shard_buffer.go

/*
Copyright 2019 The Vitess Authors.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/

package buffer

import (
	"context"
	"fmt"
	"runtime/debug"
	"sync"
	"sync/atomic"
	"time"

	"vitess.io/vitess/go/vt/discovery"

	"vitess.io/vitess/go/vt/vtgate/errorsanitizer"

	"vitess.io/vitess/go/vt/log"
	"vitess.io/vitess/go/vt/logutil"
	topodatapb "vitess.io/vitess/go/vt/proto/topodata"
	"vitess.io/vitess/go/vt/topo/topoproto"
	"vitess.io/vitess/go/vt/vterrors"
)

// bufferState represents the different states a shardBuffer object can be in.
type bufferState string

const (
	// stateIdle means no failover is currently in progress.
	stateIdle bufferState = "IDLE"
	// stateBuffering is the phase when a failover is in progress.
	stateBuffering bufferState = "BUFFERING"
	// stateDraining is the phase when a failover ended and the queue is drained.
	stateDraining bufferState = "DRAINING"
)

// entry is created per buffered request.
type entry struct {
	// done will be closed by shardBuffer when the failover is over and the
	// request can be retried.
	// Any Go routine closing this channel must also remove the entry from the
	// ShardBuffer queue such that nobody else tries to close it.
	done chan struct{}

	// deadline is the time when the entry is out of the buffering window and it
	// must be canceled.
	deadline time.Time

	// err is set if the buffering failed e.g. when the entry was evicted.
	err error

	// bufferCtx wraps the request ctx and is used to track the retry of a
	// request during the drain phase. Once the retry is done, the caller
	// must cancel this context (by calling bufferCancel).
	bufferCtx    context.Context
	bufferCancel func()
}

// shardBuffer buffers requests during a failover for a particular shard.
// The object will be reused across failovers. If no failover is currently in
// progress, the state is "IDLE".
//
// Note that this object is accessed concurrently by multiple threads:
// - vtgate request threads
// - discovery.HealthCheck listener execution thread
// - timeout thread (timeout_thread.go) to evict too old buffered requests
// - drain() thread
type shardBuffer struct {
	// Immutable fields set at construction.
	buf      *Buffer
	mode     bufferMode
	keyspace string
	shard    string

	// statsKey is used to update the stats variables.
	statsKey []string
	// statsKeyJoined is all elements of "statsKey" in one string, joined by ".".
	statsKeyJoined string
	logTooRecent   *logutil.ThrottledLogger

	// mu guards the fields below.
	mu    sync.RWMutex
	state bufferState
	// queue is the list of buffered requests (ordered by arrival).
	queue []*entry
	// lastStart is the last time we saw the start of a failover.
	lastStart time.Time
	// lastEnd is the last time we saw the end of a failover.
	lastEnd time.Time
	// lastReparent is the last time we saw that the tablet alias of the PRIMARY
	// changed i.e. we definitely reparented to a different tablet.
	lastReparent time.Time
	// currentPrimary is tracked to determine when to update "lastReparent".
	currentPrimary *topodatapb.TabletAlias
	// timeoutThread will be set while a failover is in progress and the object is
	// in the BUFFERING state.
	timeoutThread *timeoutThread
	// wg tracks all pending Go routines. waitForShutdown() will use this field to
	// block on them.
	wg sync.WaitGroup
}

func newShardBufferHealthCheck(buf *Buffer, mode bufferMode, keyspace, shard string) *shardBuffer {
	statsKey := []string{keyspace, shard}
	initVariablesForShard(statsKey)

	return &shardBuffer{
		buf:            buf,
		mode:           mode,
		keyspace:       keyspace,
		shard:          shard,
		statsKey:       statsKey,
		statsKeyJoined: fmt.Sprintf("%s.%s", keyspace, shard),
		logTooRecent:   logutil.NewThrottledLogger(fmt.Sprintf("FailoverTooRecent-%v", topoproto.KeyspaceShardString(keyspace, shard)), 5*time.Second),
		state:          stateIdle,
	}
}

func (sb *shardBuffer) timeNow() time.Time {
	return sb.buf.config.now()
}

// disabled returns true if neither buffering nor the dry-run mode is enabled.
func (sb *shardBuffer) disabled() bool {
	return sb.mode == bufferModeDisabled
}

func (sb *shardBuffer) waitForFailoverEnd(ctx context.Context, keyspace, shard string, err error) (RetryDoneFunc, error) {
	// We assume if err != nil then it's always caused by a failover.
	// Other errors must be filtered at higher layers.
	failoverDetected := err != nil

