12708: Pull blocks to desired storage classes.
[arvados.git] / services / keep-balance / balance_test.go
1 // Copyright (C) The Arvados Authors. All rights reserved.
2 //
3 // SPDX-License-Identifier: AGPL-3.0
4
5 package main
6
7 import (
8         "crypto/md5"
9         "fmt"
10         "sort"
11         "strconv"
12         "testing"
13         "time"
14
15         "git.curoverse.com/arvados.git/sdk/go/arvados"
16
17         check "gopkg.in/check.v1"
18 )
19
20 // Test with Gocheck
21 func Test(t *testing.T) {
22         check.TestingT(t)
23 }
24
25 var _ = check.Suite(&balancerSuite{})
26
27 type balancerSuite struct {
28         Balancer
29         srvs            []*KeepService
30         blks            map[string]tester
31         knownRendezvous [][]int
32         signatureTTL    int64
33 }
34
35 const (
36         // index into knownRendezvous
37         known0 = 0
38 )
39
40 type slots []int
41
42 type tester struct {
43         known       int
44         desired     map[string]int
45         current     slots
46         timestamps  []int64
47         shouldPull  slots
48         shouldTrash slots
49
50         shouldPullMounts  []string
51         shouldTrashMounts []string
52 }
53
54 func (bal *balancerSuite) SetUpSuite(c *check.C) {
55         bal.knownRendezvous = nil
56         for _, str := range []string{
57                 "3eab2d5fc9681074",
58                 "097dba52e648f1c3",
59                 "c5b4e023f8a7d691",
60                 "9d81c02e76a3bf54",
61         } {
62                 var slots []int
63                 for _, c := range []byte(str) {
64                         pos, _ := strconv.ParseUint(string(c), 16, 4)
65                         slots = append(slots, int(pos))
66                 }
67                 bal.knownRendezvous = append(bal.knownRendezvous, slots)
68         }
69
70         bal.signatureTTL = 3600
71 }
72
73 func (bal *balancerSuite) SetUpTest(c *check.C) {
74         bal.srvs = make([]*KeepService, 16)
75         bal.KeepServices = make(map[string]*KeepService)
76         for i := range bal.srvs {
77                 srv := &KeepService{
78                         KeepService: arvados.KeepService{
79                                 UUID: fmt.Sprintf("zzzzz-bi6l4-%015x", i),
80                         },
81                 }
82                 srv.mounts = []*KeepMount{{
83                         KeepMount: arvados.KeepMount{
84                                 UUID: fmt.Sprintf("zzzzz-mount-%015x", i),
85                         },
86                         KeepService: srv,
87                 }}
88                 bal.srvs[i] = srv
89                 bal.KeepServices[srv.UUID] = srv
90         }
91
92         bal.MinMtime = time.Now().UnixNano() - bal.signatureTTL*1e9
93 }
94
95 func (bal *balancerSuite) TestPerfect(c *check.C) {
96         bal.try(c, tester{
97                 desired:     map[string]int{"default": 2},
98                 current:     slots{0, 1},
99                 shouldPull:  nil,
100                 shouldTrash: nil})
101 }
102
103 func (bal *balancerSuite) TestDecreaseRepl(c *check.C) {
104         bal.try(c, tester{
105                 desired:     map[string]int{"default": 2},
106                 current:     slots{0, 2, 1},
107                 shouldTrash: slots{2}})
108 }
109
110 func (bal *balancerSuite) TestDecreaseReplToZero(c *check.C) {
111         bal.try(c, tester{
112                 desired:     map[string]int{"default": 0},
113                 current:     slots{0, 1, 3},
114                 shouldTrash: slots{0, 1, 3}})
115 }
116
117 func (bal *balancerSuite) TestIncreaseRepl(c *check.C) {
118         bal.try(c, tester{
119                 desired:    map[string]int{"default": 4},
120                 current:    slots{0, 1},
121                 shouldPull: slots{2, 3}})
122 }
123
124 func (bal *balancerSuite) TestSkipReadonly(c *check.C) {
125         bal.srvList(0, slots{3})[0].ReadOnly = true
126         bal.try(c, tester{
127                 desired:    map[string]int{"default": 4},
128                 current:    slots{0, 1},
129                 shouldPull: slots{2, 4}})
130 }
131
132 func (bal *balancerSuite) TestFixUnbalanced(c *check.C) {
133         bal.try(c, tester{
134                 desired:    map[string]int{"default": 2},
135                 current:    slots{2, 0},
136                 shouldPull: slots{1}})
137         bal.try(c, tester{
138                 desired:    map[string]int{"default": 2},
139                 current:    slots{2, 7},
140                 shouldPull: slots{0, 1}})
