frostfs-node/pkg/local_object_storage/engine/evacuate.go

package engine

import (
	"context"
	"errors"
	"fmt"
	"strings"
	"sync/atomic"

	"git.frostfs.info/TrueCloudLab/frostfs-node/internal/logs"
	"git.frostfs.info/TrueCloudLab/frostfs-node/pkg/core/object"
	meta "git.frostfs.info/TrueCloudLab/frostfs-node/pkg/local_object_storage/metabase"
	"git.frostfs.info/TrueCloudLab/frostfs-node/pkg/local_object_storage/shard"
	"git.frostfs.info/TrueCloudLab/frostfs-node/pkg/local_object_storage/util/logicerr"
	tracingPkg "git.frostfs.info/TrueCloudLab/frostfs-node/pkg/tracing"
	"git.frostfs.info/TrueCloudLab/frostfs-node/pkg/util"
	"git.frostfs.info/TrueCloudLab/frostfs-observability/tracing"
	objectSDK "git.frostfs.info/TrueCloudLab/frostfs-sdk-go/object"
	oid "git.frostfs.info/TrueCloudLab/frostfs-sdk-go/object/id"
	"git.frostfs.info/TrueCloudLab/hrw"
	"go.opentelemetry.io/otel/attribute"
	"go.opentelemetry.io/otel/trace"
	"go.uber.org/zap"
)

var (
	ErrMustBeReadOnly = logicerr.New("shard must be in read-only mode")

	evacuationOperationLogField = zap.String("operation", "evacuation")
)

// EvacuateScope is an evacuation scope. Keep in sync with pkg/services/control/service.proto.
type EvacuateScope uint32

var (
	EvacuateScopeObjects EvacuateScope = 1
	EvacuateScopeTrees   EvacuateScope = 2
)

func (s EvacuateScope) String() string {
	var sb strings.Builder
	first := true
	if s&EvacuateScopeObjects == EvacuateScopeObjects {
		if !first {
			sb.WriteString(";")
		}
		sb.WriteString("objects")
		first = false
	}
	if s&EvacuateScopeTrees == EvacuateScopeTrees {
		if !first {
			sb.WriteString(";")
		}
		sb.WriteString("trees")
	}
	return sb.String()
}

// EvacuateShardPrm represents parameters for the EvacuateShard operation.
type EvacuateShardPrm struct {
	ShardID        []*shard.ID
	ObjectsHandler func(context.Context, oid.Address, *objectSDK.Object) error
	IgnoreErrors   bool
	Async          bool
	Scope          EvacuateScope
}

// EvacuateShardRes represents result of the EvacuateShard operation.
type EvacuateShardRes struct {
	objEvacuated *atomic.Uint64
	objTotal     *atomic.Uint64
	objFailed    *atomic.Uint64
	objSkipped   *atomic.Uint64

	trEvacuated *atomic.Uint64
	trTotal     *atomic.Uint64
	trFailed    *atomic.Uint64
}

// NewEvacuateShardRes creates new EvacuateShardRes instance.
func NewEvacuateShardRes() *EvacuateShardRes {
	return &EvacuateShardRes{
		objEvacuated: new(atomic.Uint64),
		objTotal:     new(atomic.Uint64),
		objFailed:    new(atomic.Uint64),
		objSkipped:   new(atomic.Uint64),
		trEvacuated:  new(atomic.Uint64),
		trTotal:      new(atomic.Uint64),
		trFailed:     new(atomic.Uint64),
	}
}

// ObjectsEvacuated returns amount of evacuated objects.
// Objects for which handler returned no error are also assumed evacuated.
func (p *EvacuateShardRes) ObjectsEvacuated() uint64 {
	if p == nil {
		return 0
	}
	return p.objEvacuated.Load()
}

// ObjectsTotal returns total count objects to evacuate.
func (p *EvacuateShardRes) ObjectsTotal() uint64 {
	if p == nil {
		return 0
	}
	return p.objTotal.Load()
}

// ObjectsFailed returns count of failed objects to evacuate.
func (p *EvacuateShardRes) ObjectsFailed() uint64 {
	if p == nil {
		return 0
	}
	return p.objFailed.Load()
}

// ObjectsSkipped returns count of skipped objects.
func (p *EvacuateShardRes) ObjectsSkipped() uint64 {
	if p == nil {
		return 0
	}
	return p.objSkipped.Load()
}