	// Fast path (read lock): Check if we should NOT buffer a request.
	sb.mu.RLock()
	if !sb.shouldBufferLocked(failoverDetected) {
		// No buffering required. Return early.
		sb.mu.RUnlock()
		return nil, nil
	}
	sb.mu.RUnlock()

	// Buffering required. Acquire write lock.
	sb.mu.Lock()
	// Re-check state because it could have changed in the meantime.
	if !sb.shouldBufferLocked(failoverDetected) {
		// Buffering no longer required. Return early.
		sb.mu.Unlock()
		return nil, nil
	}

	// Start buffering if failover is not detected yet.
	if sb.state == stateIdle {
		// Do not buffer if last failover is too recent. This is the case if:
		// a) buffering was stopped recently
		// OR
		// b) we did not buffer, but observed a reparent very recently
		now := sb.timeNow()
		minTimeBetweenFailovers := sb.buf.config.MinTimeBetweenFailovers

		// a) Buffering was stopped recently.
		// This can happen when we stop buffering while MySQL is not ready yet
		// (read-only mode is not cleared yet on the new primary).
		lastBufferingStopped := now.Sub(sb.lastEnd)
		if !sb.lastEnd.IsZero() && lastBufferingStopped < minTimeBetweenFailovers {
			sb.mu.Unlock()
			msg := "NOT starting buffering"
			if sb.mode == bufferModeDryRun {
				msg = "Dry-run: Would NOT have started buffering"
			}

			sb.logTooRecent.Infof("%v for shard: %s because the last failover which triggered buffering is too recent (%v < %v)."+
				" (A failover was detected by this seen error: %v.)",
				msg, topoproto.KeyspaceShardString(keyspace, shard), lastBufferingStopped, minTimeBetweenFailovers, err)

			statsKeyWithReason := append(sb.statsKey, string(skippedLastFailoverTooRecent))
			requestsSkipped.Add(statsKeyWithReason, 1)
			return nil, nil
		}

		// b) The PRIMARY was reparented recently (but we did not buffer it.)
		// This can happen when we see the end of the reparent *before* the first
		// request failure caused by the reparent. This is possible if the QPS is
		// very low. If we do not skip buffering here, we would start buffering but
		// not stop because we already observed the promotion of the new primary.
		lastReparentAgo := now.Sub(sb.lastReparent)
		if !sb.lastReparent.IsZero() && lastReparentAgo < minTimeBetweenFailovers {
			sb.mu.Unlock()
			msg := "NOT starting buffering"
			if sb.mode == bufferModeDryRun {
				msg = "Dry-run: Would NOT have started buffering"
			}

			sb.logTooRecent.Infof("%v for shard: %s because the last reparent is too recent (%v < %v)."+
				" (A failover was detected by this seen error: %v.)",
				msg, topoproto.KeyspaceShardString(keyspace, shard), lastReparentAgo, minTimeBetweenFailovers, err)

			statsKeyWithReason := append(sb.statsKey, string(skippedLastReparentTooRecent))
			requestsSkipped.Add(statsKeyWithReason, 1)
			return nil, nil
		}

		sb.startBufferingLocked(err)
	}

	if sb.mode == bufferModeDryRun {
		sb.mu.Unlock()
		// Dry-run. Do not actually buffer the request and return early.
		lastRequestsDryRunMax.Add(sb.statsKey, 1)
		requestsBufferedDryRun.Add(sb.statsKey, 1)
		return nil, nil
	}