141         // if only one of the pulls succeeds, we'll see this next:
142         bal.try(c, tester{
143                 desired:     map[string]int{"default": 2},
144                 current:     slots{2, 1, 7},
145                 shouldPull:  slots{0},
146                 shouldTrash: slots{7}})
147         // if both pulls succeed, we'll see this next:
148         bal.try(c, tester{
149                 desired:     map[string]int{"default": 2},
150                 current:     slots{2, 0, 1, 7},
151                 shouldTrash: slots{2, 7}})
152
153         // unbalanced + excessive replication => pull + trash
154         bal.try(c, tester{
155                 desired:     map[string]int{"default": 2},
156                 current:     slots{2, 5, 7},
157                 shouldPull:  slots{0, 1},
158                 shouldTrash: slots{7}})
159 }
160
161 func (bal *balancerSuite) TestMultipleReplicasPerService(c *check.C) {
162         for _, srv := range bal.srvs {
163                 for i := 0; i < 3; i++ {
164                         m := *(srv.mounts[0])
165                         srv.mounts = append(srv.mounts, &m)
166                 }
167         }
168         bal.try(c, tester{
169                 desired:    map[string]int{"default": 2},
170                 current:    slots{0, 0},
171                 shouldPull: slots{1}})
172         bal.try(c, tester{
173                 desired:    map[string]int{"default": 2},
174                 current:    slots{2, 2},
175                 shouldPull: slots{0, 1}})
176         bal.try(c, tester{
177                 desired:     map[string]int{"default": 2},
178                 current:     slots{0, 0, 1},
179                 shouldTrash: slots{0}})
180         bal.try(c, tester{
181                 desired:     map[string]int{"default": 2},
182                 current:     slots{1, 1, 0},
183                 shouldTrash: slots{1}})
184         bal.try(c, tester{
185                 desired:     map[string]int{"default": 2},
186                 current:     slots{1, 0, 1, 0, 2},
187                 shouldTrash: slots{0, 1, 2}})
188         bal.try(c, tester{
189                 desired:     map[string]int{"default": 2},
190                 current:     slots{1, 1, 1, 0, 2},
191                 shouldTrash: slots{1, 1, 2}})
192         bal.try(c, tester{
193                 desired:     map[string]int{"default": 2},
194                 current:     slots{1, 1, 2},
195                 shouldPull:  slots{0},
196                 shouldTrash: slots{1}})
197         bal.try(c, tester{
198                 desired:     map[string]int{"default": 2},
199                 current:     slots{1, 1, 0},
200                 timestamps:  []int64{12345678, 12345678, 12345679},
201                 shouldTrash: nil})
202         bal.try(c, tester{
203                 desired:    map[string]int{"default": 2},
204                 current:    slots{1, 1},
205                 shouldPull: slots{0}})
206 }
207
208 func (bal *balancerSuite) TestIncreaseReplTimestampCollision(c *check.C) {
209         // For purposes of increasing replication, we assume identical
210         // replicas are distinct.
211         bal.try(c, tester{
212                 desired:    map[string]int{"default": 4},
213                 current:    slots{0, 1},
214                 timestamps: []int64{12345678, 12345678},
215                 shouldPull: slots{2, 3}})
216 }
217
218 func (bal *balancerSuite) TestDecreaseReplTimestampCollision(c *check.C) {
219         // For purposes of decreasing replication, we assume identical
220         // replicas are NOT distinct.
221         bal.try(c, tester{
222                 desired:    map[string]int{"default": 2},
223                 current:    slots{0, 1, 2},
224                 timestamps: []int64{12345678, 12345678, 12345678}})
225         bal.try(c, tester{
226                 desired:    map[string]int{"default": 2},
227                 current:    slots{0, 1, 2},
228                 timestamps: []int64{12345678, 10000000, 10000000}})
229 }
230
231 func (bal *balancerSuite) TestDecreaseReplBlockTooNew(c *check.C) {
232         oldTime := bal.MinMtime - 3600
233         newTime := bal.MinMtime + 3600
234         // The excess replica is too new to delete.
235         bal.try(c, tester{
236                 desired:    map[string]int{"default": 2},
237                 current:    slots{0, 1, 2},
238                 timestamps: []int64{oldTime, newTime, newTime + 1}})
239         // The best replicas are too new to delete, but the excess
240         // replica is old enough.
