[#965] policer: Implement continuous replication

Signed-off-by: Alex Vanin <alexey@nspcc.ru>

cmd/neofs-node/config.go

@@ -197,6 +197,10 @@ type cfgLocalStorage struct {
 
 type cfgObjectRoutines struct {
 	putRemote *ants.Pool
+
+	putRemoteCapacity int
+
+	replication *ants.Pool
 }
 
 type cfgControlService struct {
@@ -449,10 +453,13 @@ func initObjectPool(cfg *config.Config) (pool cfgObjectRoutines) {
 
 	optNonBlocking := ants.WithNonblocking(true)
 
-	pool.putRemote, err = ants.NewPool(objectconfig.Put(cfg).PoolSizeRemote(), optNonBlocking)
-	if err != nil {
-		fatalOnErr(err)
-	}
+	pool.putRemoteCapacity = objectconfig.Put(cfg).PoolSizeRemote()
+
+	pool.putRemote, err = ants.NewPool(pool.putRemoteCapacity, optNonBlocking)
+	fatalOnErr(err)
+
+	pool.replication, err = ants.NewPool(pool.putRemoteCapacity)
+	fatalOnErr(err)
 
 	return pool
 }
@@ -486,3 +493,10 @@ func (c *cfg) bootstrap() error {
 func (c *cfg) needBootstrap() bool {
 	return c.cfgNetmap.needBootstrap
 }
+
+// ObjectServiceLoad implements system loader interface for policer component.
+// It is calculated as size/capacity ratio of "remote object put" worker.
+// Returns float value between 0.0 and 1.0.
+func (c *cfg) ObjectServiceLoad() float64 {
+	return float64(c.cfgObject.pool.putRemote.Running()) / float64(c.cfgObject.pool.putRemoteCapacity)
+}

cmd/neofs-node/object.go

@@ -17,7 +17,6 @@ import (
 	morphClient "github.com/nspcc-dev/neofs-node/pkg/morph/client"
 	cntrwrp "github.com/nspcc-dev/neofs-node/pkg/morph/client/container/wrapper"
 	nmwrp "github.com/nspcc-dev/neofs-node/pkg/morph/client/netmap/wrapper"
-	"github.com/nspcc-dev/neofs-node/pkg/morph/event"
 	objectTransportGRPC "github.com/nspcc-dev/neofs-node/pkg/network/transport/object/grpc"
 	objectService "github.com/nspcc-dev/neofs-node/pkg/services/object"
 	"github.com/nspcc-dev/neofs-node/pkg/services/object/acl"
@@ -232,8 +231,6 @@ func initObjectService(c *cfg) {
 
 	c.workers = append(c.workers, repl)
 
-	ch := make(chan *policer.Task, 1)
-
 	pol := policer.New(
 		policer.WithLogger(c.log),
 		policer.WithLocalStorage(ls),
@@ -241,9 +238,6 @@ func initObjectService(c *cfg) {
 		policer.WithPlacementBuilder(
 			placement.NewNetworkMapSourceBuilder(c.cfgObject.netMapSource),
 		),
-		policer.WithWorkScope(100),
-		policer.WithExpansionRate(10),
-		policer.WithTrigger(ch),
 		policer.WithRemoteHeader(
 			headsvc.NewRemoteHeader(keyStorage, clientConstructor),
 		),
@@ -260,18 +254,11 @@ func initObjectService(c *cfg) {
 				)
 			}
 		}),
+		policer.WithMaxCapacity(c.cfgObject.pool.putRemoteCapacity),
+		policer.WithPool(c.cfgObject.pool.replication),
+		policer.WithNodeLoader(c),
 	)
 
-	addNewEpochNotificationHandler(c, func(ev event.Event) {
-		select {
-		case ch <- new(policer.Task):
-		case <-c.ctx.Done():
-			close(ch)
-		default:
-			c.log.Info("policer is busy")
-		}
-	})
-
 	traverseGen := util.NewTraverserGenerator(c.cfgObject.netMapSource, c.cfgObject.cnrSource, c)
 
 	c.workers = append(c.workers, pol)

pkg/services/policer/policer.go

@@ -1,9 +1,9 @@
 package policer
 
 import (
-	"sync"
 	"time"
 
+	lru "github.com/hashicorp/golang-lru"
 	"github.com/nspcc-dev/neofs-node/pkg/core/container"
 	"github.com/nspcc-dev/neofs-node/pkg/core/netmap"
 	"github.com/nspcc-dev/neofs-node/pkg/local_object_storage/engine"
@@ -12,15 +12,22 @@ import (
 	"github.com/nspcc-dev/neofs-node/pkg/services/replicator"
 	"github.com/nspcc-dev/neofs-node/pkg/util/logger"
 	"github.com/nspcc-dev/neofs-sdk-go/object"
+	"github.com/panjf2000/ants/v2"
 	"go.uber.org/zap"
 )
 
