability, read-after-write guarantee\\
			\vspace{1em}
			$\to$ CRDTs, monotonicity, read and write quorums
	\end{enumerate}
\end{frame}

\section{Problem 1: placing data}

\begin{frame}
	\frametitle{Key-value stores, upgraded: the Dynamo model}
	\textbf{Two keys:}
	\begin{itemize}
		\item Partition key: used to divide data into partitions {\small (a.k.a.~shards)}
		\item Sort key: used to identify items inside a partition
	\end{itemize}

	\vspace{1em}

	\begin{center}
	\begin{tabular}{|l|l|p{3cm}|}
		\hline
		\textbf{Partition key: bucket} & \textbf{Sort key: filename} & \textbf{Value} \\
		\hline
		\hline
		\texttt{website} & \texttt{index.html} & (file data) \\
		\hline
		\texttt{website} & \texttt{img/logo.svg} & (file data) \\
		\hline
		\texttt{website} & \texttt{download/index.html} & (file data) \\
		\hline
		\hline
		\texttt{backup} & \texttt{borg/index.2822} & (file data) \\
		\hline
		\texttt{backup} & \texttt{borg/data/2/2329} & (file data) \\
		\hline
		\texttt{backup} & \texttt{borg/data/2/2680} & (file data) \\
		\hline
		\hline
		\texttt{private} & \texttt{qq3a2nbe1qjq0ebbvo6ocsp6co} & (file data) \\
		\hline
	\end{tabular}
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Key-value stores, upgraded: the Dynamo model}
	\begin{itemize}
		\item Data with different partition keys is stored independently,\\
			on a different set of nodes\\
			\vspace{.5em}
			$\to$ no easy way to list all partition keys\\
			$\to$ no cross-shard transactions\\
			\vspace{2em}
		\item Placing data: hash the partition key, select nodes accordingly\\
			\vspace{.5em}
			$\to$ distributed hash table (DHT)
			\vspace{2em}
		\item For a given value of the partition key, items can be listed using their sort keys
	\end{itemize}
\end{frame}


\begin{frame}
	\frametitle{Issues with consistent hashing}
	\begin{itemize}
		\item Consistent hashing doesn't dispatch data based on geographical location of nodes
			\vspace{1em}
		\item<2-> Geographically aware adaptation, try 1:\\
			data quantities not well balanced between nodes
			\vspace{1em}
		\item<3-> Geographically aware adaptation, try 2:\\
			too many reshuffles when adding/removing nodes
	\end{itemize}
\end{frame}


\section{Problem 2: ensuring consistency}

\begin{frame}
	\frametitle{Consensus vs weak consistency}

	\hspace{1em}
	\begin{minipage}{7cm}
		\textbf{Consensus-based systems:}
		\vspace{1em}
		\begin{itemize}
			\item \textbf{Leader-based:} a leader is elected to coordinate
				all reads and writes
				\vspace{1em}
			\item \textbf{Linearizability} of all operations\\
				(strongest consistency guarantee)
				\vspace{1em}
			\item Any sequential specification can be implemented as a \textbf{replicated state machine}
				\vspace{1em}
			\item \textbf{Costly}, the leader is a bottleneck;
				leader elections on failure take time
		\end{itemize}
	\end{minipage}
	\hfill
	\begin{minipage}{7cm} \visible<2->{
		\textbf{Weakly consistent systems:}
		\vspace{1em}
		\begin{itemize}
			\item \textbf{Nodes are equivalent}, any node
				can originate a read or write operation
				\vspace{1em}
			\item \textbf{Read-after-write consistency} with quorums,
				eventual consistency without
				\vspace{1em}
			\item \textbf{Operations have to commute}, i.e.~we
				can only implement CRDTs
				\vspace{1em}
			\item \textbf{Fast}, no single bottleneck;\\
				works the same with offline nodes
		\end{itemize}
	} \end{minipage}
	\hspace{1em}
\end{frame}

\begin{frame}
	\frametitle{Consensus vs weak consistency}
	\begin{center}
		\textbf{From a theoretical point of view:}\\

	\end{center}
	\vspace{2em}

	\hspace{1em}
	\begin{minipage}{6.5cm}
		\underline{Consensus-based systems:}

		\vspace{1em}

		Require \textbf{additional assumptions} such as a fault detector or a strong RNG\\
		(FLP impossibility theorem)
	\end{minipage}
	\hfill
	\begin{minipage}{6.5cm}
		\underline{Weakly consistent systems:}