241         bal.try(c, tester{
242                 desired:     map[string]int{"default": 2},
243                 current:     slots{0, 1, 2},
244                 timestamps:  []int64{newTime, newTime + 1, oldTime},
245                 shouldTrash: slots{2}})
246 }
247
248 func (bal *balancerSuite) TestChangeStorageClasses(c *check.C) {
249         // For known blocks 0/1/2/3, server 9 is slot 9/1/14/0 in
250         // probe order. For these tests we give it two mounts, one
251         // with classes=[special], one with
252         // classes=[special,special2].
253         bal.srvs[9].mounts = []*KeepMount{{
254                 KeepMount: arvados.KeepMount{
255                         Replication:    1,
256                         StorageClasses: []string{"special"},
257                         UUID:           "zzzzz-mount-special00000009",
258                         DeviceID:       "9-special",
259                 },
260                 KeepService: bal.srvs[9],
261         }, {
262                 KeepMount: arvados.KeepMount{
263                         Replication:    1,
264                         StorageClasses: []string{"special", "special2"},
265                         UUID:           "zzzzz-mount-special20000009",
266                         DeviceID:       "9-special-and-special2",
267                 },
268                 KeepService: bal.srvs[9],
269         }}
270         // For known blocks 0/1/2/3, server 13 (d) is slot 5/3/11/1 in
271         // probe order. We give it two mounts, one with
272         // classes=[special3], one with classes=[default].
273         bal.srvs[13].mounts = []*KeepMount{{
274                 KeepMount: arvados.KeepMount{
275                         Replication:    1,
276                         StorageClasses: []string{"special2"},
277                         UUID:           "zzzzz-mount-special2000000d",
278                         DeviceID:       "13-special2",
279                 },
280                 KeepService: bal.srvs[13],
281         }, {
282                 KeepMount: arvados.KeepMount{
283                         Replication:    1,
284                         StorageClasses: []string{"default"},
285                         UUID:           "zzzzz-mount-00000000000000d",
286                         DeviceID:       "13-default",
287                 },
288                 KeepService: bal.srvs[13],
289         }}
290         // Pull to slot 9 because that's the only server with the
291         // desired class "special".
292         bal.try(c, tester{
293                 known:            0,
294                 desired:          map[string]int{"default": 2, "special": 1},
295                 current:          slots{0, 1},
296                 shouldPull:       slots{9},
297                 shouldPullMounts: []string{"zzzzz-mount-special00000009"}})
298         // If some storage classes are not satisfied, don't trash any
299         // excess replicas. (E.g., if someone desires repl=1 on
300         // class=durable, and we have two copies on class=volatile, we
301         // should wait for pull to succeed before trashing anything).
302         bal.try(c, tester{
303                 known:            0,
304                 desired:          map[string]int{"special": 1},
305                 current:          slots{0, 1},
306                 shouldPull:       slots{9},
307                 shouldPullMounts: []string{"zzzzz-mount-special00000009"}})
308         // Once storage classes are satisfied, trash excess replicas
309         // that appear earlier in probe order but aren't needed to
310         // satisfy the desired classes.
311         bal.try(c, tester{
312                 known:       0,
313                 desired:     map[string]int{"special": 1},
314                 current:     slots{0, 1, 9},
315                 shouldTrash: slots{0, 1}})
316         // Pull to slot 5, the best server with class "special2".
317         bal.try(c, tester{
318                 known:            0,
319                 desired:          map[string]int{"special2": 1},
320                 current:          slots{0, 1},
321                 shouldPull:       slots{5},
322                 shouldPullMounts: []string{"zzzzz-mount-special2000000d"}})
323         // Pull to slot 5 and 9 to get replication 2 in desired class
324         // "special2".
325         bal.try(c, tester{
326                 known:            0,
327                 desired:          map[string]int{"special2": 2},
328                 current:          slots{0, 1},
329                 shouldPull:       slots{5, 9},
330                 shouldPullMounts: []string{"zzzzz-mount-special20000009", "zzzzz-mount-special2000000d"}})
331         // Slot 0 has a replica in "default", slot 1 has a replica
332         // in "special"; we need another replica in "default", i.e.,
333         // on slot 2.
334         bal.try(c, tester{
335                 known:      1,
336                 desired:    map[string]int{"default": 2, "special": 1},
337                 current:    slots{0, 1},
338                 shouldPull: slots{2}})
339         // Pull to best probe position 0 (despite wrong storage class)
340         // if it's impossible to achieve desired replication in the
341         // desired class (only slots 1 and 3 have special2).
