storj/satellite/metabase/segmentloop/service.go

// Copyright (C) 2021 Storj Labs, Inc.
// See LICENSE for copying information.

package segmentloop

import (
	"context"
	"errors"
	"fmt"
	"time"

	"github.com/spacemonkeygo/monkit/v3"
	"github.com/zeebo/errs"
	"go.uber.org/zap"
	"golang.org/x/time/rate"

	"storj.io/common/errs2"
	"storj.io/common/sync2"
	"storj.io/storj/satellite/metabase"
)

const batchsizeLimit = 5000

var (
	mon = monkit.Package()

	// Error is a standard error class for this component.
	Error = errs.Class("segments loop")
	// ErrClosed is a loop closed error.
	ErrClosed = Error.New("loop closed")
)

// Segment contains information about segment metadata which will be received by observers.
type Segment metabase.LoopSegmentEntry

// Inline returns true if segment is inline.
func (s Segment) Inline() bool {
	return s.Redundancy.IsZero() && len(s.Pieces) == 0
}

// Expired checks if segment expired relative to now.
func (s *Segment) Expired(now time.Time) bool {
	return s.ExpiresAt != nil && s.ExpiresAt.Before(now)
}

// Observer is an interface defining an observer that can subscribe to the segments loop.
//
// architecture: Observer
type Observer interface {
	LoopStarted(context.Context, LoopInfo) error
	RemoteSegment(context.Context, *Segment) error
	InlineSegment(context.Context, *Segment) error
}

// LoopInfo contains information about the current loop.
type LoopInfo struct {
	Started time.Time
}

// NullObserver is an observer that does nothing. This is useful for joining
// and ensuring the segments loop runs once before you use a real observer.
type NullObserver struct{}

// LoopStarted is called at each loop start.
func (NullObserver) LoopStarted(context.Context, LoopInfo) error {
	return nil
}

// RemoteSegment implements the Observer interface.
func (NullObserver) RemoteSegment(context.Context, *Segment) error {
	return nil
}

// InlineSegment implements the Observer interface.
func (NullObserver) InlineSegment(context.Context, *Segment) error {
	return nil
}

type observerContext struct {
	immediate bool
	trigger   bool
	observer  Observer

	ctx  context.Context
	done chan error

	remote *monkit.DurationDist
	inline *monkit.DurationDist
}

func newObserverContext(ctx context.Context, obs Observer) *observerContext {
	name := fmt.Sprintf("%T", obs)
	key := monkit.NewSeriesKey("observer").WithTag("name", name)

	return &observerContext{
		observer: obs,

		ctx:  ctx,
		done: make(chan error),

		inline: monkit.NewDurationDist(key.WithTag("pointer_type", "inline")),
		remote: monkit.NewDurationDist(key.WithTag("pointer_type", "remote")),
	}
}

func (observer *observerContext) RemoteSegment(ctx context.Context, segment *Segment) error {
	start := time.Now()
	defer func() { observer.remote.Insert(time.Since(start)) }()

	return observer.observer.RemoteSegment(ctx, segment)
}

func (observer *observerContext) InlineSegment(ctx context.Context, segment *Segment) error {
	start := time.Now()
	defer func() { observer.inline.Insert(time.Since(start)) }()

	return observer.observer.InlineSegment(ctx, segment)
}

func (observer *observerContext) HandleError(err error) bool {
	if err != nil {
		observer.done <- err
		observer.Finish()
		return true
	}
	return false
}

func (observer *observerContext) Finish() {
	close(observer.done)

	name := fmt.Sprintf("%T", observer.observer)
	stats := allObserverStatsCollectors.GetStats(name)
	stats.Observe(observer)
}

func (observer *observerContext) Wait() error {
	return <-observer.done
}

// Config contains configurable values for the segments loop.
type Config struct {
	CoalesceDuration time.Duration `help:"how long to wait for new observers before starting iteration" releaseDefault:"5s" devDefault:"5s" testDefault:"1s"`
	RateLimit        float64       `help:"rate limit (default is 0 which is unlimited segments per second)" default:"0"`
	ListLimit        int           `help:"how many items to query in a batch" default:"2500"`