	// Buffer request.
	entry, err := sb.bufferRequestLocked(ctx)
	sb.mu.Unlock()
	if err != nil {
		return nil, err
	}
	return sb.wait(ctx, entry)
}

// shouldBufferLocked returns true if the current request should be buffered
// (based on the current state and whether the request detected a failover).
func (sb *shardBuffer) shouldBufferLocked(failoverDetected bool) bool {
	switch s := sb.state; {
	case s == stateIdle && !failoverDetected:
		// No failover in progress.
		return false
	case s == stateIdle && failoverDetected:
		// Not buffering yet, but new failover detected.
		return true
	case s == stateBuffering:
		// Failover in progress.
		return true
	case s == stateDraining && !failoverDetected:
		// Draining. Non-failover related requests can pass through.
		return false
	case s == stateDraining && failoverDetected:
		// Possible race between request which saw failover-related error and the
		// end of the failover. Do not buffer and let vtgate retry immediately.
		return false
	}
	panic("BUG: All possible states must be covered by the switch expression above.")
}

func (sb *shardBuffer) startBufferingLocked(err error) {
	// Reset monitoring data from previous failover.
	lastRequestsInFlightMax.Set(sb.statsKey, 0)
	lastRequestsDryRunMax.Set(sb.statsKey, 0)
	failoverDurationSumMs.Reset(sb.statsKey)

	sb.lastStart = sb.timeNow()
	sb.logErrorIfStateNotLocked(stateIdle)
	sb.state = stateBuffering
	sb.queue = make([]*entry, 0)

	sb.timeoutThread = newTimeoutThread(sb, sb.buf.config.MaxFailoverDuration)
	sb.timeoutThread.start()
	msg := "Starting buffering"
	if sb.mode == bufferModeDryRun {
		msg = "Dry-run: Would have started buffering"
	}
	starts.Add(sb.statsKey, 1)
	log.Infof("%v for shard: %s (window: %v, size: %v, max failover duration: %v) (A failover was detected by this seen error: %v.)",
		msg,
		topoproto.KeyspaceShardString(sb.keyspace, sb.shard),
		sb.buf.config.Window,
		sb.buf.config.Size,
		sb.buf.config.MaxFailoverDuration,
		errorsanitizer.NormalizeError(err.Error()),
	)
}

// logErrorIfStateNotLocked logs an error if the current state is not "state".
// We do not panic/crash the process here because it is expected that a wrong
// state is less severe than (potentially) crash-looping all vtgates.
// Note: The prefix "Locked" is not related to the state. Instead, it stresses
// that "sb.mu" must be locked before calling the method.
func (sb *shardBuffer) logErrorIfStateNotLocked(state bufferState) {
	if sb.state != state {
		log.Errorf("BUG: Buffer state should be '%v' and not '%v'. Full state of buffer object: %#v Stacktrace:\n%s", state, sb.state, sb, debug.Stack())
	}
}

// bufferRequest creates a new entry in the queue for a request which
// should be buffered.
// It returns *entry which can be used as input for shardBuffer.cancel(). This
// is useful for canceled RPCs (e.g. due to deadline exceeded) which want to
// give up their spot in the buffer. It also holds the "bufferCancel" function.
// If buffering fails e.g. due to a full buffer, an error is returned.
func (sb *shardBuffer) bufferRequestLocked(ctx context.Context) (*entry, error) {
	if !sb.buf.bufferSizeSema.TryAcquire(1) {
		// Buffer is full. Evict the oldest entry and buffer this request instead.
		if len(sb.queue) == 0 {
			// Overall buffer is full, but this shard's queue is empty. That means
			// there is at least one other shard failing over as well which consumes
			// the whole buffer.
			statsKeyWithReason := append(sb.statsKey, string(skippedBufferFull))
			requestsSkipped.Add(statsKeyWithReason, 1)
			return nil, bufferFullError
		}

