1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
20 "git.curoverse.com/arvados.git/sdk/go/arvados"
21 "git.curoverse.com/arvados.git/sdk/go/keepclient"
24 // CheckConfig returns an error if anything is wrong with the given
25 // config and runOptions.
26 func CheckConfig(config Config, runOptions RunOptions) error {
27 if len(config.KeepServiceList.Items) > 0 && config.KeepServiceTypes != nil {
28 return fmt.Errorf("cannot specify both KeepServiceList and KeepServiceTypes in config")
30 if !runOptions.Once && config.RunPeriod == arvados.Duration(0) {
31 return fmt.Errorf("you must either use the -once flag, or specify RunPeriod in config")
36 // Balancer compares the contents of keepstore servers with the
37 // collections stored in Arvados, and issues pull/trash requests
38 // needed to get (closer to) the optimal data layout.
40 // In the optimal data layout: every data block referenced by a
41 // collection is replicated at least as many times as desired by the
42 // collection; there are no unreferenced data blocks older than
43 // BlobSignatureTTL; and all N existing replicas of a given data block
44 // are in the N best positions in rendezvous probe order.
45 type Balancer struct {
47 KeepServices map[string]*KeepService
48 DefaultReplication int
55 mountsByClass map[string]map[*KeepMount]bool
57 serviceRoots map[string]string
63 // Run performs a balance operation using the given config and
64 // runOptions, and returns RunOptions suitable for passing to a
65 // subsequent balance operation.
67 // Run should only be called once on a given Balancer object.
71 // runOptions, err = (&Balancer{}).Run(config, runOptions)
72 func (bal *Balancer) Run(config Config, runOptions RunOptions) (nextRunOptions RunOptions, err error) {
73 nextRunOptions = runOptions
75 bal.Dumper = runOptions.Dumper
76 bal.Logger = runOptions.Logger
77 if bal.Logger == nil {
78 bal.Logger = log.New(os.Stderr, "", log.LstdFlags)
81 defer timeMe(bal.Logger, "Run")()
83 if len(config.KeepServiceList.Items) > 0 {
84 err = bal.SetKeepServices(config.KeepServiceList)
86 err = bal.DiscoverKeepServices(&config.Client, config.KeepServiceTypes)
92 for _, srv := range bal.KeepServices {
93 err = srv.discoverMounts(&config.Client)
99 if err = bal.CheckSanityEarly(&config.Client); err != nil {
102 rs := bal.rendezvousState()
103 if runOptions.CommitTrash && rs != runOptions.SafeRendezvousState {
104 if runOptions.SafeRendezvousState != "" {
105 bal.logf("notice: KeepServices list has changed since last run")
107 bal.logf("clearing existing trash lists, in case the new rendezvous order differs from previous run")
108 if err = bal.ClearTrashLists(&config.Client); err != nil {
111 // The current rendezvous state becomes "safe" (i.e.,
112 // OK to compute changes for that state without
113 // clearing existing trash lists) only now, after we
114 // succeed in clearing existing trash lists.
115 nextRunOptions.SafeRendezvousState = rs
117 if err = bal.GetCurrentState(&config.Client, config.CollectionBatchSize, config.CollectionBuffers); err != nil {
120 bal.ComputeChangeSets()
121 bal.PrintStatistics()
122 if err = bal.CheckSanityLate(); err != nil {
125 if runOptions.CommitPulls {
126 err = bal.CommitPulls(&config.Client)
128 // Skip trash if we can't pull. (Too cautious?)
132 if runOptions.CommitTrash {
133 err = bal.CommitTrash(&config.Client)
138 // SetKeepServices sets the list of KeepServices to operate on.
139 func (bal *Balancer) SetKeepServices(srvList arvados.KeepServiceList) error {
140 bal.KeepServices = make(map[string]*KeepService)
141 for _, srv := range srvList.Items {
142 bal.KeepServices[srv.UUID] = &KeepService{
144 ChangeSet: &ChangeSet{},
150 // DiscoverKeepServices sets the list of KeepServices by calling the
151 // API to get a list of all services, and selecting the ones whose
152 // ServiceType is in okTypes.