+// NodeLoader provides application load statistics.
+type nodeLoader interface {
+	// ObjectServiceLoad returns object service load value in [0:1] range.
+	ObjectServiceLoad() float64
+}
+
 // Policer represents the utility that verifies
 // compliance with the object storage policy.
 type Policer struct {
 	*cfg
 
-	prevTask prevTask
+	cache *lru.Cache
 }
 
 // Option is an option for Policer constructor.
@@ -33,12 +40,8 @@ type RedundantCopyCallback func(*object.Address)
 type cfg struct {
 	headTimeout time.Duration
 
-	workScope workScope
-
 	log *logger.Logger
 
-	trigger <-chan *Task
-
 	jobQueue jobQueue
 
 	cnrSrc container.Source
@@ -52,11 +55,25 @@ type cfg struct {
 	replicator *replicator.Replicator
 
 	cbRedundantCopy RedundantCopyCallback
+
+	taskPool *ants.Pool
+
+	loader nodeLoader
+
+	maxCapacity int
+
+	batchSize, cacheSize uint32
+
+	rebalanceFreq, evictDuration time.Duration
 }
 
 func defaultCfg() *cfg {
 	return &cfg{
-		log: zap.L(),
+		log:           zap.L(),
+		batchSize:     10,
+		cacheSize:     200_000, // should not allocate more than 200 MiB
+		rebalanceFreq: 1 * time.Second,
+		evictDuration: 30 * time.Second,
 	}
 }
 
@@ -70,12 +87,14 @@ func New(opts ...Option) *Policer {
 
 	c.log = c.log.With(zap.String("component", "Object Policer"))
 
+	cache, err := lru.New(int(c.cacheSize))
+	if err != nil {
+		panic(err)
+	}
+
 	return &Policer{
-		cfg: c,
-		prevTask: prevTask{
-			cancel: func() {},
-			wait:   new(sync.WaitGroup),
-		},
+		cfg:   c,
+		cache: cache,
 	}
 }
 
@@ -86,27 +105,6 @@ func WithHeadTimeout(v time.Duration) Option {
 	}
 }
 
-// WithWorkScope returns option to set job work scope value of Policer.
-func WithWorkScope(v int) Option {
-	return func(c *cfg) {
-		c.workScope.val = v
-	}
-}
-
-// WithExpansionRate returns option to set expansion rate of Policer's works scope (in %).
-func WithExpansionRate(v int) Option {
-	return func(c *cfg) {
-		c.workScope.expRate = v
-	}
-}
-
-// WithTrigger returns option to set triggering channel of Policer.
-func WithTrigger(v <-chan *Task) Option {
-	return func(c *cfg) {
-		c.trigger = v
-	}
-}
-
 // WithLogger returns option to set Logger of Policer.
 func WithLogger(v *logger.Logger) Option {
 	return func(c *cfg) {
@@ -164,3 +162,26 @@ func WithRedundantCopyCallback(cb RedundantCopyCallback) Option {
 		c.cbRedundantCopy = cb
 	}
 }
+
+// WithMaxCapacity returns option to set max capacity
+// that can be set to the pool.
+func WithMaxCapacity(cap int) Option {
+	return func(c *cfg) {
+		c.maxCapacity = cap
+	}
+}
+
+// WithPool returns option to set pool for
+// policy and replication operations.
+func WithPool(p *ants.Pool) Option {
+	return func(c *cfg) {
+		c.taskPool = p
+	}
+}
+
+// WithNodeLoader returns option to set NeoFS node load source.
+func WithNodeLoader(l nodeLoader) Option {
+	return func(c *cfg) {
+		c.loader = l
+	}
+}

pkg/services/policer/process.go

@@ -2,113 +2,84 @@ package policer
 
 import (
 	"context"
-	"sync"
+	"errors"
+	"time"
 
+	"github.com/nspcc-dev/neofs-node/pkg/local_object_storage/engine"
+	"github.com/nspcc-dev/neofs-sdk-go/object"
 	"go.uber.org/zap"
 )
 
-// Task represents group of Policer tact parameters.
-type Task struct{}
-
-type prevTask struct {
-	undone int
-
-	cancel context.CancelFunc
-
-	wait *sync.WaitGroup
-}
-
-type workScope struct {
-	val int
-
-	expRate int // in %
-}
-
 func (p *Policer) Run(ctx context.Context) {
 	defer func() {
 		p.log.Info("routine stopped")
 	}()
 
-	p.log.Info("process routine",
-		zap.Int("work scope value", p.workScope.val),
-		zap.Int("expansion rate (%)", p.workScope.val),
-		zap.Duration("head timeout", p.headTimeout),
+	go p.poolCapacityWorker(ctx)
+	p.shardPolicyWorker(ctx)
+}
+
+func (p *Policer) shardPolicyWorker(ctx context.Context) {
+	var (
+		addrs  []*object.Address
+		cursor *engine.Cursor
+		err    error
 	)
 