		\vspace{1em}
		
		Can be implemented in \textbf{any\\asynchronous message passing\\distributed system} with node crashes
	\end{minipage}
	\hspace{1em}

	\vspace{3em}
	\begin{center}
		They represent \textbf{different classes of computational capability}\\
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Consensus vs weak consistency}
	\begin{center}
		\textbf{The same objects cannot be implemented in both models.}
	\end{center}
	\vspace{2em}

	\hspace{1em}
	\begin{minipage}{6.5cm}
		\underline{Consensus-based systems:}

		\vspace{1em}

		\textbf{Any sequential specification}\\~

		\vspace{1em}
		\textbf{Easier to program for}: just write your program as if it were sequential on a single machine

	\end{minipage}
	\hfill
	\begin{minipage}{6.5cm}
		\underline{Weakly consistent systems:}

		\vspace{1em}
		
		\textbf{Only CRDTs}\\(conflict-free replicated data types)

		\vspace{1em}
		Part of the complexity is \textbf{reported to the consumer of the API}\\~
	\end{minipage}
	\hspace{1em}
\end{frame}

\begin{frame}
	\frametitle{Understanding the power of consensus}
	\textbf{Consensus:} an API with a single operation, $propose(x)$
	\begin{enumerate}
		\item nodes all call $propose(x)$ with their proposed value;
		\item nodes all receive the same value as a return value, which is one of the proposed values
	\end{enumerate}
	\vspace{1em}
	
	\visible<2->{
		\textbf{Equivalent to} a distributed algorithm that gives a total order on all requests
	}
	\vspace{1em}

	\visible<3->{
		\textbf{Implemented by} this simple replicated state machine:
		\vspace{.5em}
		\begin{figure}
			\centering
			\def\svgwidth{.5\textwidth}
			\large
			\import{assets/}{consensus.pdf_tex}
		\end{figure}
		\vspace{1em}
	}
\end{frame}

\begin{frame}
	\frametitle{Can my object be implemented without consensus?}
	\underline{Given the specification of an API:}
	\vspace{2em}
	\begin{itemize}
		\item \textbf{Using this API, we can implement the consensus object} (the $propose$ function)\\
			$\to$ the API is equivalent to consensus/total ordering of messages\\
			$\to$ the API cannot be implemented in a weakly consistent system
			\vspace{2em}
		\item<2-> \textbf{This API can be implemented using only weak primitives}\\
			(e.g. in the asynchronous message passing model with no further assumption)\\
			$\to$ the API is strictly weaker than consensus\\
			$\to$ we can implement it in Garage!
	\end{itemize}
\end{frame}


\begin{frame}
	\frametitle{What can we implement without consensus?}
	\begin{itemize}
		\item Any \textbf{conflict-free replicated data type} (CRDT)
			\vspace{1em}
		\item<2-> Non-transactional key-value stores such as S3 are equivalent to a simple CRDT:\\
			a map of \textbf{last-writer-wins registers} (each key is its own CRDT)
			\vspace{1em}
		\item<3-> \textbf{Read-after-write consistency} can be implemented
			using quorums on read and write operations
			\vspace{1em}
		\item<4-> \textbf{Monotonicity of reads} can be implemented with repair-on-read\\
			(makes reads more costly, not implemented in Garage)
	\end{itemize}
\end{frame}

\begin{frame}
	\frametitle{CRDTs and quorums: read-after-write consistency}
	\begin{figure}
		\centering
		\def\svgwidth{.8\textwidth}
		\only<1>{\import{assets/}{lattice1.pdf_tex}}%
		\only<2>{\import{assets/}{lattice2.pdf_tex}}%
		\only<3>{\import{assets/}{lattice3.pdf_tex}}%
		\only<4>{\import{assets/}{lattice4.pdf_tex}}%
		\only<5>{\import{assets/}{lattice5.pdf_tex}}%
		\only<6>{\import{assets/}{lattice6.pdf_tex}}%
		\only<7>{\import{assets/}{lattice7.pdf_tex}}%
		\only<8>{\import{assets/}{lattice8.pdf_tex}}%
	\end{figure}
\end{frame}

\begin{frame}
	\frametitle{CRDTs and quorums: read-after-write consistency}
	\textbf{Property:} If node $A$ did an operation $write(x)$ and received an OK response,\\
		\hspace{2cm} and node $B$ starts an operation $read()$ after $A$ received OK,\\
		\hspace{2cm} then $B$ will read a value $x' \sqsupseteq x$.