153 func (bal *Balancer) DiscoverKeepServices(c *arvados.Client, okTypes []string) error {
154 bal.KeepServices = make(map[string]*KeepService)
155 ok := make(map[string]bool)
156 for _, t := range okTypes {
159 return c.EachKeepService(func(srv arvados.KeepService) error {
160 if ok[srv.ServiceType] {
161 bal.KeepServices[srv.UUID] = &KeepService{
163 ChangeSet: &ChangeSet{},
166 bal.logf("skipping %v with service type %q", srv.UUID, srv.ServiceType)
172 // CheckSanityEarly checks for configuration and runtime errors that
173 // can be detected before GetCurrentState() and ComputeChangeSets()
176 // If it returns an error, it is pointless to run GetCurrentState or
177 // ComputeChangeSets: after doing so, the statistics would be
178 // meaningless and it would be dangerous to run any Commit methods.
179 func (bal *Balancer) CheckSanityEarly(c *arvados.Client) error {
180 u, err := c.CurrentUser()
182 return fmt.Errorf("CurrentUser(): %v", err)
184 if !u.IsActive || !u.IsAdmin {
185 return fmt.Errorf("current user (%s) is not an active admin user", u.UUID)
187 for _, srv := range bal.KeepServices {
188 if srv.ServiceType == "proxy" {
189 return fmt.Errorf("config error: %s: proxy servers cannot be balanced", srv)
195 // rendezvousState returns a fingerprint (e.g., a sorted list of
196 // UUID+host+port) of the current set of keep services.
197 func (bal *Balancer) rendezvousState() string {
198 srvs := make([]string, 0, len(bal.KeepServices))
199 for _, srv := range bal.KeepServices {
200 srvs = append(srvs, srv.String())
203 return strings.Join(srvs, "; ")
206 // ClearTrashLists sends an empty trash list to each keep
207 // service. Calling this before GetCurrentState avoids races.
209 // When a block appears in an index, we assume that replica will still
210 // exist after we delete other replicas on other servers. However,
211 // it's possible that a previous rebalancing operation made different
212 // decisions (e.g., servers were added/removed, and rendezvous order
213 // changed). In this case, the replica might already be on that
214 // server's trash list, and it might be deleted before we send a
215 // replacement trash list.
217 // We avoid this problem if we clear all trash lists before getting
218 // indexes. (We also assume there is only one rebalancing process
219 // running at a time.)
220 func (bal *Balancer) ClearTrashLists(c *arvados.Client) error {
221 for _, srv := range bal.KeepServices {
222 srv.ChangeSet = &ChangeSet{}
224 return bal.CommitTrash(c)
227 // GetCurrentState determines the current replication state, and the
228 // desired replication level, for every block that is either
229 // retrievable or referenced.
231 // It determines the current replication state by reading the block index
232 // from every known Keep service.
234 // It determines the desired replication level by retrieving all
235 // collection manifests in the database (API server).
237 // It encodes the resulting information in BlockStateMap.
238 func (bal *Balancer) GetCurrentState(c *arvados.Client, pageSize, bufs int) error {
239 defer timeMe(bal.Logger, "GetCurrentState")()
240 bal.BlockStateMap = NewBlockStateMap()
242 dd, err := c.DiscoveryDocument()
246 bal.DefaultReplication = dd.DefaultCollectionReplication
247 bal.MinMtime = time.Now().UnixNano() - dd.BlobSignatureTTL*1e9
249 errs := make(chan error, 2+len(bal.KeepServices))
250 wg := sync.WaitGroup{}
252 // Start one goroutine for each KeepService: retrieve the
253 // index, and add the returned blocks to BlockStateMap.
254 for _, srv := range bal.KeepServices {
256 go func(srv *KeepService) {
258 bal.logf("%s: retrieve indexes", srv)
259 for _, mount := range srv.mounts {
260 bal.logf("%s: retrieve index", mount)
261 idx, err := srv.IndexMount(c, mount.UUID, "")
263 errs <- fmt.Errorf("%s: retrieve index: %v", mount, err)
267 // Some other goroutine encountered an
268 // error -- any further effort here
272 bal.logf("%s: add %d replicas to map", mount, len(idx))
273 bal.BlockStateMap.AddReplicas(mount, idx)
274 bal.logf("%s: done", mount)
276 bal.logf("%s: done", srv)
280 // collQ buffers incoming collections so we can start fetching
281 // the next page without waiting for the current page to
282 // finish processing.
283 collQ := make(chan arvados.Collection, bufs)
285 // Start a goroutine to process collections. (We could use a
286 // worker pool here, but even with a single worker we already
287 // process collections much faster than we can retrieve them.)
291 for coll := range collQ {
292 err := bal.addCollection(coll)
303 // Start a goroutine to retrieve all collections from the
304 // Arvados database and send them to collQ for processing.
