storj/satellite/repair/checker/checker.go

587 lines
21 KiB
Go
Raw Normal View History

2019-01-24 20:15:10 +00:00
// Copyright (C) 2019 Storj Labs, Inc.
// See LICENSE for copying information.
package checker
import (
"context"
"time"
"github.com/spacemonkeygo/monkit/v3"
2019-01-23 19:58:44 +00:00
"github.com/zeebo/errs"
"go.uber.org/zap"
"golang.org/x/sync/errgroup"
"storj.io/common/errs2"
"storj.io/common/pb"
"storj.io/common/storj"
"storj.io/common/sync2"
"storj.io/storj/satellite/internalpb"
"storj.io/storj/satellite/metabase"
"storj.io/storj/satellite/metabase/metaloop"
"storj.io/storj/satellite/metainfo"
"storj.io/storj/satellite/overlay"
"storj.io/storj/satellite/repair"
"storj.io/storj/satellite/repair/irreparable"
"storj.io/storj/satellite/repair/queue"
)
2019-01-23 19:58:44 +00:00
// Error is a standard error class for this package.
var (
Error = errs.Class("repair checker")
2019-01-23 19:58:44 +00:00
mon = monkit.Package()
)
// Checker contains the information needed to do checks for missing pieces.
2019-09-10 14:24:16 +01:00
//
// architecture: Chore
type Checker struct {
logger *zap.Logger
repairQueue queue.RepairQueue
irrdb irreparable.DB
metabase metainfo.MetabaseDB
metaLoop *metaloop.Service
nodestate *ReliabilityCache
statsCollector *statsCollector
repairOverrides RepairOverridesMap
nodeFailureRate float64
Loop *sync2.Cycle
IrreparableLoop *sync2.Cycle
}
// NewChecker creates a new instance of checker.
func NewChecker(logger *zap.Logger, repairQueue queue.RepairQueue, irrdb irreparable.DB, metabase metainfo.MetabaseDB, metaLoop *metaloop.Service, overlay *overlay.Service, config Config) *Checker {
return &Checker{
logger: logger,
repairQueue: repairQueue,
irrdb: irrdb,
metabase: metabase,
metaLoop: metaLoop,
nodestate: NewReliabilityCache(overlay, config.ReliabilityCacheStaleness),
statsCollector: newStatsCollector(),
repairOverrides: config.RepairOverrides.GetMap(),
nodeFailureRate: config.NodeFailureRate,
Loop: sync2.NewCycle(config.Interval),
IrreparableLoop: sync2.NewCycle(config.IrreparableInterval),
}
}
// Run the checker loop.
func (checker *Checker) Run(ctx context.Context) (err error) {
defer mon.Task()(&ctx)(&err)
group, ctx := errgroup.WithContext(ctx)
group.Go(func() error {
return checker.Loop.Run(ctx, checker.IdentifyInjuredSegments)
})
group.Go(func() error {
return checker.IrreparableLoop.Run(ctx, checker.IrreparableProcess)
})
return group.Wait()
}
satellite/repair: use survivability model for segment health The chief segment health models we've come up with are the "immediate danger" model and the "survivability" model. The former calculates the chance of losing a segment becoming lost in the next time period (using the CDF of the binomial distribution to estimate the chance of x nodes failing in that period), while the latter estimates the number of iterations for which a segment can be expected to survive (using the mean of the negative binomial distribution). The immediate danger model was a promising one for comparing segment health across segments with different RS parameters, as it is more precisely what we want to prevent, but it turns out that practically all segments in production have infinite health, as the chance of losing segments with any reasonable estimate of node failure rate is smaller than DBL_EPSILON, the smallest possible difference from 1.0 representable in a float64 (about 1e-16). Leaving aside the wisdom of worrying about the repair of segments that have less than a 1e-16 chance of being lost, we want to be extremely conservative and proactive in our repair efforts, and the health of the segments we have been repairing thus far also evaluates to infinity under the immediate danger model. Thus, we find ourselves reaching for an alternative. Dr. Ben saves the day: the survivability model is a reasonably close approximation of the immediate danger model, and even better, it is far simpler to calculate and yields manageable values for real-world segments. The downside to it is that it requires as input an estimate of the total number of active nodes. This change replaces the segment health calculation to use the survivability model, and reinstates the call to SegmentHealth() where it was reverted. It gets estimates for the total number of active nodes by leveraging the reliability cache. Change-Id: Ia5d9b9031b9f6cf0fa7b9005a7011609415527dc
2020-12-08 04:18:00 +00:00
// getNodesEstimate updates the estimate of the total number of nodes. It is guaranteed
// to return a number greater than 0 when the error is nil.
