storj/satellite/repair/repairer/repairer.go
Moby von Briesen a933bcc99a satellite/repair/repairer/ec.go: add option for downloading pieces onto disk instead of in memory during repair
Add flag to satellite repairer, "InMemoryRepair" that allows the
satellite to decide whether to download the entire segment being
repaired into memory (this is what the satellite already does), or to
download it into temporary files on disk that will be read from in the
upload phase of repair.

This should help with handling high repair traffic on satellites that
cannot afford to spend 64mb of memory per repair worker.

Updates tests to test repair for both in memory and to disk.

Change-Id: Iddf591e165621497c98533d45bfea3c28b08a194
2020-03-27 16:41:00 +00:00

210 lines
7.7 KiB
Go

// Copyright (C) 2019 Storj Labs, Inc.
// See LICENSE for copying information.
package repairer
import (
"context"
"time"
"github.com/spacemonkeygo/monkit/v3"
"github.com/zeebo/errs"
"go.uber.org/zap"
"golang.org/x/sync/semaphore"
"storj.io/common/memory"
"storj.io/common/pb"
"storj.io/common/sync2"
"storj.io/storj/satellite/repair/irreparable"
"storj.io/storj/satellite/repair/queue"
"storj.io/storj/storage"
)
// Error is a standard error class for this package.
var (
Error = errs.Class("repairer error")
mon = monkit.Package()
)
// Config contains configurable values for repairer
type Config struct {
MaxRepair int `help:"maximum segments that can be repaired concurrently" releaseDefault:"5" devDefault:"1"`
Interval time.Duration `help:"how frequently repairer should try and repair more data" releaseDefault:"5m0s" devDefault:"1m0s"`
Timeout time.Duration `help:"time limit for uploading repaired pieces to new storage nodes" default:"5m0s"`
DownloadTimeout time.Duration `help:"time limit for downloading pieces from a node for repair" default:"5m0s"`
TotalTimeout time.Duration `help:"time limit for an entire repair job, from queue pop to upload completion" default:"45m"`
MaxBufferMem memory.Size `help:"maximum buffer memory (in bytes) to be allocated for read buffers" default:"4M"`
MaxExcessRateOptimalThreshold float64 `help:"ratio applied to the optimal threshold to calculate the excess of the maximum number of repaired pieces to upload" default:"0.05"`
InMemoryRepair bool `help:"whether to download pieces for repair in memory (true) or download to disk (false)" default:"false"`
}
// Service contains the information needed to run the repair service
//
// architecture: Worker
type Service struct {
log *zap.Logger
queue queue.RepairQueue
config *Config
JobLimiter *semaphore.Weighted
Loop *sync2.Cycle
repairer *SegmentRepairer
irrDB irreparable.DB
}
// NewService creates repairing service
func NewService(log *zap.Logger, queue queue.RepairQueue, config *Config, repairer *SegmentRepairer, irrDB irreparable.DB) *Service {
return &Service{
log: log,
queue: queue,
config: config,
JobLimiter: semaphore.NewWeighted(int64(config.MaxRepair)),
Loop: sync2.NewCycle(config.Interval),
repairer: repairer,
irrDB: irrDB,
}
}
// Close closes resources
func (service *Service) Close() error { return nil }
// WaitForPendingRepairs waits for all ongoing repairs to complete.
//
// NB: this assumes that service.config.MaxRepair will never be changed once this Service instance
// is initialized. If that is not a valid assumption, we should keep a copy of its initial value to
// use here instead.
func (service *Service) WaitForPendingRepairs() {
// Acquire and then release the entire capacity of the semaphore, ensuring that
// it is completely empty (or, at least it was empty at some point).
//
// No error return is possible here; context.Background() can't be canceled
_ = service.JobLimiter.Acquire(context.Background(), int64(service.config.MaxRepair))
service.JobLimiter.Release(int64(service.config.MaxRepair))
}
// Run runs the repairer service
func (service *Service) Run(ctx context.Context) (err error) {
defer mon.Task()(&ctx)(&err)
// Wait for all repairs to complete
defer service.WaitForPendingRepairs()
return service.Loop.Run(ctx, service.processWhileQueueHasItems)