308 err = EachCollection(c, pageSize,
309 func(coll arvados.Collection) error {
312 // some other GetCurrentState
313 // error happened: no point
316 return fmt.Errorf("")
319 }, func(done, total int) {
320 bal.logf("collections: %d/%d", done, total)
335 func (bal *Balancer) addCollection(coll arvados.Collection) error {
336 blkids, err := coll.SizedDigests()
339 bal.errors = append(bal.errors, fmt.Errorf("%v: %v", coll.UUID, err))
343 repl := bal.DefaultReplication
344 if coll.ReplicationDesired != nil {
345 repl = *coll.ReplicationDesired
347 debugf("%v: %d block x%d", coll.UUID, len(blkids), repl)
348 bal.BlockStateMap.IncreaseDesired(coll.StorageClassesDesired, repl, blkids)
352 // ComputeChangeSets compares, for each known block, the current and
353 // desired replication states. If it is possible to get closer to the
354 // desired state by copying or deleting blocks, it adds those changes
355 // to the relevant KeepServices' ChangeSets.
357 // It does not actually apply any of the computed changes.
358 func (bal *Balancer) ComputeChangeSets() {
359 // This just calls balanceBlock() once for each block, using a
360 // pool of worker goroutines.
361 defer timeMe(bal.Logger, "ComputeChangeSets")()
362 bal.setupLookupTables()
364 type balanceTask struct {
365 blkid arvados.SizedDigest
368 workers := runtime.GOMAXPROCS(-1)
369 todo := make(chan balanceTask, workers)
371 bal.BlockStateMap.Apply(func(blkid arvados.SizedDigest, blk *BlockState) {
379 results := make(chan balanceResult, workers)
381 var wg sync.WaitGroup
382 for i := 0; i < workers; i++ {
385 for work := range todo {
386 results <- bal.balanceBlock(work.blkid, work.blk)
394 bal.collectStatistics(results)
397 func (bal *Balancer) setupLookupTables() {
398 bal.serviceRoots = make(map[string]string)
399 bal.classes = []string{"default"}
400 bal.mountsByClass = map[string]map[*KeepMount]bool{"default": {}}
402 for _, srv := range bal.KeepServices {
403 bal.serviceRoots[srv.UUID] = srv.UUID
404 for _, mnt := range srv.mounts {
407 // All mounts on a read-only service are
408 // effectively read-only.
409 mnt.ReadOnly = mnt.ReadOnly || srv.ReadOnly
411 if len(mnt.StorageClasses) == 0 {
412 bal.mountsByClass["default"][mnt] = true
415 for _, class := range mnt.StorageClasses {
416 if mbc := bal.mountsByClass[class]; mbc == nil {
417 bal.classes = append(bal.classes, class)
418 bal.mountsByClass[class] = map[*KeepMount]bool{mnt: true}
425 // Consider classes in lexicographic order to avoid flapping
426 // between balancing runs. The outcome of the "prefer a mount
427 // we're already planning to use for a different storage
428 // class" case in balanceBlock depends on the order classes
430 sort.Strings(bal.classes)
440 var changeName = map[int]string{
443 changeTrash: "trash",
447 type balanceResult struct {
449 blkid arvados.SizedDigest
452 classState map[string]balancedBlockState
455 // balanceBlock compares current state to desired state for a single
456 // block, and makes the appropriate ChangeSet calls.
457 func (bal *Balancer) balanceBlock(blkid arvados.SizedDigest, blk *BlockState) balanceResult {
458 debugf("balanceBlock: %v %+v", blkid, blk)
461 mnt *KeepMount // never nil
462 repl *Replica // replica already stored here (or nil)
463 want bool // we should pull/leave a replica here
466 // Build a list of all slots (one per mounted volume).
467 slots := make([]slot, 0, bal.mounts)
468 for _, srv := range bal.KeepServices {
469 for _, mnt := range srv.mounts {
471 for r := range blk.Replicas {
472 if blk.Replicas[r].KeepMount == mnt {
473 repl = &blk.Replicas[r]
476 // Initial value of "want" is "have, and can't
477 // delete". These untrashable replicas get
478 // prioritized when sorting slots: otherwise,
479 // non-optimal readonly copies would cause us
481 slots = append(slots, slot{
484 want: repl != nil && (mnt.ReadOnly || repl.Mtime >= bal.MinMtime),
489 uuids := keepclient.NewRootSorter(bal.serviceRoots, string(blkid[:32])).GetSortedRoots()
490 srvRendezvous := make(map[*KeepService]int, len(uuids))
491 for i, uuid := range uuids {
492 srv := bal.KeepServices[uuid]
493 srvRendezvous[srv] = i
496 // Below we set underreplicated=true if we find any storage
497 // class that's currently underreplicated -- in that case we
498 // won't want to trash any replicas.