// TreesEvacuated returns amount of evacuated trees.
func (p *EvacuateShardRes) TreesEvacuated() uint64 {
	if p == nil {
		return 0
	}
	return p.trEvacuated.Load()
}

// TreesTotal returns total count trees to evacuate.
func (p *EvacuateShardRes) TreesTotal() uint64 {
	if p == nil {
		return 0
	}
	return p.trTotal.Load()
}

// TreesFailed returns count of failed trees to evacuate.
func (p *EvacuateShardRes) TreesFailed() uint64 {
	if p == nil {
		return 0
	}
	return p.trFailed.Load()
}

// DeepCopy returns deep copy of result instance.
func (p *EvacuateShardRes) DeepCopy() *EvacuateShardRes {
	if p == nil {
		return nil
	}

	res := &EvacuateShardRes{
		objEvacuated: new(atomic.Uint64),
		objTotal:     new(atomic.Uint64),
		objFailed:    new(atomic.Uint64),
		objSkipped:   new(atomic.Uint64),
		trEvacuated:  new(atomic.Uint64),
		trTotal:      new(atomic.Uint64),
		trFailed:     new(atomic.Uint64),
	}

	res.objEvacuated.Store(p.objEvacuated.Load())
	res.objTotal.Store(p.objTotal.Load())
	res.objFailed.Store(p.objFailed.Load())
	res.objSkipped.Store(p.objSkipped.Load())
	res.trTotal.Store(p.trTotal.Load())
	res.trEvacuated.Store(p.trEvacuated.Load())
	res.trFailed.Store(p.trFailed.Load())
	return res
}

const defaultEvacuateBatchSize = 100

type pooledShard struct {
	hashedShard
	pool util.WorkerPool
}

var errMustHaveTwoShards = errors.New("must have at least 1 spare shard")

// Evacuate moves data from one shard to the others.
// The shard being moved must be in read-only mode.
func (e *StorageEngine) Evacuate(ctx context.Context, prm EvacuateShardPrm) (*EvacuateShardRes, error) {
	select {
	case <-ctx.Done():
		return nil, ctx.Err()
	default:
	}

	shardIDs := make([]string, len(prm.ShardID))
	for i := range prm.ShardID {
		shardIDs[i] = prm.ShardID[i].String()
	}

	ctx, span := tracing.StartSpanFromContext(ctx, "StorageEngine.Evacuate",
		trace.WithAttributes(
			attribute.StringSlice("shardIDs", shardIDs),
			attribute.Bool("async", prm.Async),
			attribute.Bool("ignoreErrors", prm.IgnoreErrors),
			attribute.Stringer("scope", prm.Scope),
		))
	defer span.End()

	shards, weights, err := e.getActualShards(shardIDs, prm)
	if err != nil {
		return nil, err
	}

	shardsToEvacuate := make(map[string]*shard.Shard)
	for i := range shardIDs {
		for j := range shards {
			if shards[j].ID().String() == shardIDs[i] {
				shardsToEvacuate[shardIDs[i]] = shards[j].Shard
			}
		}
	}

	res := NewEvacuateShardRes()
	ctx = ctxOrBackground(ctx, prm.Async)
	eg, egCtx, err := e.evacuateLimiter.TryStart(ctx, shardIDs, res)
	if err != nil {
		return nil, err
	}

	eg.Go(func() error {
		return e.evacuateShards(egCtx, shardIDs, prm, res, shards, weights, shardsToEvacuate)
	})

	if prm.Async {
		return nil, nil
	}

	return res, eg.Wait()
}

func ctxOrBackground(ctx context.Context, background bool) context.Context {
	if background {
		return context.Background()
	}
	return ctx
}

func (e *StorageEngine) evacuateShards(ctx context.Context, shardIDs []string, prm EvacuateShardPrm, res *EvacuateShardRes,
	shards []pooledShard, weights []float64, shardsToEvacuate map[string]*shard.Shard,
) error {
	var err error
	ctx, span := tracing.StartSpanFromContext(ctx, "StorageEngine.evacuateShards",
		trace.WithAttributes(
			attribute.StringSlice("shardIDs", shardIDs),
			attribute.Bool("async", prm.Async),
			attribute.Bool("ignoreErrors", prm.IgnoreErrors),
			attribute.Stringer("scope", prm.Scope),
		))

	defer func() {
		span.End()
		e.evacuateLimiter.Complete(err)
	}()

	e.log.Info(logs.EngineStartedShardsEvacuation, zap.Strings("shard_ids", shardIDs), evacuationOperationLogField,
		zap.String("trace_id", tracingPkg.GetTraceID(ctx)), zap.Stringer("scope", prm.Scope))