//
// We can't calculate this upon first starting a Checker, because there may not be any
// nodes yet. We expect that there will be nodes before there are segments, though.
func (checker *Checker) getNodesEstimate(ctx context.Context) (int, error) {
// this should be safe to call frequently; it is an efficient caching lookup.
totalNumNodes, err := checker.nodestate.NumNodes(ctx)
if err != nil {
// We could proceed here by returning the last good value, or by returning a fallback
// constant estimate, like "20000", and we'd probably be fine, but it would be better
// not to have that happen silently for too long. Also, if we can't get this from the
// database, we probably can't modify the injured segments queue, so it won't help to
// proceed with this repair operation.
return 0, err
}
if totalNumNodes == 0 {
return 0, Error.New("segment health is meaningless: there are no nodes")
}
return totalNumNodes, nil
}
// RefreshReliabilityCache forces refreshing node online status cache.
func (checker *Checker) RefreshReliabilityCache(ctx context.Context) error {
return checker.nodestate.Refresh(ctx)
}
// Close halts the Checker loop.
func (checker *Checker) Close() error {
checker.Loop.Close()
return nil
}
Satellite Peer (#1034) * add satellite peer * Add overlay * reorganize kademlia * add RunRefresh * add refresh to storagenode.Peer * add discovery * add agreements and metainfo * rename * add datarepair checker * add repair * add todo notes for audit * add testing interface * add into testplanet * fixes * fix compilation errors * fix compilation errors * make testplanet run * remove audit refrences * ensure that audit tests run * dev * checker tests compilable * fix discovery * fix compilation * fix * fix * dev * fix * disable auth * fixes * revert go.mod/sum * fix linter errors * fix * fix copyright * Add address param for SN dashboard (#1076) * Rename storj-sdk to storj-sim (#1078) * Storagenode logs and config improvements (#1075) * Add more info to SN logs * remove config-dir from user config * add output where config was stored * add message for successful connection * fix linter * remove storage.path from user config * resolve config path * move success message to info * log improvements * Remove captplanet (#1070) * pkg/server: include production cert (#1082) Change-Id: Ie8e6fe78550be83c3bd797db7a1e58d37c684792 * Generate Payments Report (#1079) * memory.Size: autoformat sizes based on value entropy (#1081) * Jj/bytes (#1085) * run tally and rollup * sets dev default tally and rollup intervals * nonessential storj-sim edits (#1086) * Closing context doesn't stop storage node (#1084) * Print when cancelled * Close properly * Don't log nil * Don't print error when closing dashboard * Fix panic in inspector if ping fails (#1088) * Consolidate identity management to identity cli commands (#1083) * Consolidate identity management: Move identity cretaion/signing out of storagenode setup command. * fixes * linters * Consolidate identity management: Move identity cretaion/signing out of storagenode setup command. * fixes * sava backups before saving signed certs * add "-prebuilt-test-cmds" test flag * linters * prepare cli tests for travis * linter fixes * more fixes * linter gods * sp/sdk/sim * remove ca.difficulty * remove unused difficulty * return setup to its rightful place * wip travis * Revert "wip travis" This reverts commit 56834849dcf066d3cc0a4f139033fc3f6d7188ca. * typo in travis.yaml * remove tests * remove more * make it only create one identity at a time for consistency * add config-dir for consitency * add identity creation to storj-sim * add flags * simplify * fix nolint and compile * prevent overwrite and pass difficulty, concurrency, and parent creds * goimports
2019-01-18 13:54:08 +00:00
// IdentifyInjuredSegments checks for missing pieces off of the metainfo and overlay.