	AsOfSystemInterval time.Duration `help:"as of system interval" releaseDefault:"-5m" devDefault:"-1us" testDefault:"-1us"`

	SuspiciousProcessedRatio float64 `help:"ratio where to consider processed count as supicious" default:"0.03"`
}

// MetabaseDB contains iterators for the metabase data.
type MetabaseDB interface {
	// Now returns the time on the database.
	Now(ctx context.Context) (time.Time, error)
	// IterateLoopStreams iterates through all streams passed in as arguments.
	IterateLoopSegments(ctx context.Context, opts metabase.IterateLoopSegments, fn func(context.Context, metabase.LoopSegmentsIterator) error) (err error)

	// GetTableStats gathers statistics about the tables.
	GetTableStats(context.Context, metabase.GetTableStats) (metabase.TableStats, error)
}

// Service is a segments loop service.
//
// architecture: Service
type Service struct {
	log        *zap.Logger
	config     Config
	metabaseDB MetabaseDB
	join       chan *observerContext
	done       chan struct{}
}

// New creates a new segments loop service.
func New(log *zap.Logger, config Config, metabaseDB MetabaseDB) *Service {
	return &Service{
		log:        log,
		metabaseDB: metabaseDB,
		config:     config,
		join:       make(chan *observerContext),
		done:       make(chan struct{}),
	}
}

// Join will join the looper for one full cycle until completion and then returns.
// Joining will trigger a new iteration after coalesce duration.
// On ctx cancel the observer will return without completely finishing.
// Only on full complete iteration it will return nil.
// Safe to be called concurrently.
func (loop *Service) Join(ctx context.Context, observer Observer) (err error) {
	return loop.joinObserver(ctx, true, observer)
}

// Monitor will join the looper for one full cycle until completion and then returns.
// Joining with monitoring won't trigger after coalesce duration.
// On ctx cancel the observer will return without completely finishing.
// Only on full complete iteration it will return nil.
// Safe to be called concurrently.
func (loop *Service) Monitor(ctx context.Context, observer Observer) (err error) {
	return loop.joinObserver(ctx, false, observer)
}

// joinObserver will join the looper for one full cycle until completion and then returns.
// On ctx cancel the observer will return without completely finishing.
// Only on full complete iteration it will return nil.
// Safe to be called concurrently.
func (loop *Service) joinObserver(ctx context.Context, trigger bool, obs Observer) (err error) {
	defer mon.Task()(&ctx)(&err)

	obsctx := newObserverContext(ctx, obs)
	obsctx.immediate = sync2.IsManuallyTriggeredCycle(ctx)
	obsctx.trigger = trigger || obsctx.immediate

	select {
	case loop.join <- obsctx:
	case <-ctx.Done():
		return ctx.Err()
	case <-loop.done:
		return ErrClosed
	}

	return obsctx.Wait()
}

// Run starts the looping service.
// It can only be called once, otherwise a panic will occur.
func (loop *Service) Run(ctx context.Context) (err error) {
	defer mon.Task()(&ctx)(&err)

	for {
		err := loop.RunOnce(ctx)
		if err != nil {
			loop.log.Error("segment loop failure", zap.Error(err))

			if errs2.IsCanceled(err) {
				return err
			}
			if ctx.Err() != nil {
				return errs.Combine(err, ctx.Err())
			}

			mon.Event("segmentloop_error") //mon:locked
		}
	}
}

// Close closes the looping services.
func (loop *Service) Close() (err error) {
	close(loop.done)
	return nil
}