499 underreplicated := false
501 classState := make(map[string]balancedBlockState, len(bal.classes))
502 unsafeToDelete := make(map[int64]bool, len(slots))
503 for _, class := range bal.classes {
504 desired := blk.Desired[class]
507 for _, slot := range slots {
508 if slot.repl != nil && bal.mountsByClass[class][slot.mnt] {
512 classState[class] = balancedBlockState{
514 surplus: have - desired,
521 // Sort the slots by desirability.
522 sort.Slice(slots, func(i, j int) bool {
523 si, sj := slots[i], slots[j]
524 if classi, classj := bal.mountsByClass[class][si.mnt], bal.mountsByClass[class][sj.mnt]; classi != classj {
525 // Prefer a mount that satisfies the
527 return bal.mountsByClass[class][si.mnt]
528 } else if wanti, wantj := si.want, si.want; wanti != wantj {
529 // Prefer a mount that will have a
530 // replica no matter what we do here
531 // -- either because it already has an
532 // untrashable replica, or because we
533 // already need it to satisfy a
534 // different storage class.
536 } else if orderi, orderj := srvRendezvous[si.mnt.KeepService], srvRendezvous[sj.mnt.KeepService]; orderi != orderj {
537 // Prefer a better rendezvous
539 return orderi < orderj
540 } else if repli, replj := si.repl != nil, sj.repl != nil; repli != replj {
541 // Prefer a mount that already has a
545 // If pull/trash turns out to be
546 // needed, distribute the
547 // new/remaining replicas uniformly
548 // across qualifying mounts on a given
550 return rendezvousLess(si.mnt.DeviceID, sj.mnt.DeviceID, blkid)
554 // Servers and mounts (with or without existing
555 // replicas) that are part of the best achievable
556 // layout for this storage class.
557 wantSrv := map[*KeepService]bool{}
558 wantMnt := map[*KeepMount]bool{}
559 // Positions (with existing replicas) that have been
560 // protected (via unsafeToDelete) to ensure we don't
561 // reduce replication below desired level when
562 // trashing replicas that aren't optimal positions for
563 // any storage class.
564 protMnt := map[*KeepMount]bool{}
566 // trySlot tries using a slot to meet requirements,
567 // and returns true if all requirements are met.
568 trySlot := func(i int) bool {
570 if len(protMnt) < desired && slot.repl != nil {
571 unsafeToDelete[slot.repl.Mtime] = true
572 protMnt[slot.mnt] = true
574 if len(wantMnt) < desired && (slot.repl != nil || !slot.mnt.ReadOnly) {
576 wantSrv[slot.mnt.KeepService] = true
577 wantMnt[slot.mnt] = true
579 return len(protMnt) >= desired && len(wantMnt) >= desired
582 // First try to achieve desired replication without
583 // using the same server twice.
585 for i := 0; i < len(slots) && !done; i++ {
586 if !wantSrv[slots[i].mnt.KeepService] {
591 // If that didn't suffice, do another pass without the
592 // "distinct services" restriction. (Achieving the
593 // desired volume replication on fewer than the
594 // desired number of services is better than
595 // underreplicating.)
596 for i := 0; i < len(slots) && !done; i++ {
600 if !underreplicated {
602 for _, slot := range slots {
603 if slot.repl == nil || !bal.mountsByClass[class][slot.mnt] {
606 if safe++; safe >= desired {
610 underreplicated = safe < desired
613 // set the unachievable flag if there aren't enough
614 // slots offering the relevant storage class. (This is
615 // as easy as checking slots[desired] because we
616 // already sorted the qualifying slots to the front.)
617 if desired >= len(slots) || !bal.mountsByClass[class][slots[desired].mnt] {
618 cs := classState[class]
619 cs.unachievable = true
620 classState[class] = cs
624 // TODO: If multiple replicas are trashable, prefer the oldest
625 // replica that doesn't have a timestamp collision with
629 for _, slot := range slots {
633 if slot.repl != nil {
639 for _, slot := range slots {
640 // TODO: request a Touch if Mtime is duplicated.