	err = e.getTotalObjectsCount(ctx, shardsToEvacuate, res)
	if err != nil {
		e.log.Error(logs.EngineShardsEvacuationFailedToCount, zap.Strings("shard_ids", shardIDs), zap.Error(err), evacuationOperationLogField,
			zap.String("trace_id", tracingPkg.GetTraceID(ctx)), zap.Stringer("scope", prm.Scope))
		return err
	}

	for _, shardID := range shardIDs {
		if err = e.evacuateShard(ctx, shardID, prm, res, shards, weights, shardsToEvacuate); err != nil {
			e.log.Error(logs.EngineFinishedWithErrorShardsEvacuation, zap.Error(err), zap.Strings("shard_ids", shardIDs), evacuationOperationLogField,
				zap.String("trace_id", tracingPkg.GetTraceID(ctx)), zap.Stringer("scope", prm.Scope))
			return err
		}
	}

	e.log.Info(logs.EngineFinishedSuccessfullyShardsEvacuation,
		zap.Strings("shard_ids", shardIDs),
		evacuationOperationLogField,
		zap.Uint64("total_objects", res.ObjectsTotal()),
		zap.Uint64("evacuated_objects", res.ObjectsEvacuated()),
		zap.Uint64("failed_objects", res.ObjectsFailed()),
		zap.Uint64("skipped_objects", res.ObjectsSkipped()),
		zap.Uint64("total_trees", res.TreesTotal()),
		zap.Uint64("evacuated_trees", res.TreesEvacuated()),
		zap.Uint64("failed_trees", res.TreesFailed()),
	)
	return nil
}

func (e *StorageEngine) getTotalObjectsCount(ctx context.Context, shardsToEvacuate map[string]*shard.Shard, res *EvacuateShardRes) error {
	ctx, span := tracing.StartSpanFromContext(ctx, "StorageEngine.getTotalObjectsCount")
	defer span.End()

	for _, sh := range shardsToEvacuate {
		cnt, err := sh.LogicalObjectsCount(ctx)
		if err != nil {
			if errors.Is(err, shard.ErrDegradedMode) {
				continue
			}
			return err
		}
		res.objTotal.Add(cnt)
	}
	return nil
}

func (e *StorageEngine) evacuateShard(ctx context.Context, shardID string, prm EvacuateShardPrm, res *EvacuateShardRes,
	shards []pooledShard, weights []float64, shardsToEvacuate map[string]*shard.Shard,
) error {
	ctx, span := tracing.StartSpanFromContext(ctx, "StorageEngine.evacuateShard",
		trace.WithAttributes(
			attribute.String("shardID", shardID),
		))
	defer span.End()

	var listPrm shard.ListWithCursorPrm
	listPrm.WithCount(defaultEvacuateBatchSize)

	sh := shardsToEvacuate[shardID]

	var c *meta.Cursor
	for {
		listPrm.WithCursor(c)

		// TODO (@fyrchik): #1731 this approach doesn't work in degraded modes
		//  because ListWithCursor works only with the metabase.
		listRes, err := sh.ListWithCursor(ctx, listPrm)
		if err != nil {
			if errors.Is(err, meta.ErrEndOfListing) || errors.Is(err, shard.ErrDegradedMode) {
				break
			}
			e.log.Error(logs.EngineShardsEvacuationFailedToListObjects, zap.String("shard_id", shardID), zap.Error(err), evacuationOperationLogField,
				zap.String("trace_id", tracingPkg.GetTraceID(ctx)))
			return err
		}

		if err = e.evacuateObjects(ctx, sh, listRes.AddressList(), prm, res, shards, weights, shardsToEvacuate); err != nil {
			return err
		}

		c = listRes.Cursor()
	}
	return nil
}

func (e *StorageEngine) getActualShards(shardIDs []string, prm EvacuateShardPrm) ([]pooledShard, []float64, error) {
	e.mtx.RLock()
	defer e.mtx.RUnlock()

	for i := range shardIDs {
		sh, ok := e.shards[shardIDs[i]]
		if !ok {
			return nil, nil, errShardNotFound
		}

		if !sh.GetMode().ReadOnly() {
			return nil, nil, ErrMustBeReadOnly
		}
	}

	if len(e.shards)-len(shardIDs) < 1 && prm.ObjectsHandler == nil {
		return nil, nil, errMustHaveTwoShards
	}