	\vspace{1em}

	\hspace{1em}
	\begin{minipage}{6.8cm}
		\textbf{Algorithm $write(x)$:}
		\begin{enumerate}
			\item Broadcast $write(x)$ to all nodes
			\item Wait for $k > n/2$ nodes to reply OK
			\item Return OK
		\end{enumerate}
	\end{minipage}
	\hfill
	\begin{minipage}{6.8cm}
		\vspace{1em}
		\textbf{Algorithm $read()$:}
		\begin{enumerate}
			\item Broadcast $read()$ to all nodes
			\item Wait for $k > n/2$ nodes to reply\\
				with values $x_1, \dots, x_k$
			\item Return $x_1 \sqcup \dots \sqcup x_k$
		\end{enumerate}
	\end{minipage}
	\hspace{1em}

	\vspace{2em}
	\textbf{Why does it work?} There is at least one node at the intersection between the two sets of nodes that replied to each request, that ``saw'' $x$ before the $read()$ started ($x_i \sqsupseteq x$).
\end{frame}

\begin{frame}
	\frametitle{CRDTs and quorums: monotonic-reads consistency}
	\begin{figure}
		\centering
		\def\svgwidth{.8\textwidth}
		\only<1>{\import{assets/}{latticeB_1.pdf_tex}}%
		\only<2>{\import{assets/}{latticeB_2.pdf_tex}}%
		\only<3>{\import{assets/}{latticeB_3.pdf_tex}}%
		\only<4>{\import{assets/}{latticeB_4.pdf_tex}}%
		\only<5>{\import{assets/}{latticeB_5.pdf_tex}}%
		\only<6>{\import{assets/}{latticeB_6.pdf_tex}}%
		\only<7>{\import{assets/}{latticeB_7.pdf_tex}}%
		\only<8>{\import{assets/}{latticeB_8.pdf_tex}}%
		\only<9>{\import{assets/}{latticeB_9.pdf_tex}}%
		\only<10>{\import{assets/}{latticeB_10.pdf_tex}}%
	\end{figure}
\end{frame}

\begin{frame}
	\frametitle{CRDTs and quorums: monotonic-reads consistency}
	\textbf{Property:} If node $A$ did an operation $read()$ and received $x$ as a response,\\
		\hspace{2cm} and node $B$ starts an operation $read()$ after $A$ received $x$,\\
		\hspace{2cm} then $B$ will read a value $x' \sqsupseteq x$.

	\vspace{1em}

	\textbf{Algorithm $monotonic\_read()$:} {\small (a.k.a. repair-on-read)}
	\begin{enumerate}
		\item Broadcast $read()$ to all nodes
		\item Wait for $k > n/2$ nodes to reply with values $x_1, \dots, x_k$
		\item If $x_i \ne x_j$ for some nodes $i$ and $j$,\\
			\hspace{1cm}then call $write(x_1 \sqcup \dots \sqcup x_k)$ and wait for OK from $k' > n/2$ nodes
		\item Return $x_1 \sqcup \dots \sqcup x_k$
	\end{enumerate}

	\vspace{1em}

	This makes reads slower in some cases, and is \textbf{not implemented in Garage}.
\end{frame}