643 case !underreplicated && slot.repl != nil && !slot.want && !unsafeToDelete[slot.repl.Mtime]:
644 slot.mnt.KeepService.AddTrash(Trash{
646 Mtime: slot.repl.Mtime,
650 case len(blk.Replicas) == 0:
652 case slot.repl == nil && slot.want && !slot.mnt.ReadOnly:
653 slot.mnt.KeepService.AddPull(Pull{
655 From: blk.Replicas[0].KeepMount.KeepService,
662 if bal.Dumper != nil {
664 if slot.repl != nil {
665 mtime = slot.repl.Mtime
667 srv := slot.mnt.KeepService
668 changes = append(changes, fmt.Sprintf("%s:%d/%s=%s,%d", srv.ServiceHost, srv.ServicePort, slot.mnt.UUID, changeName[change], mtime))
671 if bal.Dumper != nil {
672 bal.Dumper.Printf("%s have=%d want=%v %s", blkid, have, want, strings.Join(changes, " "))
674 return balanceResult{
679 classState: classState,
683 type blocksNBytes struct {
689 func (bb blocksNBytes) String() string {
690 return fmt.Sprintf("%d replicas (%d blocks, %d bytes)", bb.replicas, bb.blocks, bb.bytes)
693 type balancerStats struct {
698 underrep blocksNBytes
699 unachievable blocksNBytes
700 justright blocksNBytes
706 classStats map[string]replicationStats
709 type replicationStats struct {
713 unachievable blocksNBytes
716 type balancedBlockState struct {
722 func (bal *Balancer) collectStatistics(results <-chan balanceResult) {
724 s.replHistogram = make([]int, 2)
725 s.classStats = make(map[string]replicationStats, len(bal.classes))
726 for result := range results {
727 surplus := result.have - result.want
728 bytes := result.blkid.Size()
730 for class, state := range result.classState {
731 cs := s.classStats[class]
732 if state.unachievable {
733 cs.unachievable.blocks++
734 cs.unachievable.bytes += bytes
736 if state.desired > 0 {
737 cs.desired.replicas += state.desired
739 cs.desired.bytes += bytes * int64(state.desired)
741 if state.surplus > 0 {
742 cs.surplus.replicas += state.surplus
744 cs.surplus.bytes += bytes * int64(state.surplus)
745 } else if state.surplus < 0 {
746 cs.short.replicas += -state.surplus
748 cs.short.bytes += bytes * int64(-state.surplus)
750 s.classStats[class] = cs
754 case result.have == 0 && result.want > 0:
755 s.lost.replicas -= surplus
757 s.lost.bytes += bytes * int64(-surplus)
759 s.underrep.replicas -= surplus
761 s.underrep.bytes += bytes * int64(-surplus)
762 case surplus > 0 && result.want == 0:
763 counter := &s.garbage
764 for _, r := range result.blk.Replicas {
765 if r.Mtime >= bal.MinMtime {
770 counter.replicas += surplus
772 counter.bytes += bytes * int64(surplus)
774 s.overrep.replicas += surplus
776 s.overrep.bytes += bytes * int64(len(result.blk.Replicas)-result.want)
778 s.justright.replicas += result.want
780 s.justright.bytes += bytes * int64(result.want)
784 s.desired.replicas += result.want
786 s.desired.bytes += bytes * int64(result.want)
788 if len(result.blk.Replicas) > 0 {
789 s.current.replicas += len(result.blk.Replicas)
791 s.current.bytes += bytes * int64(len(result.blk.Replicas))
794 for len(s.replHistogram) <= len(result.blk.Replicas) {
795 s.replHistogram = append(s.replHistogram, 0)
797 s.replHistogram[len(result.blk.Replicas)]++
799 for _, srv := range bal.KeepServices {
800 s.pulls += len(srv.ChangeSet.Pulls)
801 s.trashes += len(srv.ChangeSet.Trashes)
806 // PrintStatistics writes statistics about the computed changes to
807 // bal.Logger. It should not be called until ComputeChangeSets has
809 func (bal *Balancer) PrintStatistics() {
811 bal.logf("%s lost (0=have<want)", bal.stats.lost)
812 bal.logf("%s underreplicated (0<have<want)", bal.stats.underrep)
813 bal.logf("%s just right (have=want)", bal.stats.justright)
814 bal.logf("%s overreplicated (have>want>0)", bal.stats.overrep)
815 bal.logf("%s unreferenced (have>want=0, new)", bal.stats.unref)
816 bal.logf("%s garbage (have>want=0, old)", bal.stats.garbage)
817 for _, class := range bal.classes {
