1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
28 "git.arvados.org/arvados.git/lib/controller/dblock"
29 "git.arvados.org/arvados.git/sdk/go/arvados"
30 "git.arvados.org/arvados.git/sdk/go/keepclient"
31 "github.com/jmoiron/sqlx"
32 "github.com/sirupsen/logrus"
35 // Balancer compares the contents of keepstore servers with the
36 // collections stored in Arvados, and issues pull/trash requests
37 // needed to get (closer to) the optimal data layout.
39 // In the optimal data layout: every data block referenced by a
40 // collection is replicated at least as many times as desired by the
41 // collection; there are no unreferenced data blocks older than
42 // BlobSignatureTTL; and all N existing replicas of a given data block
43 // are in the N best positions in rendezvous probe order.
44 type Balancer struct {
46 Logger logrus.FieldLogger
47 Dumper logrus.FieldLogger
54 KeepServices map[string]*KeepService
55 DefaultReplication int
60 mountsByClass map[string]map[*KeepMount]bool
62 serviceRoots map[string]string
69 // Run performs a balance operation using the given config and
70 // runOptions, and returns RunOptions suitable for passing to a
71 // subsequent balance operation.
73 // Run should only be called once on a given Balancer object.
74 func (bal *Balancer) Run(ctx context.Context, client *arvados.Client, cluster *arvados.Cluster, runOptions RunOptions) (nextRunOptions RunOptions, err error) {
75 nextRunOptions = runOptions
77 bal.logf("acquiring active lock")
78 if !dblock.KeepBalanceActive.Lock(ctx, func(context.Context) (*sqlx.DB, error) { return bal.DB, nil }) {
82 defer dblock.KeepBalanceActive.Unlock()
84 defer bal.time("sweep", "wall clock time to run one full sweep")()
86 ctx, cancel := context.WithDeadline(ctx, time.Now().Add(cluster.Collections.BalanceTimeout.Duration()))
89 go bal.reportMemorySize(ctx)
92 if bal.LostBlocksFile != "" {
93 tmpfn := bal.LostBlocksFile + ".tmp"
94 lbFile, err = os.OpenFile(tmpfn, os.O_CREATE|os.O_WRONLY, 0777)
99 err = syscall.Flock(int(lbFile.Fd()), syscall.LOCK_EX|syscall.LOCK_NB)
104 // Remove the tempfile only if we didn't get
105 // as far as successfully renaming it.
110 bal.lostBlocks = lbFile
112 bal.lostBlocks = ioutil.Discard
115 err = bal.DiscoverKeepServices(client)
120 for _, srv := range bal.KeepServices {
121 err = srv.discoverMounts(client)
128 if err = bal.CheckSanityEarly(client); err != nil {
132 // On a big site, indexing and sending trash/pull lists can
133 // take much longer than the usual 5 minute client
134 // timeout. From here on, we rely on the context deadline
135 // instead, aborting the entire operation if any part takes
139 rs := bal.rendezvousState()
140 if runOptions.CommitTrash && rs != runOptions.SafeRendezvousState {
141 if runOptions.SafeRendezvousState != "" {
142 bal.logf("notice: KeepServices list has changed since last run")
144 bal.logf("clearing existing trash lists, in case the new rendezvous order differs from previous run")
145 if err = bal.ClearTrashLists(ctx, client); err != nil {
148 // The current rendezvous state becomes "safe" (i.e.,
149 // OK to compute changes for that state without
150 // clearing existing trash lists) only now, after we
151 // succeed in clearing existing trash lists.
152 nextRunOptions.SafeRendezvousState = rs
155 if err = bal.GetCurrentState(ctx, client, cluster.Collections.BalanceCollectionBatch, cluster.Collections.BalanceCollectionBuffers); err != nil {
158 bal.ComputeChangeSets()
159 bal.PrintStatistics()
160 if err = bal.CheckSanityLate(); err != nil {
168 err = os.Rename(bal.LostBlocksFile+".tmp", bal.LostBlocksFile)
174 if runOptions.CommitPulls {
175 err = bal.CommitPulls(ctx, client)
177 // Skip trash if we can't pull. (Too cautious?)
181 if runOptions.CommitTrash {
182 err = bal.CommitTrash(ctx, client)
187 if runOptions.CommitConfirmedFields {
188 err = bal.updateCollections(ctx, client, cluster)
196 // SetKeepServices sets the list of KeepServices to operate on.