}
// processWhileQueueHasItems keeps calling process() until the queue is empty or something
// else goes wrong in fetching from the queue.
func (service *Service) processWhileQueueHasItems(ctx context.Context) error {
for {
err := service.process(ctx)
if err != nil {
if storage.ErrEmptyQueue.Has(err) {
return nil
}
service.log.Error("process", zap.Error(Error.Wrap(err)))
return err
}
}
}
// process picks items from repair queue and spawns a repair worker
func (service *Service) process(ctx context.Context) (err error) {
defer mon.Task()(&ctx)(&err)
// wait until we are allowed to spawn a new job
if err := service.JobLimiter.Acquire(ctx, 1); err != nil {
return err
}
// IMPORTANT: this timeout must be started before service.queue.Select(), in case
// service.queue.Select() takes some non-negligible amount of time, so that we can depend on
// repair jobs being given up within some set interval after the time in the 'attempted'
// column in the queue table.
//
// This is the reason why we are using a semaphore in this somewhat awkward way instead of
// using a simpler sync2.Limiter pattern. We don't want this timeout to include the waiting
// time from the semaphore acquisition, but it _must_ include the queue fetch time. At the
// same time, we don't want to do the queue pop in a separate goroutine, because we want to
// return from service.Run when queue fetch fails.
ctx, cancel := context.WithTimeout(ctx, service.config.TotalTimeout)
seg, err := service.queue.Select(ctx)
if err != nil {
service.JobLimiter.Release(1)
cancel()
return err
}
service.log.Info("Retrieved segment from repair queue", zap.Binary("Segment", seg.GetPath()))
// this goroutine inherits the JobLimiter semaphore acquisition and is now responsible
// for releasing it.
go func() {
defer service.JobLimiter.Release(1)
defer cancel()
if err := service.worker(ctx, seg); err != nil {
service.log.Error("repair worker failed:", zap.Binary("Segment", seg.GetPath()), zap.Error(err))
}
}()
return nil
}
func (service *Service) worker(ctx context.Context, seg *pb.InjuredSegment) (err error) {
defer mon.Task()(&ctx)(&err)
workerStartTime := time.Now().UTC()
service.log.Info("Limiter running repair on segment",
zap.Binary("Segment", seg.GetPath()),
zap.String("Segment Path", string(seg.GetPath())))
// note that shouldDelete is used even in the case where err is not null
shouldDelete, err := service.repairer.Repair(ctx, string(seg.GetPath()))
if shouldDelete {
if irreparableErr, ok := err.(*irreparableError); ok {
service.log.Error("segment could not be repaired! adding to irreparableDB for more attention",
zap.Error(err),
zap.Binary("segment", seg.GetPath()))
segmentInfo := &pb.IrreparableSegment{
Path: seg.GetPath(),
SegmentDetail: irreparableErr.segmentInfo,
LostPieces: irreparableErr.piecesRequired - irreparableErr.piecesAvailable,
LastRepairAttempt: time.Now().Unix(),
RepairAttemptCount: int64(1),
}
if err := service.irrDB.IncrementRepairAttempts(ctx, segmentInfo); err != nil {
service.log.Error("failed to add segment to irreparableDB! will leave in repair queue", zap.Error(err))
shouldDelete = false
}
} else if err != nil {
service.log.Error("unexpected error repairing segment!",
zap.Error(err),
zap.Binary("segment", seg.GetPath()))
} else {
service.log.Info("removing repaired segment from repair queue",
zap.Binary("Segment", seg.GetPath()))
}
if shouldDelete {
delErr := service.queue.Delete(ctx, seg)
if delErr != nil {
err = errs.Combine(err, Error.New("failed to remove segment from queue: %v", delErr))
}
}
}
if err != nil {
return Error.Wrap(err)
}
repairedTime := time.Now().UTC()
timeForRepair := repairedTime.Sub(workerStartTime)
mon.FloatVal("time_for_repair").Observe(timeForRepair.Seconds()) //locked
insertedTime := seg.GetInsertedTime()
// do not send metrics if segment was added before the InsertedTime field was added
if !insertedTime.IsZero() {
timeSinceQueued := workerStartTime.Sub(insertedTime)
mon.FloatVal("time_since_checker_queue").Observe(timeSinceQueued.Seconds()) //locked
}
return nil
}