// RunOnce goes through segments one time and sends information to observers.
//
// It is not safe to call this concurrently with Run.
func (loop *Service) RunOnce(ctx context.Context) (err error) {
	defer mon.Task()(&ctx)(&err) //mon:locked

	coalesceTimer := time.NewTimer(loop.config.CoalesceDuration)
	defer coalesceTimer.Stop()
	stopTimer(coalesceTimer)

	earlyExit := make(chan *observerContext)
	earlyExitDone := make(chan struct{})
	monitorEarlyExit := func(obs *observerContext) {
		select {
		case <-obs.ctx.Done():
			select {
			case <-earlyExitDone:
			case earlyExit <- obs:
			}
		case <-earlyExitDone:
		}
	}

	timerStarted := false
	observers := []*observerContext{}

waitformore:
	for {
		select {
		// when the coalesce timer hits, we have waited enough for observers to join.
		case <-coalesceTimer.C:
			break waitformore

		// wait for a new observer to join.
		case obsctx := <-loop.join:
			// when the observer triggers the loop and it's the first one,
			// then start the coalescing timer.
			if obsctx.trigger {
				if !timerStarted {
					coalesceTimer.Reset(loop.config.CoalesceDuration)
					timerStarted = true
				}
			}

			observers = append(observers, obsctx)
			go monitorEarlyExit(obsctx)

			if obsctx.immediate {
				break waitformore
			}

		// remove an observer from waiting when it's canceled before the loop starts.
		case obsctx := <-earlyExit:
			for i, obs := range observers {
				if obs == obsctx {
					observers = append(observers[:i], observers[i+1:]...)
					break
				}
			}

			obsctx.HandleError(obsctx.ctx.Err())

			// reevalute, whether we acually need to start the loop.
			timerShouldRun := false
			for _, obs := range observers {
				timerShouldRun = timerShouldRun || obs.trigger
			}

			if !timerShouldRun && timerStarted {
				stopTimer(coalesceTimer)
			}

		// when ctx done happens we can finish all the waiting observers.
		case <-ctx.Done():
			close(earlyExitDone)
			errorObservers(observers, ctx.Err())
			return ctx.Err()
		}
	}
	close(earlyExitDone)

	return loop.iterateDatabase(ctx, observers)
}

func stopTimer(t *time.Timer) {
	t.Stop()
	// drain if it contains something
	select {
	case <-t.C:
	default:
	}
}

// Wait waits for run to be finished.
// Safe to be called concurrently.
func (loop *Service) Wait() {
	<-loop.done
}

var errNoObservers = errs.New("no observers")

func (loop *Service) iterateDatabase(ctx context.Context, observers []*observerContext) (err error) {
	defer mon.Task()(&ctx)(&err)

	defer func() {
		if err != nil {
			errorObservers(observers, err)
			return
		}
		finishObservers(observers)
	}()

	before, err := loop.metabaseDB.GetTableStats(ctx, metabase.GetTableStats{
		AsOfSystemInterval: loop.config.AsOfSystemInterval,
	})
	if err != nil {
		return Error.Wrap(err)
	}

	var processed processedStats
	processed, observers, err = loop.iterateSegments(ctx, observers)
	if errors.Is(err, errNoObservers) {
		return nil
	}
	if err != nil {
		return Error.Wrap(err)
	}

	after, err := loop.metabaseDB.GetTableStats(ctx, metabase.GetTableStats{
		AsOfSystemInterval: loop.config.AsOfSystemInterval,
	})
	if err != nil {
		return Error.Wrap(err)
	}

	if err := loop.verifyCount(before.SegmentCount, after.SegmentCount, processed.segments); err != nil {
		return Error.Wrap(err)
	}

	return err
}

func (loop *Service) verifyCount(before, after, processed int64) error {
	low, high := before, after
	if low > high {
		low, high = high, low
	}

	var deltaFromBounds int64
	var ratio float64
	if processed < low {
		deltaFromBounds = low - processed
		// +1 to avoid division by zero
		ratio = float64(deltaFromBounds) / float64(low+1)
	} else if processed > high {
		deltaFromBounds = processed - high
		// +1 to avoid division by zero
		ratio = float64(deltaFromBounds) / float64(high+1)
	}

	mon.IntVal("segmentloop_verify_before").Observe(before)
	mon.IntVal("segmentloop_verify_after").Observe(after)
	mon.IntVal("segmentloop_verify_processed").Observe(processed)
	mon.IntVal("segmentloop_verify_outside").Observe(deltaFromBounds)
	mon.FloatVal("segmentloop_verify_outside_ratio").Observe(ratio)