197 func (bal *Balancer) SetKeepServices(srvList arvados.KeepServiceList) error {
198 bal.KeepServices = make(map[string]*KeepService)
199 for _, srv := range srvList.Items {
200 bal.KeepServices[srv.UUID] = &KeepService{
202 ChangeSet: &ChangeSet{},
208 // DiscoverKeepServices sets the list of KeepServices by calling the
209 // API to get a list of all services, and selecting the ones whose
210 // ServiceType is "disk"
211 func (bal *Balancer) DiscoverKeepServices(c *arvados.Client) error {
212 bal.KeepServices = make(map[string]*KeepService)
213 return c.EachKeepService(func(srv arvados.KeepService) error {
214 if srv.ServiceType == "disk" {
215 bal.KeepServices[srv.UUID] = &KeepService{
217 ChangeSet: &ChangeSet{},
220 bal.logf("skipping %v with service type %q", srv.UUID, srv.ServiceType)
226 func (bal *Balancer) cleanupMounts() {
227 rwdev := map[string]*KeepService{}
228 for _, srv := range bal.KeepServices {
229 for _, mnt := range srv.mounts {
231 rwdev[mnt.UUID] = srv
235 // Drop the readonly mounts whose device is mounted RW
237 for _, srv := range bal.KeepServices {
238 var dedup []*KeepMount
239 for _, mnt := range srv.mounts {
240 if !mnt.AllowWrite && rwdev[mnt.UUID] != nil {
241 bal.logf("skipping srv %s readonly mount %q because same volume is mounted read-write on srv %s", srv, mnt.UUID, rwdev[mnt.UUID])
243 dedup = append(dedup, mnt)
248 for _, srv := range bal.KeepServices {
249 for _, mnt := range srv.mounts {
250 if mnt.Replication <= 0 {
251 log.Printf("%s: mount %s reports replication=%d, using replication=1", srv, mnt.UUID, mnt.Replication)
258 // CheckSanityEarly checks for configuration and runtime errors that
259 // can be detected before GetCurrentState() and ComputeChangeSets()
262 // If it returns an error, it is pointless to run GetCurrentState or
263 // ComputeChangeSets: after doing so, the statistics would be
264 // meaningless and it would be dangerous to run any Commit methods.
265 func (bal *Balancer) CheckSanityEarly(c *arvados.Client) error {
266 u, err := c.CurrentUser()
268 return fmt.Errorf("CurrentUser(): %v", err)
270 if !u.IsActive || !*u.IsAdmin {
271 return fmt.Errorf("current user (%s) is not an active admin user", u.UUID)
273 for _, srv := range bal.KeepServices {
274 if srv.ServiceType == "proxy" {
275 return fmt.Errorf("config error: %s: proxy servers cannot be balanced", srv)
278 for _, c := range bal.ChunkPrefix {
279 if !strings.ContainsRune("0123456789abcdef", c) {
280 return fmt.Errorf("invalid char %q in chunk prefix %q: only lowercase hex digits make sense", string(c), bal.ChunkPrefix)
283 if len(bal.ChunkPrefix) > 32 {
284 return fmt.Errorf("invalid chunk prefix %q: longer than a block hash", bal.ChunkPrefix)
287 mountProblem := false
288 type deviceMount struct {
292 deviceMounted := map[string]deviceMount{} // DeviceID -> mount
293 for _, srv := range bal.KeepServices {
294 for _, mnt := range srv.mounts {
295 if first, dup := deviceMounted[mnt.DeviceID]; dup && first.mnt.UUID != mnt.UUID && mnt.DeviceID != "" {
296 bal.logf("config error: device %s is mounted with multiple volume UUIDs: %s on %s, and %s on %s",
298 first.mnt.UUID, first.srv,
303 deviceMounted[mnt.DeviceID] = deviceMount{srv, mnt}
307 return errors.New("cannot continue with config errors (see above)")
310 var checkPage arvados.CollectionList
311 if err = c.RequestAndDecode(&checkPage, "GET", "arvados/v1/collections", nil, arvados.ResourceListParams{
315 IncludeOldVersions: true,
316 Filters: []arvados.Filter{{
323 } else if n := checkPage.ItemsAvailable; n > 0 {
324 return fmt.Errorf("%d collections exist with null modified_at; cannot fetch reliably", n)
330 // rendezvousState returns a fingerprint (e.g., a sorted list of
331 // UUID+host+port) of the current set of keep services.
332 func (bal *Balancer) rendezvousState() string {
333 srvs := make([]string, 0, len(bal.KeepServices))
334 for _, srv := range bal.KeepServices {
335 srvs = append(srvs, srv.String())
338 return strings.Join(srvs, "; ")
341 // ClearTrashLists sends an empty trash list to each keep
342 // service. Calling this before GetCurrentState avoids races.
344 // When a block appears in an index, we assume that replica will still
345 // exist after we delete other replicas on other servers. However,
346 // it's possible that a previous rebalancing operation made different
347 // decisions (e.g., servers were added/removed, and rendezvous order
348 // changed). In this case, the replica might already be on that
349 // server's trash list, and it might be deleted before we send a
350 // replacement trash list.