func (checker *Checker) IdentifyInjuredSegments(ctx context.Context) (err error) {
defer mon.Task()(&ctx)(&err)
startTime := time.Now()
observer := &checkerObserver{
satellite/repair: use survivability model for segment health The chief segment health models we've come up with are the "immediate danger" model and the "survivability" model. The former calculates the chance of losing a segment becoming lost in the next time period (using the CDF of the binomial distribution to estimate the chance of x nodes failing in that period), while the latter estimates the number of iterations for which a segment can be expected to survive (using the mean of the negative binomial distribution). The immediate danger model was a promising one for comparing segment health across segments with different RS parameters, as it is more precisely what we want to prevent, but it turns out that practically all segments in production have infinite health, as the chance of losing segments with any reasonable estimate of node failure rate is smaller than DBL_EPSILON, the smallest possible difference from 1.0 representable in a float64 (about 1e-16). Leaving aside the wisdom of worrying about the repair of segments that have less than a 1e-16 chance of being lost, we want to be extremely conservative and proactive in our repair efforts, and the health of the segments we have been repairing thus far also evaluates to infinity under the immediate danger model. Thus, we find ourselves reaching for an alternative. Dr. Ben saves the day: the survivability model is a reasonably close approximation of the immediate danger model, and even better, it is far simpler to calculate and yields manageable values for real-world segments. The downside to it is that it requires as input an estimate of the total number of active nodes. This change replaces the segment health calculation to use the survivability model, and reinstates the call to SegmentHealth() where it was reverted. It gets estimates for the total number of active nodes by leveraging the reliability cache. Change-Id: Ia5d9b9031b9f6cf0fa7b9005a7011609415527dc
2020-12-08 04:18:00 +00:00
repairQueue: checker.repairQueue,
irrdb: checker.irrdb,
nodestate: checker.nodestate,
statsCollector: checker.statsCollector,
monStats: aggregateStats{},
repairOverrides: checker.repairOverrides,
nodeFailureRate: checker.nodeFailureRate,
getNodesEstimate: checker.getNodesEstimate,
log: checker.logger,
}
err = checker.metaLoop.Join(ctx, observer)
if err != nil {
if !errs2.IsCanceled(err) {
checker.logger.Error("IdentifyInjuredSegments error", zap.Error(err))
}
return err
}
// remove all segments which were not seen as unhealthy by this checker iteration
healthyDeleted, err := checker.repairQueue.Clean(ctx, startTime)
if err != nil {
return Error.Wrap(err)
}
checker.statsCollector.collectAggregates()
mon.IntVal("remote_files_checked").Observe(observer.monStats.objectsChecked) //mon:locked
mon.IntVal("remote_segments_checked").Observe(observer.monStats.remoteSegmentsChecked) //mon:locked
mon.IntVal("remote_segments_failed_to_check").Observe(observer.monStats.remoteSegmentsFailedToCheck) //mon:locked
mon.IntVal("remote_segments_needing_repair").Observe(observer.monStats.remoteSegmentsNeedingRepair) //mon:locked
mon.IntVal("new_remote_segments_needing_repair").Observe(observer.monStats.newRemoteSegmentsNeedingRepair) //mon:locked
mon.IntVal("remote_segments_lost").Observe(observer.monStats.remoteSegmentsLost) //mon:locked
mon.IntVal("remote_files_lost").Observe(int64(len(observer.monStats.remoteSegmentInfo))) //mon:locked
mon.IntVal("remote_segments_over_threshold_1").Observe(observer.monStats.remoteSegmentsOverThreshold[0]) //mon:locked
mon.IntVal("remote_segments_over_threshold_2").Observe(observer.monStats.remoteSegmentsOverThreshold[1]) //mon:locked
mon.IntVal("remote_segments_over_threshold_3").Observe(observer.monStats.remoteSegmentsOverThreshold[2]) //mon:locked
mon.IntVal("remote_segments_over_threshold_4").Observe(observer.monStats.remoteSegmentsOverThreshold[3]) //mon:locked
mon.IntVal("remote_segments_over_threshold_5").Observe(observer.monStats.remoteSegmentsOverThreshold[4]) //mon:locked
mon.IntVal("healthy_segments_removed_from_queue").Observe(healthyDeleted) //mon:locked
allUnhealthy := observer.monStats.remoteSegmentsNeedingRepair + observer.monStats.remoteSegmentsFailedToCheck
allChecked := observer.monStats.remoteSegmentsChecked
allHealthy := allChecked - allUnhealthy
mon.FloatVal("remote_segments_healthy_percentage").Observe(100 * float64(allHealthy) / float64(allChecked)) //mon:locked
return nil
}
// checks for a object location in slice.