	// If we have very few items from the bounds, then it's expected and the ratio does not capture it well.
	const minimumDeltaThreshold = 100
	if deltaFromBounds < minimumDeltaThreshold {
		return nil
	}

	if ratio > loop.config.SuspiciousProcessedRatio {
		return Error.New("processed count looks suspicious: before:%v after:%v processed:%v ratio:%v threshold:%v", before, after, processed, ratio, loop.config.SuspiciousProcessedRatio)
	}

	return nil
}

type processedStats struct {
	segments int64
}

func (loop *Service) iterateSegments(ctx context.Context, observers []*observerContext) (processed processedStats, _ []*observerContext, err error) {
	defer mon.Task()(&ctx)(&err)

	rateLimiter := rate.NewLimiter(rate.Limit(loop.config.RateLimit), 1)

	if loop.config.RateLimit == 0 {
		rateLimiter = rate.NewLimiter(rate.Inf, 1)
	}

	limit := loop.config.ListLimit
	if limit <= 0 || limit > batchsizeLimit {
		limit = batchsizeLimit
	}

	startingTime, err := loop.metabaseDB.Now(ctx)
	if err != nil {
		return processed, observers, Error.Wrap(err)
	}

	observers = withObservers(ctx, observers, func(ctx context.Context, observer *observerContext) bool {
		err := observer.observer.LoopStarted(ctx, LoopInfo{Started: startingTime})
		return !observer.HandleError(err)
	})

	if len(observers) == 0 {
		return processed, observers, errNoObservers
	}

	err = loop.metabaseDB.IterateLoopSegments(ctx, metabase.IterateLoopSegments{
		BatchSize:          limit,
		AsOfSystemTime:     startingTime,
		AsOfSystemInterval: loop.config.AsOfSystemInterval,
	}, func(ctx context.Context, iterator metabase.LoopSegmentsIterator) error {
		defer mon.TaskNamed("iterateLoopSegmentsCB")(&ctx)(&err)

		var entry metabase.LoopSegmentEntry
		for iterator.Next(ctx, &entry) {
			if err := ctx.Err(); err != nil {
				return err
			}

			timer := mon.Timer("iterateLoopSegmentsRateLimit").Start()
			if err := rateLimiter.Wait(ctx); err != nil {
				// We don't really execute concurrent batches so we should never
				// exceed the burst size of 1 and this should never happen.
				// We can also enter here if the context is cancelled.
				timer.Stop()
				return err
			}
			timer.Stop()

			observers = withObservers(ctx, observers, func(ctx context.Context, observer *observerContext) bool {
				segment := Segment(entry)
				return !observer.HandleError(handleSegment(ctx, observer, &segment))
			})
			if len(observers) == 0 {
				return errNoObservers
			}

			processed.segments++
			mon.IntVal("segmentsProcessed").Observe(processed.segments) //mon:locked
		}
		return nil
	})

	return processed, observers, err
}

func withObservers(ctx context.Context, observers []*observerContext, handleObserver func(ctx context.Context, observer *observerContext) bool) []*observerContext {
	defer mon.Task()(&ctx)(nil)
	nextObservers := observers[:0]
	for _, observer := range observers {
		keepObserver := handleObserver(ctx, observer)
		if keepObserver {
			nextObservers = append(nextObservers, observer)
		}
	}
	return nextObservers
}

func handleSegment(ctx context.Context, observer *observerContext, segment *Segment) (err error) {
	defer mon.Task()(&ctx)(&err)

	if segment.Inline() {
		if err := observer.InlineSegment(ctx, segment); err != nil {
			return err
		}
	} else {
		if err := observer.RemoteSegment(ctx, segment); err != nil {
			return err
		}
	}

	return observer.ctx.Err()
}

func finishObservers(observers []*observerContext) {
	for _, observer := range observers {
		observer.Finish()
	}
}

func errorObservers(observers []*observerContext, err error) {
	for _, observer := range observers {
		observer.HandleError(err)
	}
}