\begin{frame}
	\frametitle{A hard problem: layout changes}
	\begin{itemize}
		\item We rely on quorums $k > n/2$ within each partition:\\
			$$n=3,~~~~~~~k\ge 2$$
		\item<2-> When rebalancing, the set of nodes responsible for a partition can change:\\
			$$\{n_A, n_B, n_C\} \to \{n_A, n_D, n_E\}$$
			\vspace{.01em}
		\item<3-> During the rebalancing, $D$ and $E$ don't yet have the data,\\
			~~~~~~~~~~~~~~~~~~~and $B$ and $C$ want to get rid of the data to free up space\\
			\vspace{.2em}
			$\to$ quorums only within the new set of nodes don't work\\
			$\to$ how to coordinate? \textbf{currently, we don't...}

	\end{itemize}
\end{frame}

\section{Operating big Garage clusters}

\begin{frame}
	\frametitle{Operating Garage}
	\begin{center}
		\only<1-2>{
			\includegraphics[width=.9\linewidth]{assets/scr_garage_status.png}
			\\\vspace{1em}
			\visible<2>{\includegraphics[width=.85\linewidth]{assets/scr_garage_status_broken.png}}
		}
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Garage's architecture}
	\begin{center}
		\only<1>{\includegraphics[width=.45\linewidth]{assets/garage.drawio.pdf}}%
		\only<2>{\includegraphics[width=.6\linewidth]{assets/garage_sync.drawio.pdf}}%
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Digging deeper}
	\begin{center}
		\only<1>{\includegraphics[width=.9\linewidth]{assets/scr_garage_stats.png}}
		\only<2>{\includegraphics[width=.6\linewidth]{assets/scr_garage_worker_list.png}}
		\only<3>{\includegraphics[width=.6\linewidth]{assets/scr_garage_worker_get.png}}
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Potential limitations and bottlenecks}
	\begin{itemize}
		\item Global:
			\begin{itemize}
				\item Max. $\sim$100 nodes per cluster (excluding gateways)
			\end{itemize}
			\vspace{1em}
		\item Metadata:
			\begin{itemize}
				\item One big bucket = bottleneck, object list on 3 nodes only
			\end{itemize}
			\vspace{1em}
		\item Block manager:
			\begin{itemize}
				\item Lots of small files on disk
				\item Processing the resync queue can be slow
				\item Multi-HDD support not yet released (soon!)
			\end{itemize}
	\end{itemize}
\end{frame}

\begin{frame}
	\frametitle{Deployment advice for very large clusters}
	\begin{itemize}
		\item Metadata storage:
			\begin{itemize}
				\item ZFS mirror (x2) on fast NVMe
				\item Use LMDB storage engine
			\end{itemize}
			\vspace{.5em}
		\item Data block storage:
			\begin{itemize}
				\item Wait for v0.9 with multi-HDD support
				\item XFS on individual drives
				\item Increase block size (1MB $\to$ 10MB, requires more RAM and good networking)
				\item Tune \texttt{resync-tranquility} and \texttt{resync-worker-count} dynamically
			\end{itemize}
			\vspace{.5em}
		\item Other :
			\begin{itemize}
				\item Split data over several buckets
				\item Use less than 100 storage nodes
				\item Use gateway nodes
			\end{itemize}
			\vspace{.5em}
	\end{itemize}
	Current deployments: $< 10$ TB, we don't have much experience with more
\end{frame}

\section{Going further than the S3 API}

\begin{frame}
	\frametitle{Using Garage for everything}
	\begin{center}
		\only<1>{\includegraphics[width=.8\linewidth]{assets/slideB1.png}}%
		\only<2>{\includegraphics[width=.8\linewidth]{assets/slideB2.png}}%
		\only<3>{\includegraphics[width=.8\linewidth]{assets/slideB3.png}}%
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{K2V Design}
	\begin{itemize}
		\item A new, custom, minimal API\\
			\vspace{.5em}
			\begin{itemize}
				\item Single-item operations
				\item Operations on ranges and batches of items
				\item Polling operations to help implement a PubSub pattern
			\end{itemize}
			\vspace{1em}
		\item<2-> Exposes the partitoning mechanism of Garage\\
			K2V = partition key / sort key / value (like Dynamo)
			\vspace{1em}
		\item<3-> Weakly consistent, CRDT-friendly\\
			$\to$ no support for transactions (not ACID)
			\vspace{1em}
		\item<4-> Cryptography-friendly: values are binary blobs
	\end{itemize}
\end{frame}

\begin{frame}
	\frametitle{Handling concurrent values}
	\textbf{How to handle concurrency?} Example:
	\vspace{1em}
	\begin{enumerate}