func containsObjectLocation(a []metabase.ObjectLocation, x metabase.ObjectLocation) bool {
for _, n := range a {
if x == n {
return true
}
}
return false
}
func (checker *Checker) updateIrreparableSegmentStatus(ctx context.Context, key metabase.SegmentKey, redundancy storj.RedundancyScheme, creationDate time.Time, pieces metabase.Pieces) (err error) {
defer mon.Task()(&ctx)(&err)
if len(pieces) == 0 {
checker.logger.Debug("no pieces on remote segment")
return nil
}
missingPieces, err := checker.nodestate.MissingPieces(ctx, creationDate, pieces)
if err != nil {
return errs.Combine(Error.New("error getting missing pieces"), err)
}
numHealthy := int32(len(pieces) - len(missingPieces))
repairThreshold := int32(redundancy.RepairShares)
pbRedundancy := &pb.RedundancyScheme{
MinReq: int32(redundancy.RequiredShares),
RepairThreshold: int32(redundancy.RepairShares),
SuccessThreshold: int32(redundancy.OptimalShares),
Total: int32(redundancy.TotalShares),
}
overrideValue := checker.repairOverrides.GetOverrideValuePB(pbRedundancy)
if overrideValue != 0 {
repairThreshold = overrideValue
}
satellite/repair: use survivability model for segment health The chief segment health models we've come up with are the "immediate danger" model and the "survivability" model. The former calculates the chance of losing a segment becoming lost in the next time period (using the CDF of the binomial distribution to estimate the chance of x nodes failing in that period), while the latter estimates the number of iterations for which a segment can be expected to survive (using the mean of the negative binomial distribution). The immediate danger model was a promising one for comparing segment health across segments with different RS parameters, as it is more precisely what we want to prevent, but it turns out that practically all segments in production have infinite health, as the chance of losing segments with any reasonable estimate of node failure rate is smaller than DBL_EPSILON, the smallest possible difference from 1.0 representable in a float64 (about 1e-16). Leaving aside the wisdom of worrying about the repair of segments that have less than a 1e-16 chance of being lost, we want to be extremely conservative and proactive in our repair efforts, and the health of the segments we have been repairing thus far also evaluates to infinity under the immediate danger model. Thus, we find ourselves reaching for an alternative. Dr. Ben saves the day: the survivability model is a reasonably close approximation of the immediate danger model, and even better, it is far simpler to calculate and yields manageable values for real-world segments. The downside to it is that it requires as input an estimate of the total number of active nodes. This change replaces the segment health calculation to use the survivability model, and reinstates the call to SegmentHealth() where it was reverted. It gets estimates for the total number of active nodes by leveraging the reliability cache. Change-Id: Ia5d9b9031b9f6cf0fa7b9005a7011609415527dc
2020-12-08 04:18:00 +00:00
totalNumNodes, err := checker.getNodesEstimate(ctx)
if err != nil {
return Error.New("could not get estimate of total number of nodes: %w", err)
}
// we repair when the number of healthy pieces is less than or equal to the repair threshold and is greater or equal to
// minimum required pieces in redundancy
// except for the case when the repair and success thresholds are the same (a case usually seen during testing)
//
// If the segment is suddenly entirely healthy again, we don't need to repair and we don't need to
// keep it in the irreparabledb queue either.
if numHealthy >= int32(redundancy.RequiredShares) && numHealthy <= repairThreshold && numHealthy < int32(redundancy.OptimalShares) {
segmentHealth := repair.SegmentHealth(int(numHealthy), int(redundancy.RequiredShares), totalNumNodes, checker.nodeFailureRate)
_, err = checker.repairQueue.Insert(ctx, &internalpb.InjuredSegment{
Path: key,
LostPieces: missingPieces,
InsertedTime: time.Now().UTC(),
}, segmentHealth)
if err != nil {
return errs.Combine(Error.New("error adding injured segment to queue"), err)
}
// delete always returns nil when something was deleted and also when element didn't exists
err = checker.irrdb.Delete(ctx, key)
if err != nil {
checker.logger.Error("error deleting entry from irreparable db: ", zap.Error(err))
}
} else if numHealthy < int32(redundancy.RequiredShares) && numHealthy < repairThreshold {
// make an entry into the irreparable table
segmentInfo := &internalpb.IrreparableSegment{
Path: key,
LostPieces: int32(len(missingPieces)),
LastRepairAttempt: time.Now().Unix(),
RepairAttemptCount: int64(1),
}
// add the entry if new or update attempt count if already exists
err := checker.irrdb.IncrementRepairAttempts(ctx, segmentInfo)
if err != nil {
return errs.Combine(Error.New("error handling irreparable segment to queue"), err)
}
} else if numHealthy > repairThreshold || numHealthy >= int32(redundancy.OptimalShares) {
err = checker.irrdb.Delete(ctx, key)
if err != nil {
return Error.New("error removing segment from irreparable queue: %v", err)
}
}
return nil
}
var _ metaloop.Observer = (*checkerObserver)(nil)
2019-09-10 14:24:16 +01:00
// checkerObserver implements the metainfo loop Observer interface.