	// We must have all shards, to have correct information about their
	// indexes in a sorted slice and set appropriate marks in the metabase.
	// Evacuated shard is skipped during put.
	shards := make([]pooledShard, 0, len(e.shards))
	for id := range e.shards {
		shards = append(shards, pooledShard{
			hashedShard: hashedShard(e.shards[id]),
			pool:        e.shardPools[id],
		})
	}

	weights := make([]float64, 0, len(shards))
	for i := range shards {
		weights = append(weights, e.shardWeight(shards[i].Shard))
	}

	return shards, weights, nil
}

func (e *StorageEngine) evacuateObjects(ctx context.Context, sh *shard.Shard, toEvacuate []object.AddressWithType, prm EvacuateShardPrm, res *EvacuateShardRes,
	shards []pooledShard, weights []float64, shardsToEvacuate map[string]*shard.Shard,
) error {
	ctx, span := tracing.StartSpanFromContext(ctx, "StorageEngine.evacuateObjects",
		trace.WithAttributes(
			attribute.Int("objects_count", len(toEvacuate)),
		))
	defer span.End()

	for i := range toEvacuate {
		select {
		case <-ctx.Done():
			return ctx.Err()
		default:
		}
		addr := toEvacuate[i].Address

		var getPrm shard.GetPrm
		getPrm.SetAddress(addr)

		getRes, err := sh.Get(ctx, getPrm)
		if err != nil {
			if prm.IgnoreErrors {
				res.objFailed.Add(1)
				continue
			}
			e.log.Error(logs.EngineShardsEvacuationFailedToReadObject, zap.String("address", addr.EncodeToString()), zap.Error(err), evacuationOperationLogField,
				zap.String("trace_id", tracingPkg.GetTraceID(ctx)))
			return err
		}

		evacuatedLocal, err := e.tryEvacuateObjectLocal(ctx, addr, getRes.Object(), sh, shards, weights, shardsToEvacuate, res)
		if err != nil {
			return err
		}

		if evacuatedLocal {
			continue
		}

		if prm.ObjectsHandler == nil {
			// Do not check ignoreErrors flag here because
			// ignoring errors on put make this command kinda useless.
			return fmt.Errorf("%w: %s", errPutShard, toEvacuate[i])
		}

		err = prm.ObjectsHandler(ctx, addr, getRes.Object())
		if err != nil {
			e.log.Error(logs.EngineShardsEvacuationFailedToMoveObject, zap.String("address", addr.EncodeToString()), zap.Error(err), evacuationOperationLogField,
				zap.String("trace_id", tracingPkg.GetTraceID(ctx)))
			return err
		}
		res.objEvacuated.Add(1)
	}
	return nil
}

func (e *StorageEngine) tryEvacuateObjectLocal(ctx context.Context, addr oid.Address, object *objectSDK.Object, sh *shard.Shard,
	shards []pooledShard, weights []float64, shardsToEvacuate map[string]*shard.Shard, res *EvacuateShardRes,
) (bool, error) {
	hrw.SortHasherSliceByWeightValue(shards, weights, hrw.StringHash(addr.EncodeToString()))
	for j := range shards {
		select {
		case <-ctx.Done():
			return false, ctx.Err()
		default:
		}

		if _, ok := shardsToEvacuate[shards[j].ID().String()]; ok {
			continue
		}
		switch e.putToShard(ctx, shards[j].hashedShard, j, shards[j].pool, addr, object).status {
		case putToShardSuccess:
			res.objEvacuated.Add(1)
			e.log.Debug(logs.EngineObjectIsMovedToAnotherShard,
				zap.Stringer("from", sh.ID()),
				zap.Stringer("to", shards[j].ID()),
				zap.Stringer("addr", addr),
				evacuationOperationLogField,
				zap.String("trace_id", tracingPkg.GetTraceID(ctx)))
			return true, nil
		case putToShardExists, putToShardRemoved:
			res.objSkipped.Add(1)
			return true, nil
		default:
			continue
		}
	}

	return false, nil
}

func (e *StorageEngine) GetEvacuationState(ctx context.Context) (*EvacuationState, error) {
	select {
	case <-ctx.Done():
		return nil, ctx.Err()
	default:
	}

	return e.evacuateLimiter.GetState(), nil
}

func (e *StorageEngine) EnqueRunningEvacuationStop(ctx context.Context) error {
	select {
	case <-ctx.Done():
		return ctx.Err()
	default:
	}

	return e.evacuateLimiter.CancelIfRunning()
}