352 // We avoid this problem if we clear all trash lists before getting
353 // indexes. (We also assume there is only one rebalancing process
354 // running at a time.)
355 func (bal *Balancer) ClearTrashLists(ctx context.Context, c *arvados.Client) error {
356 for _, srv := range bal.KeepServices {
357 srv.ChangeSet = &ChangeSet{}
359 return bal.CommitTrash(ctx, c)
362 // GetCurrentState determines the current replication state, and the
363 // desired replication level, for every block that is either
364 // retrievable or referenced.
366 // It determines the current replication state by reading the block index
367 // from every known Keep service.
369 // It determines the desired replication level by retrieving all
370 // collection manifests in the database (API server).
372 // It encodes the resulting information in BlockStateMap.
373 func (bal *Balancer) GetCurrentState(ctx context.Context, c *arvados.Client, pageSize, bufs int) error {
374 ctx, cancel := context.WithCancel(ctx)
377 defer bal.time("get_state", "wall clock time to get current state")()
378 bal.BlockStateMap = NewBlockStateMap()
380 dd, err := c.DiscoveryDocument()
384 bal.DefaultReplication = dd.DefaultCollectionReplication
385 bal.MinMtime = time.Now().UnixNano() - dd.BlobSignatureTTL*1e9
387 errs := make(chan error, 1)
388 wg := sync.WaitGroup{}
390 // When a device is mounted more than once, we will get its
391 // index only once, and call AddReplicas on all of the mounts.
392 // equivMount keys are the mounts that will be indexed, and
393 // each value is a list of mounts to apply the received index
395 equivMount := map[*KeepMount][]*KeepMount{}
396 // deviceMount maps each device ID to the one mount that will
397 // be indexed for that device.
398 deviceMount := map[string]*KeepMount{}
399 for _, srv := range bal.KeepServices {
400 for _, mnt := range srv.mounts {
401 equiv := deviceMount[mnt.UUID]
404 deviceMount[mnt.UUID] = equiv
406 equivMount[equiv] = append(equivMount[equiv], mnt)
410 // Start one goroutine for each (non-redundant) mount:
411 // retrieve the index, and add the returned blocks to
413 for _, mounts := range equivMount {
415 go func(mounts []*KeepMount) {
417 bal.logf("mount %s: retrieve index from %s", mounts[0], mounts[0].KeepService)
418 idx, err := mounts[0].KeepService.IndexMount(ctx, c, mounts[0].UUID, bal.ChunkPrefix)
421 case errs <- fmt.Errorf("%s: retrieve index: %v", mounts[0], err):
428 // Some other goroutine encountered an
429 // error -- any further effort here
433 for _, mount := range mounts {
434 bal.logf("%s: add %d entries to map", mount, len(idx))
435 bal.BlockStateMap.AddReplicas(mount, idx)
436 bal.logf("%s: added %d entries to map at %dx (%d replicas)", mount, len(idx), mount.Replication, len(idx)*mount.Replication)
438 bal.logf("mount %s: index done", mounts[0])
442 collQ := make(chan arvados.Collection, bufs)
444 // Retrieve all collections from the database and send them to
449 err = EachCollection(ctx, bal.DB, c,
450 func(coll arvados.Collection) error {
453 // some other GetCurrentState
454 // error happened: no point
457 return fmt.Errorf("")
460 }, func(done, total int) {
461 bal.logf("collections: %d/%d", done, total)
473 // Parse manifests from collQ and pass the block hashes to
474 // BlockStateMap to track desired replication.
475 for i := 0; i < runtime.NumCPU(); i++ {
479 for coll := range collQ {
480 err := bal.addCollection(coll)
481 if err != nil || len(errs) > 0 {
489 atomic.AddInt64(&bal.collScanned, 1)
501 func (bal *Balancer) addCollection(coll arvados.Collection) error {
502 blkids, err := coll.SizedDigests()
504 return fmt.Errorf("%v: %v", coll.UUID, err)
506 repl := bal.DefaultReplication
507 if coll.ReplicationDesired != nil {
508 repl = *coll.ReplicationDesired
510 if bal.ChunkPrefix != "" {
511 // Throw out blocks that don't match the requested
512 // prefix. (We save a bit of GC work here by
513 // preallocating based on each hex digit in
514 // ChunkPrefix reducing the expected size of the
515 // filtered set by ~16x.)
516 filtered := make([]arvados.SizedDigest, 0, len(blkids)>>(4*len(bal.ChunkPrefix)-1))
517 for _, blkid := range blkids {
518 if strings.HasPrefix(string(blkid), bal.ChunkPrefix) {
519 filtered = append(filtered, blkid)
524 bal.Logger.Debugf("%v: %d blocks x%d", coll.UUID, len(blkids), repl)
525 // Pass pdh to IncreaseDesired only if LostBlocksFile is being
526 // written -- otherwise it's just a waste of memory.