2019-09-10 14:24:16 +01:00
//
// architecture: Observer
type checkerObserver struct {
satellite/repair: use survivability model for segment health The chief segment health models we've come up with are the "immediate danger" model and the "survivability" model. The former calculates the chance of losing a segment becoming lost in the next time period (using the CDF of the binomial distribution to estimate the chance of x nodes failing in that period), while the latter estimates the number of iterations for which a segment can be expected to survive (using the mean of the negative binomial distribution). The immediate danger model was a promising one for comparing segment health across segments with different RS parameters, as it is more precisely what we want to prevent, but it turns out that practically all segments in production have infinite health, as the chance of losing segments with any reasonable estimate of node failure rate is smaller than DBL_EPSILON, the smallest possible difference from 1.0 representable in a float64 (about 1e-16). Leaving aside the wisdom of worrying about the repair of segments that have less than a 1e-16 chance of being lost, we want to be extremely conservative and proactive in our repair efforts, and the health of the segments we have been repairing thus far also evaluates to infinity under the immediate danger model. Thus, we find ourselves reaching for an alternative. Dr. Ben saves the day: the survivability model is a reasonably close approximation of the immediate danger model, and even better, it is far simpler to calculate and yields manageable values for real-world segments. The downside to it is that it requires as input an estimate of the total number of active nodes. This change replaces the segment health calculation to use the survivability model, and reinstates the call to SegmentHealth() where it was reverted. It gets estimates for the total number of active nodes by leveraging the reliability cache. Change-Id: Ia5d9b9031b9f6cf0fa7b9005a7011609415527dc
2020-12-08 04:18:00 +00:00
repairQueue queue.RepairQueue
irrdb irreparable.DB
nodestate *ReliabilityCache
statsCollector *statsCollector
monStats aggregateStats // TODO(cam): once we verify statsCollector reports data correctly, remove this
repairOverrides RepairOverridesMap
nodeFailureRate float64
getNodesEstimate func(ctx context.Context) (int, error)
log *zap.Logger
// we need to delay counting objects to ensure they get associated with the correct redundancy only once
objectCounted bool
}
func (obs *checkerObserver) getStatsByRS(redundancy storj.RedundancyScheme) *stats {
rsString := getRSString(obs.loadRedundancy(redundancy))
return obs.statsCollector.getStatsByRS(rsString)
}
func (obs *checkerObserver) loadRedundancy(redundancy storj.RedundancyScheme) (int, int, int, int) {
repair := int(redundancy.RepairShares)
overrideValue := obs.repairOverrides.GetOverrideValue(redundancy)
if overrideValue != 0 {
repair = int(overrideValue)
}
return int(redundancy.RequiredShares), repair, int(redundancy.OptimalShares), int(redundancy.TotalShares)
}
// LoopStarted is called at each start of a loop.