818 cs := bal.stats.classStats[class]
820 bal.logf("storage class %q: %s desired", class, cs.desired)
821 bal.logf("storage class %q: %s short", class, cs.short)
822 bal.logf("storage class %q: %s surplus", class, cs.surplus)
823 bal.logf("storage class %q: %s unachievable", class, cs.unachievable)
826 bal.logf("%s total commitment (excluding unreferenced)", bal.stats.desired)
827 bal.logf("%s total usage", bal.stats.current)
829 for _, srv := range bal.KeepServices {
830 bal.logf("%s: %v\n", srv, srv.ChangeSet)
833 bal.printHistogram(60)
837 func (bal *Balancer) printHistogram(hashColumns int) {
838 bal.logf("Replication level distribution (counting N replicas on a single server as N):")
840 for _, count := range bal.stats.replHistogram {
841 if maxCount < count {
845 hashes := strings.Repeat("#", hashColumns)
846 countWidth := 1 + int(math.Log10(float64(maxCount+1)))
847 scaleCount := 10 * float64(hashColumns) / math.Floor(1+10*math.Log10(float64(maxCount+1)))
848 for repl, count := range bal.stats.replHistogram {
849 nHashes := int(scaleCount * math.Log10(float64(count+1)))
850 bal.logf("%2d: %*d %s", repl, countWidth, count, hashes[:nHashes])
854 // CheckSanityLate checks for configuration and runtime errors after
855 // GetCurrentState() and ComputeChangeSets() have finished.
857 // If it returns an error, it is dangerous to run any Commit methods.
858 func (bal *Balancer) CheckSanityLate() error {
859 if bal.errors != nil {
860 for _, err := range bal.errors {
861 bal.logf("deferred error: %v", err)
863 return fmt.Errorf("cannot proceed safely after deferred errors")
866 if bal.collScanned == 0 {
867 return fmt.Errorf("received zero collections")
871 bal.BlockStateMap.Apply(func(_ arvados.SizedDigest, blk *BlockState) {
872 for _, desired := range blk.Desired {
880 return fmt.Errorf("zero blocks have desired replication>0")
883 if dr := bal.DefaultReplication; dr < 1 {
884 return fmt.Errorf("Default replication (%d) is less than 1", dr)
887 // TODO: no two services have identical indexes
888 // TODO: no collisions (same md5, different size)
892 // CommitPulls sends the computed lists of pull requests to the
893 // keepstore servers. This has the effect of increasing replication of
894 // existing blocks that are either underreplicated or poorly
895 // distributed according to rendezvous hashing.
896 func (bal *Balancer) CommitPulls(c *arvados.Client) error {
897 return bal.commitAsync(c, "send pull list",
898 func(srv *KeepService) error {
899 return srv.CommitPulls(c)
903 // CommitTrash sends the computed lists of trash requests to the
904 // keepstore servers. This has the effect of deleting blocks that are
905 // overreplicated or unreferenced.
906 func (bal *Balancer) CommitTrash(c *arvados.Client) error {
907 return bal.commitAsync(c, "send trash list",
908 func(srv *KeepService) error {
909 return srv.CommitTrash(c)
913 func (bal *Balancer) commitAsync(c *arvados.Client, label string, f func(srv *KeepService) error) error {
914 errs := make(chan error)
915 for _, srv := range bal.KeepServices {
916 go func(srv *KeepService) {
918 defer func() { errs <- err }()
919 label := fmt.Sprintf("%s: %v", srv, label)
920 defer timeMe(bal.Logger, label)()
923 err = fmt.Errorf("%s: %v", label, err)
928 for range bal.KeepServices {
929 if err := <-errs; err != nil {
938 func (bal *Balancer) logf(f string, args ...interface{}) {
939 if bal.Logger != nil {
940 bal.Logger.Printf(f, args...)
944 // Rendezvous hash sort function. Less efficient than sorting on
945 // precomputed rendezvous hashes, but also rarely used.
946 func rendezvousLess(i, j string, blkid arvados.SizedDigest) bool {
947 a := md5.Sum([]byte(string(blkid[:32]) + i))
948 b := md5.Sum([]byte(string(blkid[:32]) + j))
949 return bytes.Compare(a[:], b[:]) < 0