528 if bal.LostBlocksFile != "" {
529 pdh = coll.PortableDataHash
531 bal.BlockStateMap.IncreaseDesired(pdh, coll.StorageClassesDesired, repl, blkids)
535 // ComputeChangeSets compares, for each known block, the current and
536 // desired replication states. If it is possible to get closer to the
537 // desired state by copying or deleting blocks, it adds those changes
538 // to the relevant KeepServices' ChangeSets.
540 // It does not actually apply any of the computed changes.
541 func (bal *Balancer) ComputeChangeSets() {
542 // This just calls balanceBlock() once for each block, using a
543 // pool of worker goroutines.
544 defer bal.time("changeset_compute", "wall clock time to compute changesets")()
545 bal.setupLookupTables()
547 type balanceTask struct {
548 blkid arvados.SizedDigest
551 workers := runtime.GOMAXPROCS(-1)
552 todo := make(chan balanceTask, workers)
554 bal.BlockStateMap.Apply(func(blkid arvados.SizedDigest, blk *BlockState) {
562 results := make(chan balanceResult, workers)
564 var wg sync.WaitGroup
565 for i := 0; i < workers; i++ {
568 for work := range todo {
569 results <- bal.balanceBlock(work.blkid, work.blk)
577 bal.collectStatistics(results)
580 func (bal *Balancer) setupLookupTables() {
581 bal.serviceRoots = make(map[string]string)
582 bal.classes = defaultClasses
583 bal.mountsByClass = map[string]map[*KeepMount]bool{"default": {}}
585 for _, srv := range bal.KeepServices {
586 bal.serviceRoots[srv.UUID] = srv.UUID
587 for _, mnt := range srv.mounts {
591 // All mounts on a read-only service
592 // are effectively read-only.
593 mnt.AllowWrite = false
596 for class := range mnt.StorageClasses {
597 if mbc := bal.mountsByClass[class]; mbc == nil {
598 bal.classes = append(bal.classes, class)
599 bal.mountsByClass[class] = map[*KeepMount]bool{mnt: true}
606 // Consider classes in lexicographic order to avoid flapping
607 // between balancing runs. The outcome of the "prefer a mount
608 // we're already planning to use for a different storage
609 // class" case in balanceBlock depends on the order classes
611 sort.Strings(bal.classes)
621 var changeName = map[int]string{
624 changeTrash: "trash",
628 type balancedBlockState struct {
635 type balanceResult struct {
637 blkid arvados.SizedDigest
639 blockState balancedBlockState
640 classState map[string]balancedBlockState
644 mnt *KeepMount // never nil
645 repl *Replica // replica already stored here (or nil)
646 want bool // we should pull/leave a replica here
649 // balanceBlock compares current state to desired state for a single
650 // block, and makes the appropriate ChangeSet calls.
651 func (bal *Balancer) balanceBlock(blkid arvados.SizedDigest, blk *BlockState) balanceResult {
652 bal.Logger.Debugf("balanceBlock: %v %+v", blkid, blk)
654 // Build a list of all slots (one per mounted volume).
655 slots := make([]slot, 0, bal.mounts)
656 for _, srv := range bal.KeepServices {
657 for _, mnt := range srv.mounts {
659 for r := range blk.Replicas {
660 if blk.Replicas[r].KeepMount == mnt {
661 repl = &blk.Replicas[r]
664 // Initial value of "want" is "have, and can't
665 // delete". These untrashable replicas get
666 // prioritized when sorting slots: otherwise,
667 // non-optimal readonly copies would cause us
669 slots = append(slots, slot{
672 want: repl != nil && !mnt.AllowTrash,
677 uuids := keepclient.NewRootSorter(bal.serviceRoots, string(blkid[:32])).GetSortedRoots()
678 srvRendezvous := make(map[*KeepService]int, len(uuids))
679 for i, uuid := range uuids {
680 srv := bal.KeepServices[uuid]
681 srvRendezvous[srv] = i
684 // Below we set underreplicated=true if we find any storage
685 // class that's currently underreplicated -- in that case we
686 // won't want to trash any replicas.
687 underreplicated := false
689 unsafeToDelete := make(map[int64]bool, len(slots))
690 for _, class := range bal.classes {
691 desired := blk.Desired[class]
696 // Sort the slots by desirability.
697 sort.Slice(slots, func(i, j int) bool {
698 si, sj := slots[i], slots[j]
699 if classi, classj := bal.mountsByClass[class][si.mnt], bal.mountsByClass[class][sj.mnt]; classi != classj {
700 // Prefer a mount that satisfies the
702 return bal.mountsByClass[class][si.mnt]
703 } else if si.want != sj.want {
704 // Prefer a mount that will have a
705 // replica no matter what we do here
706 // -- either because it already has an
707 // untrashable replica, or because we
708 // already need it to satisfy a
709 // different storage class.