func (obs *checkerObserver) LoopStarted(context.Context, metaloop.LoopInfo) (err error) {
return nil
}
func (obs *checkerObserver) RemoteSegment(ctx context.Context, segment *metaloop.Segment) (err error) {
defer mon.Task()(&ctx)(&err)
// ignore segment if expired
if segment.Expired(time.Now()) {
return nil
}
stats := obs.getStatsByRS(segment.Redundancy)
if !obs.objectCounted {
obs.objectCounted = true
stats.iterationAggregates.objectsChecked++
}
obs.monStats.remoteSegmentsChecked++
stats.iterationAggregates.remoteSegmentsChecked++
// ensure we get values, even if only zero values, so that redash can have an alert based on this
mon.Counter("checker_segments_below_min_req").Inc(0) //mon:locked
stats.segmentsBelowMinReq.Inc(0)
pieces := segment.Pieces
if len(pieces) == 0 {
obs.log.Debug("no pieces on remote segment")
return nil
}
pbPieces := make([]*pb.RemotePiece, len(pieces))
for i, piece := range pieces {
pbPieces[i] = &pb.RemotePiece{
PieceNum: int32(piece.Number),
NodeId: piece.StorageNode,
}
}
satellite/repair: use survivability model for segment health The chief segment health models we've come up with are the "immediate danger" model and the "survivability" model. The former calculates the chance of losing a segment becoming lost in the next time period (using the CDF of the binomial distribution to estimate the chance of x nodes failing in that period), while the latter estimates the number of iterations for which a segment can be expected to survive (using the mean of the negative binomial distribution). The immediate danger model was a promising one for comparing segment health across segments with different RS parameters, as it is more precisely what we want to prevent, but it turns out that practically all segments in production have infinite health, as the chance of losing segments with any reasonable estimate of node failure rate is smaller than DBL_EPSILON, the smallest possible difference from 1.0 representable in a float64 (about 1e-16). Leaving aside the wisdom of worrying about the repair of segments that have less than a 1e-16 chance of being lost, we want to be extremely conservative and proactive in our repair efforts, and the health of the segments we have been repairing thus far also evaluates to infinity under the immediate danger model. Thus, we find ourselves reaching for an alternative. Dr. Ben saves the day: the survivability model is a reasonably close approximation of the immediate danger model, and even better, it is far simpler to calculate and yields manageable values for real-world segments. The downside to it is that it requires as input an estimate of the total number of active nodes. This change replaces the segment health calculation to use the survivability model, and reinstates the call to SegmentHealth() where it was reverted. It gets estimates for the total number of active nodes by leveraging the reliability cache. Change-Id: Ia5d9b9031b9f6cf0fa7b9005a7011609415527dc
2020-12-08 04:18:00 +00:00
totalNumNodes, err := obs.getNodesEstimate(ctx)
if err != nil {
return Error.New("could not get estimate of total number of nodes: %w", err)
}
createdAt := time.Time{}
if segment.CreatedAt != nil {
createdAt = *segment.CreatedAt
}
repairedAt := time.Time{}
if segment.RepairedAt != nil {
repairedAt = *segment.RepairedAt
}
missingPieces, err := obs.nodestate.MissingPieces(ctx, createdAt, segment.Pieces)
if err != nil {
obs.monStats.remoteSegmentsFailedToCheck++
stats.iterationAggregates.remoteSegmentsFailedToCheck++
return errs.Combine(Error.New("error getting missing pieces"), err)
}
numHealthy := len(pieces) - len(missingPieces)
mon.IntVal("checker_segment_total_count").Observe(int64(len(pieces))) //mon:locked
stats.segmentTotalCount.Observe(int64(len(pieces)))
mon.IntVal("checker_segment_healthy_count").Observe(int64(numHealthy)) //mon:locked
stats.segmentHealthyCount.Observe(int64(numHealthy))
segmentAge := time.Since(createdAt)
mon.IntVal("checker_segment_age").Observe(int64(segmentAge.Seconds())) //mon:locked
stats.segmentAge.Observe(int64(segmentAge.Seconds()))
required, repairThreshold, successThreshold, _ := obs.loadRedundancy(segment.Redundancy)