		e := sb.queue[0]
		// Evict the entry. Do not release its slot in the buffer and reuse it for
		// this new request.
		// NOTE: We keep the lock to avoid racing with drain().
		// NOTE: We're not waiting until the request finishes and instead reuse its
		// slot immediately, i.e. the number of evicted requests + drained requests
		// can be bigger than the buffer size.
		sb.unblockAndWait(e, entryEvictedError, false /* releaseSlot */, false /* blockingWait */)
		sb.queue = sb.queue[1:]
		statsKeyWithReason := append(sb.statsKey, evictedBufferFull)
		requestsEvicted.Add(statsKeyWithReason, 1)
	}

	e := &entry{
		done:     make(chan struct{}),
		deadline: sb.timeNow().Add(sb.buf.config.Window),
	}
	e.bufferCtx, e.bufferCancel = context.WithCancel(ctx)
	sb.queue = append(sb.queue, e)

	if max := lastRequestsInFlightMax.Counts()[sb.statsKeyJoined]; max < int64(len(sb.queue)) {
		lastRequestsInFlightMax.Set(sb.statsKey, int64(len(sb.queue)))
	}
	requestsBuffered.Add(sb.statsKey, 1)

	if len(sb.queue) == 1 {
		sb.timeoutThread.notifyQueueNotEmpty()
	}
	return e, nil
}

// unblockAndWait unblocks a blocked request.
// If releaseSlot is true, the buffer semaphore will be decreased by 1 when
// the request retried and finished.
// If blockingWait is true, this call will block until the request retried and
// finished. This mode is used during the drain (to avoid flooding the primary)
// while the non-blocking mode is used when a) evicting a request (e.g. because
// the buffer is full or it exceeded the buffering window) or b) when the
// request was canceled from outside and we removed it.
func (sb *shardBuffer) unblockAndWait(e *entry, err error, releaseSlot, blockingWait bool) {
	// Set error such that the request will see it.
	e.err = err
	// Tell blocked request to stop waiting.
	close(e.done)

	if blockingWait {
		sb.waitForRequestFinish(e, releaseSlot, false /* async */)
	} else {
		sb.wg.Add(1)
		go sb.waitForRequestFinish(e, releaseSlot, true /* async */)
	}
}

func (sb *shardBuffer) waitForRequestFinish(e *entry, releaseSlot, async bool) {
	if async {
		defer sb.wg.Done()
	}

	// Wait for unblocked request to finish.
	<-e.bufferCtx.Done()

	// Release the slot to the buffer.
	// NOTE: We always wait for the request first, even if the calling code like
	// the buffer full eviction or the timeout thread does not block on us.
	// This way, the request's slot can only be reused after the request finished.
	if releaseSlot {
		sb.buf.bufferSizeSema.Release(1)
	}
}

// wait blocks while the request is buffered during the failover.
// See Buffer.WaitForFailoverEnd() for the API contract of the return values.
func (sb *shardBuffer) wait(ctx context.Context, e *entry) (RetryDoneFunc, error) {
	select {
	case <-ctx.Done():
		sb.remove(e)
		return nil, vterrors.Errorf(vterrors.Code(contextCanceledError), "%v: %v", contextCanceledError, ctx.Err())
	case <-e.done:
		return e.bufferCancel, e.err
	}
}

// oldestEntry returns the head of the queue or nil if the queue is empty.
func (sb *shardBuffer) oldestEntry() *entry {
	sb.mu.Lock()
	defer sb.mu.Unlock()

	if len(sb.queue) > 0 {
		return sb.queue[0]
	}
	return nil
}

// evictOldestEntry is used by timeoutThread to evict the head entry of the
// queue if it exceeded its buffering window.
func (sb *shardBuffer) evictOldestEntry(e *entry) {
	sb.mu.Lock()
	defer sb.mu.Unlock()

	if len(sb.queue) == 0 || e != sb.queue[0] {
		// Entry is already removed e.g. by remove(). Ignore it.
		return
	}