711 } else if orderi, orderj := srvRendezvous[si.mnt.KeepService], srvRendezvous[sj.mnt.KeepService]; orderi != orderj {
712 // Prefer a better rendezvous
714 return orderi < orderj
715 } else if repli, replj := si.repl != nil, sj.repl != nil; repli != replj {
716 // Prefer a mount that already has a
720 // If pull/trash turns out to be
721 // needed, distribute the
722 // new/remaining replicas uniformly
723 // across qualifying mounts on a given
725 return rendezvousLess(si.mnt.UUID, sj.mnt.UUID, blkid)
729 // Servers/mounts/devices (with or without existing
730 // replicas) that are part of the best achievable
731 // layout for this storage class.
732 wantSrv := map[*KeepService]bool{}
733 wantMnt := map[*KeepMount]bool{}
734 wantDev := map[string]bool{}
735 // Positions (with existing replicas) that have been
736 // protected (via unsafeToDelete) to ensure we don't
737 // reduce replication below desired level when
738 // trashing replicas that aren't optimal positions for
739 // any storage class.
740 protMnt := map[*KeepMount]bool{}
741 // Replication planned so far (corresponds to wantMnt).
743 // Protected replication (corresponds to protMnt).
746 // trySlot tries using a slot to meet requirements,
747 // and returns true if all requirements are met.
748 trySlot := func(i int) bool {
750 if wantMnt[slot.mnt] || wantDev[slot.mnt.UUID] {
751 // Already allocated a replica to this
752 // backend device, possibly on a
756 if replProt < desired && slot.repl != nil && !protMnt[slot.mnt] {
757 unsafeToDelete[slot.repl.Mtime] = true
758 protMnt[slot.mnt] = true
759 replProt += slot.mnt.Replication
761 if replWant < desired && (slot.repl != nil || slot.mnt.AllowWrite) {
763 wantSrv[slot.mnt.KeepService] = true
764 wantMnt[slot.mnt] = true
765 wantDev[slot.mnt.UUID] = true
766 replWant += slot.mnt.Replication
768 return replProt >= desired && replWant >= desired
771 // First try to achieve desired replication without
772 // using the same server twice.
774 for i := 0; i < len(slots) && !done; i++ {
775 if !wantSrv[slots[i].mnt.KeepService] {
780 // If that didn't suffice, do another pass without the
781 // "distinct services" restriction. (Achieving the
782 // desired volume replication on fewer than the
783 // desired number of services is better than
784 // underreplicating.)
785 for i := 0; i < len(slots) && !done; i++ {
789 if !underreplicated {
791 for _, slot := range slots {
792 if slot.repl == nil || !bal.mountsByClass[class][slot.mnt] {
795 if safe += slot.mnt.Replication; safe >= desired {
799 underreplicated = safe < desired
802 // Avoid deleting wanted replicas from devices that
803 // are mounted on multiple servers -- even if they
804 // haven't already been added to unsafeToDelete
805 // because the servers report different Mtimes.
806 for _, slot := range slots {
807 if slot.repl != nil && wantDev[slot.mnt.UUID] {
808 unsafeToDelete[slot.repl.Mtime] = true
813 // TODO: If multiple replicas are trashable, prefer the oldest
814 // replica that doesn't have a timestamp collision with
817 for i, slot := range slots {
818 // Don't trash (1) any replicas of an underreplicated
819 // block, even if they're in the wrong positions, or
820 // (2) any replicas whose Mtimes are identical to
821 // needed replicas (in case we're really seeing the
822 // same copy via different mounts).
823 if slot.repl != nil && (underreplicated || unsafeToDelete[slot.repl.Mtime]) {
828 classState := make(map[string]balancedBlockState, len(bal.classes))
829 for _, class := range bal.classes {
830 classState[class] = computeBlockState(slots, bal.mountsByClass[class], len(blk.Replicas), blk.Desired[class])
832 blockState := computeBlockState(slots, nil, len(blk.Replicas), 0)
834 // Sort the slots by rendezvous order. This ensures "trash the
835 // first of N replicas with identical timestamps" is
836 // predictable (helpful for testing) and well distributed
838 sort.Slice(slots, func(i, j int) bool {
839 si, sj := slots[i], slots[j]
840 if orderi, orderj := srvRendezvous[si.mnt.KeepService], srvRendezvous[sj.mnt.KeepService]; orderi != orderj {
841 return orderi < orderj
843 return rendezvousLess(si.mnt.UUID, sj.mnt.UUID, blkid)
850 trashedMtime = make(map[int64]bool, len(slots))
852 for _, slot := range slots {
853 // TODO: request a Touch if Mtime is duplicated.