satellite/repair: use survivability model for segment health The chief segment health models we've come up with are the "immediate danger" model and the "survivability" model. The former calculates the chance of losing a segment becoming lost in the next time period (using the CDF of the binomial distribution to estimate the chance of x nodes failing in that period), while the latter estimates the number of iterations for which a segment can be expected to survive (using the mean of the negative binomial distribution). The immediate danger model was a promising one for comparing segment health across segments with different RS parameters, as it is more precisely what we want to prevent, but it turns out that practically all segments in production have infinite health, as the chance of losing segments with any reasonable estimate of node failure rate is smaller than DBL_EPSILON, the smallest possible difference from 1.0 representable in a float64 (about 1e-16). Leaving aside the wisdom of worrying about the repair of segments that have less than a 1e-16 chance of being lost, we want to be extremely conservative and proactive in our repair efforts, and the health of the segments we have been repairing thus far also evaluates to infinity under the immediate danger model. Thus, we find ourselves reaching for an alternative. Dr. Ben saves the day: the survivability model is a reasonably close approximation of the immediate danger model, and even better, it is far simpler to calculate and yields manageable values for real-world segments. The downside to it is that it requires as input an estimate of the total number of active nodes. This change replaces the segment health calculation to use the survivability model, and reinstates the call to SegmentHealth() where it was reverted. It gets estimates for the total number of active nodes by leveraging the reliability cache. Change-Id: Ia5d9b9031b9f6cf0fa7b9005a7011609415527dc
2020-12-08 04:18:00 +00:00
segmentHealth := repair.SegmentHealth(numHealthy, required, totalNumNodes, obs.nodeFailureRate)
mon.FloatVal("checker_segment_health").Observe(segmentHealth) //mon:locked
stats.segmentHealth.Observe(segmentHealth)
key := segment.Location.Encode()
// we repair when the number of healthy pieces is less than or equal to the repair threshold and is greater or equal to
// minimum required pieces in redundancy
// except for the case when the repair and success thresholds are the same (a case usually seen during testing)
if numHealthy >= required && numHealthy <= repairThreshold && numHealthy < successThreshold {
mon.FloatVal("checker_injured_segment_health").Observe(segmentHealth) //mon:locked
stats.injuredSegmentHealth.Observe(segmentHealth)
obs.monStats.remoteSegmentsNeedingRepair++
stats.iterationAggregates.remoteSegmentsNeedingRepair++
alreadyInserted, err := obs.repairQueue.Insert(ctx, &internalpb.InjuredSegment{
Path: key,
LostPieces: missingPieces,
InsertedTime: time.Now().UTC(),
}, segmentHealth)
if err != nil {
obs.log.Error("error adding injured segment to queue", zap.Error(err))
return nil
}
if !alreadyInserted {
obs.monStats.newRemoteSegmentsNeedingRepair++
stats.iterationAggregates.newRemoteSegmentsNeedingRepair++
}
// delete always returns nil when something was deleted and also when element didn't exists
err = obs.irrdb.Delete(ctx, key)
if err != nil {
obs.log.Error("error deleting entry from irreparable db", zap.Error(err))
return nil
}
} else if numHealthy < required && numHealthy < repairThreshold {
lostSegInfo := segment.Location.Object()
if !containsObjectLocation(obs.monStats.remoteSegmentInfo, lostSegInfo) {
obs.monStats.remoteSegmentInfo = append(obs.monStats.remoteSegmentInfo, lostSegInfo)
}
if !containsObjectLocation(stats.iterationAggregates.remoteSegmentInfo, lostSegInfo) {
stats.iterationAggregates.remoteSegmentInfo = append(stats.iterationAggregates.remoteSegmentInfo, lostSegInfo)
}
var segmentAge time.Duration
if createdAt.Before(repairedAt) {
segmentAge = time.Since(repairedAt)
} else {
segmentAge = time.Since(createdAt)
}
mon.IntVal("checker_segment_time_until_irreparable").Observe(int64(segmentAge.Seconds())) //mon:locked
stats.segmentTimeUntilIrreparable.Observe(int64(segmentAge.Seconds()))
obs.monStats.remoteSegmentsLost++
stats.iterationAggregates.remoteSegmentsLost++
mon.Counter("checker_segments_below_min_req").Inc(1) //mon:locked
stats.segmentsBelowMinReq.Inc(1)
// make an entry into the irreparable table
segmentInfo := &internalpb.IrreparableSegment{