342         bal.try(c, tester{
343                 known:      1,
344                 desired:    map[string]int{"special2": 3},
345                 current:    slots{3},
346                 shouldPull: slots{0, 1}})
347         // Trash excess replica.
348         bal.try(c, tester{
349                 known:       3,
350                 desired:     map[string]int{"special": 1},
351                 current:     slots{0, 1},
352                 shouldTrash: slots{1}})
353         // Leave one copy on slot 1 because slot 0 (server 9) only
354         // gives us repl=1.
355         bal.try(c, tester{
356                 known:   3,
357                 desired: map[string]int{"special": 2},
358                 current: slots{0, 1}})
359 }
360
361 // Clear all servers' changesets, balance a single block, and verify
362 // the appropriate changes for that block have been added to the
363 // changesets.
364 func (bal *balancerSuite) try(c *check.C, t tester) {
365         bal.setupCaches()
366         blk := &BlockState{
367                 Replicas: bal.replList(t.known, t.current),
368                 Desired:  t.desired,
369         }
370         for i, t := range t.timestamps {
371                 blk.Replicas[i].Mtime = t
372         }
373         for _, srv := range bal.srvs {
374                 srv.ChangeSet = &ChangeSet{}
375         }
376         bal.balanceBlock(knownBlkid(t.known), blk)
377
378         var didPull, didTrash slots
379         var didPullMounts, didTrashMounts []string
380         for i, srv := range bal.srvs {
381                 var slot int
382                 for probeOrder, srvNum := range bal.knownRendezvous[t.known] {
383                         if srvNum == i {
384                                 slot = probeOrder
385                         }
386                 }
387                 for _, pull := range srv.Pulls {
388                         didPull = append(didPull, slot)
389                         didPullMounts = append(didPullMounts, pull.To.UUID)
390                         c.Check(pull.SizedDigest, check.Equals, knownBlkid(t.known))
391                 }
392                 for _, trash := range srv.Trashes {
393                         didTrash = append(didTrash, slot)
394                         didTrashMounts = append(didTrashMounts, trash.From.UUID)
395                         c.Check(trash.SizedDigest, check.Equals, knownBlkid(t.known))
396                 }
397         }
398
399         for _, list := range []slots{didPull, didTrash, t.shouldPull, t.shouldTrash} {
400                 sort.Sort(sort.IntSlice(list))
401         }
402         c.Check(didPull, check.DeepEquals, t.shouldPull)
403         c.Check(didTrash, check.DeepEquals, t.shouldTrash)
404         if t.shouldPullMounts != nil {
405                 sort.Strings(didPullMounts)
406                 c.Check(didPullMounts, check.DeepEquals, t.shouldPullMounts)
407         }
408         if t.shouldTrashMounts != nil {
409                 sort.Strings(didTrashMounts)
410                 c.Check(didTrashMounts, check.DeepEquals, t.shouldTrashMounts)
411         }
412 }
413
414 // srvList returns the KeepServices, sorted in rendezvous order and
415 // then selected by idx. For example, srvList(3, slots{0, 1, 4})
416 // returns the the first-, second-, and fifth-best servers for storing
417 // bal.knownBlkid(3).
418 func (bal *balancerSuite) srvList(knownBlockID int, order slots) (srvs []*KeepService) {
419         for _, i := range order {
420                 srvs = append(srvs, bal.srvs[bal.knownRendezvous[knownBlockID][i]])
421         }
422         return
423 }
424
425 // replList is like srvList but returns an "existing replicas" slice,
426 // suitable for a BlockState test fixture.
427 func (bal *balancerSuite) replList(knownBlockID int, order slots) (repls []Replica) {
428         nextMnt := map[*KeepService]int{}
429         mtime := time.Now().UnixNano() - (bal.signatureTTL+86400)*1e9
430         for _, srv := range bal.srvList(knownBlockID, order) {
431                 // round-robin repls onto each srv's mounts
432                 n := nextMnt[srv]
433                 nextMnt[srv] = (n + 1) % len(srv.mounts)
434
435                 repls = append(repls, Replica{srv.mounts[n], mtime})
436                 mtime++
437         }
438         return
439 }
440
441 // generate the same data hashes that are tested in
442 // sdk/go/keepclient/root_sorter_test.go
443 func knownBlkid(i int) arvados.SizedDigest {
444         return arvados.SizedDigest(fmt.Sprintf("%x+64", md5.Sum([]byte(fmt.Sprintf("%064x", i)))))
445 }