856 case !slot.want && slot.repl != nil && slot.repl.Mtime < bal.MinMtime:
857 if trashedMtime[slot.repl.Mtime] {
858 // Don't trash multiple replicas with
859 // identical timestamps. If they are
860 // multiple views of the same backing
861 // storage, asking both servers to
862 // trash is redundant and can cause
863 // races (see #20242). If they are
864 // distinct replicas that happen to
865 // have identical timestamps, we'll
866 // get this one on the next sweep.
869 slot.mnt.KeepService.AddTrash(Trash{
871 Mtime: slot.repl.Mtime,
875 trashedMtime[slot.repl.Mtime] = true
877 case slot.repl == nil && slot.want && len(blk.Replicas) == 0:
880 case slot.repl == nil && slot.want && slot.mnt.AllowWrite:
881 slot.mnt.KeepService.AddPull(Pull{
883 From: blk.Replicas[0].KeepMount.KeepService,
887 case slot.repl != nil:
892 if bal.Dumper != nil {
894 if slot.repl != nil {
895 mtime = slot.repl.Mtime
897 srv := slot.mnt.KeepService
898 changes = append(changes, fmt.Sprintf("%s:%d/%s=%s,%d", srv.ServiceHost, srv.ServicePort, slot.mnt.UUID, changeName[change], mtime))
901 if bal.Dumper != nil {
902 bal.Dumper.Printf("%s refs=%d needed=%d unneeded=%d pulling=%v %v %v", blkid, blk.RefCount, blockState.needed, blockState.unneeded, blockState.pulling, blk.Desired, changes)
904 return balanceResult{
908 blockState: blockState,
909 classState: classState,
913 func computeBlockState(slots []slot, onlyCount map[*KeepMount]bool, have, needRepl int) (bbs balancedBlockState) {
915 countedDev := map[string]bool{}
916 for _, slot := range slots {
917 if onlyCount != nil && !onlyCount[slot.mnt] {
920 if countedDev[slot.mnt.UUID] {
924 case slot.repl != nil && slot.want:
926 repl += slot.mnt.Replication
927 case slot.repl != nil && !slot.want:
929 repl += slot.mnt.Replication
930 case slot.repl == nil && slot.want && have > 0:
932 repl += slot.mnt.Replication
934 countedDev[slot.mnt.UUID] = true
937 bbs.unachievable = true
942 type blocksNBytes struct {
948 func (bb blocksNBytes) String() string {
949 return fmt.Sprintf("%d replicas (%d blocks, %d bytes)", bb.replicas, bb.blocks, bb.bytes)
952 type replicationStats struct {
954 unneeded blocksNBytes
956 unachievable blocksNBytes
959 type balancerStats struct {
964 underrep blocksNBytes
965 unachievable blocksNBytes
966 justright blocksNBytes
972 classStats map[string]replicationStats
974 // collectionBytes / collectionBlockBytes = deduplication ratio
975 collectionBytes int64 // sum(bytes in referenced blocks) across all collections
976 collectionBlockBytes int64 // sum(block size) across all blocks referenced by collections
977 collectionBlockRefs int64 // sum(number of blocks referenced) across all collections
978 collectionBlocks int64 // number of blocks referenced by any collection
981 func (s *balancerStats) dedupByteRatio() float64 {
982 if s.collectionBlockBytes == 0 {
985 return float64(s.collectionBytes) / float64(s.collectionBlockBytes)
988 func (s *balancerStats) dedupBlockRatio() float64 {
989 if s.collectionBlocks == 0 {
992 return float64(s.collectionBlockRefs) / float64(s.collectionBlocks)
995 func (bal *Balancer) collectStatistics(results <-chan balanceResult) {
997 s.replHistogram = make([]int, 2)
998 s.classStats = make(map[string]replicationStats, len(bal.classes))
999 for result := range results {
1000 bytes := result.blkid.Size()
1002 if rc := int64(result.blk.RefCount); rc > 0 {
1003 s.collectionBytes += rc * bytes
1004 s.collectionBlockBytes += bytes
1005 s.collectionBlockRefs += rc
1006 s.collectionBlocks++
1009 for class, state := range result.classState {
1010 cs := s.classStats[class]
1011 if state.unachievable {
1012 cs.unachievable.replicas++
1013 cs.unachievable.blocks++
1014 cs.unachievable.bytes += bytes
1016 if state.needed > 0 {
1017 cs.needed.replicas += state.needed
1019 cs.needed.bytes += bytes * int64(state.needed)