+	for {
+		addrs, cursor, err = p.jobQueue.Select(cursor, p.batchSize)
+		if err != nil {
+			if errors.Is(err, engine.ErrEndOfListing) {
+				time.Sleep(time.Second) // finished whole cycle, sleep a bit
+				continue
+			}
+			p.log.Warn("failure at object select for replication", zap.Error(err))
+		}
+
+		for i := range addrs {
+			select {
+			case <-ctx.Done():
+				return
+			default:
+				addr := addrs[i]
+				addrStr := addr.String()
+				err = p.taskPool.Submit(func() {
+					v, ok := p.cache.Get(addrStr)
+					if ok && time.Since(v.(time.Time)) < p.evictDuration {
+						return
+					}
+
+					p.processObject(ctx, addr)
+					p.cache.Add(addrStr, time.Now())
+				})
+				if err != nil {
+					p.log.Warn("pool submission", zap.Error(err))
+				}
+			}
+		}
+	}
+}
+
+func (p *Policer) poolCapacityWorker(ctx context.Context) {
+	ticker := time.NewTicker(p.rebalanceFreq)
 	for {
 		select {
 		case <-ctx.Done():
-			p.prevTask.cancel()
-
-			p.log.Warn("context is done",
-				zap.String("error", ctx.Err().Error()),
-			)
-
+			ticker.Stop()
 			return
-		case task, ok := <-p.trigger:
-			if !ok {
-				p.log.Warn("trigger channel is closed")
-
-				return
+		case <-ticker.C:
+			neofsSysLoad := p.loader.ObjectServiceLoad()
+			newCapacity := int((1.0 - neofsSysLoad) * float64(p.maxCapacity))
+			if newCapacity == 0 {
+				newCapacity++
 			}
 
-			p.prevTask.cancel()
-			p.prevTask.wait.Wait()
-
-			var taskCtx context.Context
-
-			taskCtx, p.prevTask.cancel = context.WithCancel(ctx)
-
-			go p.handleTask(taskCtx, task)
+			if p.taskPool.Cap() != newCapacity {
+				p.taskPool.Tune(newCapacity)
+				p.log.Debug("tune replication capacity",
+					zap.Float64("system_load", neofsSysLoad),
+					zap.Int("new_capacity", newCapacity))
+			}
 		}
 	}
 }
-
-func (p *Policer) handleTask(ctx context.Context, task *Task) {
-	p.prevTask.wait.Add(1)
-
-	defer func() {
-		p.prevTask.wait.Done()
-		p.log.Info("finish work",
-			zap.Int("amount of unfinished objects", p.prevTask.undone),
-		)
-	}()
-
-	var delta int
-
-	// undone - amount of objects we couldn't process in last epoch
-	if p.prevTask.undone > 0 {
-		// if there are unprocessed objects, then lower your estimation
-		delta = -p.prevTask.undone
-	} else {
-		// otherwise try to expand
-		delta = p.workScope.val * p.workScope.expRate / 100
-	}
-
-	addrs, err := p.jobQueue.Select(p.workScope.val + delta)
-	if err != nil {
-		p.log.Warn("could not select objects",
-			zap.String("error", err.Error()),
-		)
-	}
-
-	// if there are NOT enough objects to fill the pool, do not change it
-	// otherwise expand or shrink it with the delta value
-	if len(addrs) >= p.workScope.val+delta {
-		p.workScope.val += delta
-	}
-
-	p.prevTask.undone = len(addrs)
-
-	for i := range addrs {
-		select {
-		case <-ctx.Done():
-			return
-		default:
-		}
-
-		p.processObject(ctx, addrs[i])
-
-		p.prevTask.undone--
-	}
-}

pkg/services/policer/queue.go

@@ -1,8 +1,9 @@
 package policer
 
 import (
+	"fmt"
+
 	"github.com/nspcc-dev/neofs-node/pkg/local_object_storage/engine"
-	"github.com/nspcc-dev/neofs-node/pkg/util/rand"
 	"github.com/nspcc-dev/neofs-sdk-go/object"
 )
 
@@ -10,23 +11,15 @@ type jobQueue struct {
 	localStorage *engine.StorageEngine
 }
 
-func (q *jobQueue) Select(limit int) ([]*object.Address, error) {
-	// TODO: optimize the logic for selecting objects
-	// We can prioritize objects for migration, newly arrived objects, etc.
-	// It is recommended to make changes after updating the metabase
+func (q *jobQueue) Select(cursor *engine.Cursor, count uint32) ([]*object.Address, *engine.Cursor, error) {
+	prm := new(engine.ListWithCursorPrm)
+	prm.WithCursor(cursor)
+	prm.WithCount(count)
 
-	res, err := engine.List(q.localStorage, 0) // consider some limit
+	res, err := q.localStorage.ListWithCursor(prm)
 	if err != nil {
-		return nil, err
+		return nil, nil, fmt.Errorf("cannot list objects in engine: %w", err)
 	}
 
-	rand.New().Shuffle(len(res), func(i, j int) {
-		res[i], res[j] = res[j], res[i]
-	})
-
-	if len(res) < limit {
-		return res, nil
-	}
-
-	return res[:limit], nil
+	return res.AddressList(), res.Cursor(), nil
 }