		\item Client $A$ reads the initial value of a key, $x_0$
			\vspace{1em}
		\item<2-> Client $B$ also reads the initial value $x_0$ of that key
			\vspace{1em}
		\item<3-> Client $A$ modifies $x_0$, and writes a new value $x_1$
			\vspace{1em}
		\item<4-> Client $B$ also modifies $x_0$, and writes a new value $x'_1$,\\
			without having a chance to first read $x_1$\\
			\vspace{1em}
			$\to$ what should the final state be?
	\end{enumerate}
\end{frame}

\begin{frame}
	\frametitle{Handling concurrent values}
	\begin{itemize}
		\item If we keep only $x_1$ or $x'_1$, we risk \textbf{loosing application data}
			\vspace{1.5em}
		\item<2-> Values are opaque binary blobs, \textbf{K2V cannot resolve conflicts} by itself\\
			(e.g. by implementing a CRDT)
			\vspace{1.5em}
		\item<3-> Solution: \textbf{we keep both!}\\
			$\to$ the value of the key is now $\{x_1, x'_1\}$\\
			$\to$ the client application can decide how to resolve conflicts on the next read
	\end{itemize}
\end{frame}

\begin{frame}
	\frametitle{Keeping track of causality}
	How does K2V know that $x_1$ and $x'_1$ are concurrent?
	\vspace{1em}
	\begin{itemize}
		\item $read()$ returns \textbf{a set of values} and an associated \textbf{causality token}\\
			\vspace{1.5em}
		\item<2-> When calling $write()$, the client sends \textbf{the causality token from its last read}
			\vspace{1.5em}
		\item<3-> The causality token represents the set of values \textbf{already seen by the client}\\
			$\to$ those values are the \textbf{causal past} of the write operation\\
			$\to$ K2V can keep concurrent values and overwrite all ones in the causal past
			\vspace{1.5em}
		\item<4-> Internally, the causality token is \textbf{a vector clock}
	\end{itemize}
\end{frame}

\begin{frame}
	\frametitle{Application: an e-mail storage server}
	\begin{center}
		\only<1>{\includegraphics[width=.9\linewidth]{assets/aerogramme.png}}%
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{Aerogramme data model}
	\begin{center}
		\only<1->{\includegraphics[width=.4\linewidth]{assets/aerogramme_datatype.drawio.pdf}}%
	\end{center}
	\visible<2->{Aerogramme encrypts all stored values for privacy\\
	(Garage server administrators can't read your mail)}
\end{frame}

\begin{frame}
	\frametitle{Different deployment scenarios}
	\begin{center}
		\only<1>{\includegraphics[width=.9\linewidth]{assets/aerogramme_components1.drawio.pdf}}%
		\only<2>{\includegraphics[width=.9\linewidth]{assets/aerogramme_components2.drawio.pdf}}%
	\end{center}
\end{frame}

\begin{frame}
	\frametitle{A new model for building resilient software}
	How to build an application using only Garage as a data store:
	\vspace{1em}
	\begin{enumerate}
		\item Design a data model suited to K2V\\
			{\footnotesize (see Cassandra docs on porting SQL data models to Cassandra)}
			\vspace{1em}
			\begin{itemize}
				\item Use CRDTs or other eventually consistent data types (see e.g. Bayou)
					\vspace{1em}
				\item Store opaque binary blobs to provide End-to-End Encryption\\
			\end{itemize}
			\vspace{1em}
		\item<2-> Store big blobs (files) using the S3 API
			\vspace{1em}
		\item<3-> Let Garage manage sharding, replication, failover, etc.
	\end{enumerate}
\end{frame}

\section{Conclusion}

\begin{frame}
	\frametitle{Perspectives}
	\begin{itemize}
		\item Fix the consistency issue when rebalancing
			\vspace{1em}
		\item Write about Garage's architecture and properties,\\
			and about our proposed architecture for (E2EE) apps over K2V+S3
			\vspace{1em}
		\item Continue developing Garage; finish Aerogramme; build new applications...
			\vspace{1em}
		\item Anything else?
	\end{itemize}
\end{frame}

\begin{frame}
	\frametitle{Where to find us}
	\begin{center}
			\includegraphics[width=.25\linewidth]{../../logo/garage_hires.png}\\
			\vspace{-1em}
		\url{https://garagehq.deuxfleurs.fr/}\\
		\url{mailto:garagehq@deuxfleurs.fr}\\
		\texttt{\#garage:deuxfleurs.fr} on Matrix

		\vspace{1.5em}
			\includegraphics[width=.06\linewidth]{assets/rust_logo.png}
			\includegraphics[width=.13\linewidth]{assets/AGPLv3_Logo.png}
	\end{center}
\end{frame}

\end{document}

%% vim: set ts=4 sw=4 tw=0 noet spelllang=en :