1021 if state.unneeded > 0 {
1022 cs.unneeded.replicas += state.unneeded
1023 cs.unneeded.blocks++
1024 cs.unneeded.bytes += bytes * int64(state.unneeded)
1026 if state.pulling > 0 {
1027 cs.pulling.replicas += state.pulling
1029 cs.pulling.bytes += bytes * int64(state.pulling)
1031 s.classStats[class] = cs
1034 bs := result.blockState
1039 s.lost.bytes += bytes
1040 fmt.Fprintf(bal.lostBlocks, "%s", strings.SplitN(string(result.blkid), "+", 2)[0])
1041 for pdh := range result.blk.Refs {
1042 fmt.Fprintf(bal.lostBlocks, " %s", pdh)
1044 fmt.Fprint(bal.lostBlocks, "\n")
1045 case bs.pulling > 0:
1046 s.underrep.replicas += bs.pulling
1048 s.underrep.bytes += bytes * int64(bs.pulling)
1049 case bs.unachievable:
1050 s.underrep.replicas++
1052 s.underrep.bytes += bytes
1053 case bs.unneeded > 0 && bs.needed == 0:
1054 // Count as "garbage" if all replicas are old
1055 // enough to trash, otherwise count as
1057 counter := &s.garbage
1058 for _, r := range result.blk.Replicas {
1059 if r.Mtime >= bal.MinMtime {
1064 counter.replicas += bs.unneeded
1066 counter.bytes += bytes * int64(bs.unneeded)
1067 case bs.unneeded > 0:
1068 s.overrep.replicas += bs.unneeded
1070 s.overrep.bytes += bytes * int64(bs.unneeded)
1072 s.justright.replicas += bs.needed
1073 s.justright.blocks++
1074 s.justright.bytes += bytes * int64(bs.needed)
1078 s.desired.replicas += bs.needed
1080 s.desired.bytes += bytes * int64(bs.needed)
1082 if bs.needed+bs.unneeded > 0 {
1083 s.current.replicas += bs.needed + bs.unneeded
1085 s.current.bytes += bytes * int64(bs.needed+bs.unneeded)
1088 for len(s.replHistogram) <= bs.needed+bs.unneeded {
1089 s.replHistogram = append(s.replHistogram, 0)
1091 s.replHistogram[bs.needed+bs.unneeded]++
1093 for _, srv := range bal.KeepServices {
1094 s.pulls += len(srv.ChangeSet.Pulls)
1095 s.trashes += len(srv.ChangeSet.Trashes)
1098 bal.Metrics.UpdateStats(s)
1101 // PrintStatistics writes statistics about the computed changes to
1102 // bal.Logger. It should not be called until ComputeChangeSets has
1104 func (bal *Balancer) PrintStatistics() {
1106 bal.logf("%s lost (0=have<want)", bal.stats.lost)
1107 bal.logf("%s underreplicated (0<have<want)", bal.stats.underrep)
1108 bal.logf("%s just right (have=want)", bal.stats.justright)
1109 bal.logf("%s overreplicated (have>want>0)", bal.stats.overrep)
1110 bal.logf("%s unreferenced (have>want=0, new)", bal.stats.unref)
1111 bal.logf("%s garbage (have>want=0, old)", bal.stats.garbage)
1112 for _, class := range bal.classes {
1113 cs := bal.stats.classStats[class]
1115 bal.logf("storage class %q: %s needed", class, cs.needed)
1116 bal.logf("storage class %q: %s unneeded", class, cs.unneeded)
1117 bal.logf("storage class %q: %s pulling", class, cs.pulling)
1118 bal.logf("storage class %q: %s unachievable", class, cs.unachievable)
1121 bal.logf("%s total commitment (excluding unreferenced)", bal.stats.desired)
1122 bal.logf("%s total usage", bal.stats.current)
1124 for _, srv := range bal.KeepServices {
1125 bal.logf("%s: %v\n", srv, srv.ChangeSet)
1128 bal.printHistogram(60)
1132 func (bal *Balancer) printHistogram(hashColumns int) {
1133 bal.logf("Replication level distribution:")
1135 for _, count := range bal.stats.replHistogram {
1136 if maxCount < count {
1140 hashes := strings.Repeat("#", hashColumns)
1141 countWidth := 1 + int(math.Log10(float64(maxCount+1)))
1142 scaleCount := 10 * float64(hashColumns) / math.Floor(1+10*math.Log10(float64(maxCount+1)))
1143 for repl, count := range bal.stats.replHistogram {
1144 nHashes := int(scaleCount * math.Log10(float64(count+1)))
1145 bal.logf("%2d: %*d %s", repl, countWidth, count, hashes[:nHashes])
1149 // CheckSanityLate checks for configuration and runtime errors after
1150 // GetCurrentState() and ComputeChangeSets() have finished.
1152 // If it returns an error, it is dangerous to run any Commit methods.