Path: key,
LostPieces: int32(len(missingPieces)),
LastRepairAttempt: time.Now().Unix(),
RepairAttemptCount: int64(1),
}
// add the entry if new or update attempt count if already exists
err := obs.irrdb.IncrementRepairAttempts(ctx, segmentInfo)
if err != nil {
obs.log.Error("error handling irreparable segment to queue", zap.Error(err))
return nil
}
} else {
if numHealthy > repairThreshold && numHealthy <= (repairThreshold+len(obs.monStats.remoteSegmentsOverThreshold)) {
// record metrics for segments right above repair threshold
// numHealthy=repairThreshold+1 through numHealthy=repairThreshold+5
for i := range obs.monStats.remoteSegmentsOverThreshold {
if numHealthy == (repairThreshold + i + 1) {
obs.monStats.remoteSegmentsOverThreshold[i]++
break
}
}
}
if numHealthy > repairThreshold && numHealthy <= (repairThreshold+len(stats.iterationAggregates.remoteSegmentsOverThreshold)) {
// record metrics for segments right above repair threshold
// numHealthy=repairThreshold+1 through numHealthy=repairThreshold+5
for i := range stats.iterationAggregates.remoteSegmentsOverThreshold {
if numHealthy == (repairThreshold + i + 1) {
stats.iterationAggregates.remoteSegmentsOverThreshold[i]++
break
}
}
}
}
return nil
}
func (obs *checkerObserver) Object(ctx context.Context, object *metaloop.Object) (err error) {
defer mon.Task()(&ctx)(&err)
obs.monStats.objectsChecked++
// TODO: check for expired objects
if object.SegmentCount == 0 {
stats := obs.getStatsByRS(storj.RedundancyScheme{})
stats.iterationAggregates.objectsChecked++
return nil
}
obs.objectCounted = false
return nil
}
func (obs *checkerObserver) InlineSegment(ctx context.Context, segment *metaloop.Segment) (err error) {
defer mon.Task()(&ctx)(&err)
// TODO: check for expired segments
if !obs.objectCounted {
// Note: this may give false stats when an object starts with a inline segment.
obs.objectCounted = true
stats := obs.getStatsByRS(storj.RedundancyScheme{})
stats.iterationAggregates.objectsChecked++
}
return nil
}
// IrreparableProcess iterates over all items in the irreparabledb. If an item can
// now be repaired then it is added to a worker queue.
func (checker *Checker) IrreparableProcess(ctx context.Context) (err error) {
defer mon.Task()(&ctx)(&err)
const limit = 1000
lastSeenSegmentKey := metabase.SegmentKey{}
for {
segments, err := checker.irrdb.GetLimited(ctx, limit, lastSeenSegmentKey)
if err != nil {
return errs.Combine(Error.New("error reading segment from the queue"), err)
}
// zero segments returned with nil err
if len(segments) == 0 {
break
}
lastSeenSegmentKey = metabase.SegmentKey(segments[len(segments)-1].Path)
for _, segment := range segments {
var redundancy storj.RedundancyScheme
var pieces metabase.Pieces
var createAt time.Time
if segment.SegmentDetail == (&pb.Pointer{}) {
// TODO IrreparableDB will be removed in a future so we shouldn't care too much about performance
location, err := metabase.ParseSegmentKey(metabase.SegmentKey(segment.GetPath()))
if err != nil {
return err
}
object, err := checker.metabase.GetObjectLatestVersion(ctx, metabase.GetObjectLatestVersion{
ObjectLocation: location.Object(),
})
if err != nil {
return err
}
createAt = object.CreatedAt
segment, err := checker.metabase.GetSegmentByPosition(ctx, metabase.GetSegmentByPosition{
StreamID: object.StreamID,
Position: location.Position,
})
if err != nil {
return err
}
redundancy = segment.Redundancy
} else {
// skip inline segments
if segment.SegmentDetail.Remote == nil {
return nil
}
createAt = segment.SegmentDetail.CreationDate
pbRedundancy := segment.SegmentDetail.Remote.Redundancy
redundancy = storj.RedundancyScheme{
RequiredShares: int16(pbRedundancy.MinReq),
RepairShares: int16(pbRedundancy.RepairThreshold),
OptimalShares: int16(pbRedundancy.SuccessThreshold),
TotalShares: int16(pbRedundancy.Total),
ShareSize: pbRedundancy.ErasureShareSize,
}
pieces = make(metabase.Pieces, len(segment.SegmentDetail.Remote.RemotePieces))
for _, piece := range segment.SegmentDetail.Remote.RemotePieces {
pieces = append(pieces, metabase.Piece{
Number: uint16(piece.PieceNum),
StorageNode: piece.NodeId,
})
}
}
err = checker.updateIrreparableSegmentStatus(ctx,
metabase.SegmentKey(segment.GetPath()),
redundancy,
createAt,
pieces,
)
if err != nil {
checker.logger.Error("irrepair segment checker failed: ", zap.Error(err))
}
}
}
return nil
}