1153 func (bal *Balancer) CheckSanityLate() error {
1154 if bal.errors != nil {
1155 for _, err := range bal.errors {
1156 bal.logf("deferred error: %v", err)
1158 return fmt.Errorf("cannot proceed safely after deferred errors")
1161 if bal.collScanned == 0 {
1162 return fmt.Errorf("received zero collections")
1166 bal.BlockStateMap.Apply(func(_ arvados.SizedDigest, blk *BlockState) {
1167 for _, desired := range blk.Desired {
1175 return fmt.Errorf("zero blocks have desired replication>0")
1178 if dr := bal.DefaultReplication; dr < 1 {
1179 return fmt.Errorf("Default replication (%d) is less than 1", dr)
1182 // TODO: no two services have identical indexes
1183 // TODO: no collisions (same md5, different size)
1187 // CommitPulls sends the computed lists of pull requests to the
1188 // keepstore servers. This has the effect of increasing replication of
1189 // existing blocks that are either underreplicated or poorly
1190 // distributed according to rendezvous hashing.
1191 func (bal *Balancer) CommitPulls(ctx context.Context, c *arvados.Client) error {
1192 defer bal.time("send_pull_lists", "wall clock time to send pull lists")()
1193 return bal.commitAsync(c, "send pull list",
1194 func(srv *KeepService) error {
1195 return srv.CommitPulls(ctx, c)
1199 // CommitTrash sends the computed lists of trash requests to the
1200 // keepstore servers. This has the effect of deleting blocks that are
1201 // overreplicated or unreferenced.
1202 func (bal *Balancer) CommitTrash(ctx context.Context, c *arvados.Client) error {
1203 defer bal.time("send_trash_lists", "wall clock time to send trash lists")()
1204 return bal.commitAsync(c, "send trash list",
1205 func(srv *KeepService) error {
1206 return srv.CommitTrash(ctx, c)
1210 func (bal *Balancer) commitAsync(c *arvados.Client, label string, f func(srv *KeepService) error) error {
1211 errs := make(chan error)
1212 for _, srv := range bal.KeepServices {
1213 go func(srv *KeepService) {
1215 defer func() { errs <- err }()
1216 label := fmt.Sprintf("%s: %v", srv, label)
1219 err = fmt.Errorf("%s: %v", label, err)
1224 for range bal.KeepServices {
1225 if err := <-errs; err != nil {
1234 func (bal *Balancer) logf(f string, args ...interface{}) {
1235 if bal.Logger != nil {
1236 bal.Logger.Printf(f, args...)
1240 func (bal *Balancer) time(name, help string) func() {
1241 observer := bal.Metrics.DurationObserver(name+"_seconds", help)
1243 bal.Logger.Printf("%s: start", name)
1245 dur := time.Since(t0)
1246 observer.Observe(dur.Seconds())
1247 bal.Logger.Printf("%s: took %vs", name, dur.Seconds())
1251 // Log current memory usage: once now, at least once every 10 minutes,
1252 // and when memory grows by 40% since the last log. Stop when ctx is
1254 func (bal *Balancer) reportMemorySize(ctx context.Context) {
1255 buf, _ := os.ReadFile("/proc/self/smaps")
1256 m := regexp.MustCompile(`\nKernelPageSize:\s*(\d+) kB\n`).FindSubmatch(buf)
1259 pagesize, _ = strconv.ParseInt(string(m[1]), 10, 64)
1263 bal.logf("cannot log OS-reported memory size: failed to parse KernelPageSize from /proc/self/smaps")
1265 osstats := func() string {
1269 buf, _ := os.ReadFile("/proc/self/statm")
1270 fields := strings.Split(string(buf), " ")
1271 if len(fields) < 2 {
1274 virt, _ := strconv.ParseInt(fields[0], 10, 64)
1276 res, _ := strconv.ParseInt(fields[1], 10, 64)
1278 if virt == 0 || res == 0 {
1281 return fmt.Sprintf(" virt %d res %d", virt, res)
1284 var nextTime time.Time
1286 const maxInterval = time.Minute * 10
1287 const maxIncrease = 1.4
1289 ticker := time.NewTicker(time.Second)
1291 var memstats runtime.MemStats
1292 for ctx.Err() == nil {
1294 runtime.ReadMemStats(&memstats)
1295 mem := memstats.StackInuse + memstats.HeapInuse
1296 if now.After(nextTime) || mem >= nextMem {
1297 bal.logf("heap %d stack %d heapalloc %d%s", memstats.HeapInuse, memstats.StackInuse, memstats.HeapAlloc, osstats())
1298 nextMem = uint64(float64(mem) * maxIncrease)
1299 nextTime = now.Add(maxInterval)
1305 // Rendezvous hash sort function. Less efficient than sorting on
1306 // precomputed rendezvous hashes, but also rarely used.
1307 func rendezvousLess(i, j string, blkid arvados.SizedDigest) bool {
1308 a := md5.Sum([]byte(string(blkid[:32]) + i))
1309 b := md5.Sum([]byte(string(blkid[:32]) + j))
1310 return bytes.Compare(a[:], b[:]) < 0