	// Evict the entry.
	//
	// NOTE: We're not waiting for the request to finish in order to unblock the
	// timeout thread as fast as possible. However, the slot of the evicted
	// request is only returned after it has finished i.e. the buffer may stay
	// full in the meantime. This is a design tradeoff to keep things simple and
	// avoid additional pressure on the primary tablet.
	sb.unblockAndWait(e, nil /* err */, true /* releaseSlot */, false /* blockingWait */)
	sb.queue = sb.queue[1:]
	statsKeyWithReason := append(sb.statsKey, evictedWindowExceeded)
	requestsEvicted.Add(statsKeyWithReason, 1)
}

// remove must be called when the request was canceled from outside and not
// internally.
func (sb *shardBuffer) remove(toRemove *entry) {
	sb.mu.Lock()
	defer sb.mu.Unlock()

	if sb.queue == nil {
		// Queue is cleared because we're already in the DRAIN phase.
		return
	}

	// If entry is still in the queue, delete it and cancel it internally.
	for i, e := range sb.queue {
		if e == toRemove {
			// Delete entry at index "i" from slice.
			sb.queue = append(sb.queue[:i], sb.queue[i+1:]...)

			// Cancel the entry's "bufferCtx".
			// The usual drain or eviction code would unblock the request and then
			// wait for the "bufferCtx" to be done.
			// But this code path is different because it's going to return an error
			// to the request and not the "e.bufferCancel" function i.e. the request
			// cannot cancel the "bufferCtx" itself.
			// Therefore, we call "e.bufferCancel". This also avoids that the
			// context's Go routine could leak.
			e.bufferCancel()
			// Release the buffer slot and close the "e.done" channel.
			// By closing "e.done", we finish it explicitly and timeoutThread will
			// find out about it as well.
			sb.unblockAndWait(e, nil /* err */, true /* releaseSlot */, false /* blockingWait */)

			// Track it as "ContextDone" eviction.
			statsKeyWithReason := append(sb.statsKey, string(evictedContextDone))
			requestsEvicted.Add(statsKeyWithReason, 1)
			return
		}
	}

	// Entry was already removed. Keep the queue as it is.
}

func (sb *shardBuffer) recordKeyspaceEvent(alias *topodatapb.TabletAlias, stillServing bool, keyspaceEvent *discovery.KeyspaceEvent) {
	sb.mu.Lock()
	defer sb.mu.Unlock()

	log.Infof("disruption in shard %s/%s resolved (serving: %v), movetable state %#v",
		sb.keyspace, sb.shard, stillServing, keyspaceEvent.MoveTablesState)

	if !topoproto.TabletAliasEqual(alias, sb.currentPrimary) {
		if sb.currentPrimary != nil {
			sb.lastReparent = sb.timeNow()
		}
		sb.currentPrimary = alias
	}
	var reason stopReason
	var msg string

	// heuristically determine the reason why vtgate is currently buffering
	moveTablesSwitched := false
	if keyspaceEvent.MoveTablesState.State == discovery.MoveTablesSwitched {
		moveTablesSwitched = true
	}
	switch {
	case moveTablesSwitched:
		reason = stopMoveTablesSwitchingTraffic
		msg = stopMoveTablesSwitchingTrafficMessage
	case stillServing:
		reason = stopFailoverEndDetected
		msg = stopFailoverEndDetectedMessage
	default:
		reason = stopShardMissing
		msg = stopShardMissingMessage
	}
	sb.stopBufferingLocked(reason, msg)
}

func (sb *shardBuffer) stopBufferingDueToMaxDuration() {
	sb.mu.Lock()
	defer sb.mu.Unlock()

	sb.stopBufferingLocked(stopMaxFailoverDurationExceeded,
		fmt.Sprintf("stopping buffering because failover did not finish in time (%v)", sb.buf.config.MaxFailoverDuration))
}

func (sb *shardBuffer) stopBufferingLocked(reason stopReason, details string) {
	if sb.state != stateBuffering {
		return
	}

	// Stop buffering.
	sb.lastEnd = sb.timeNow()
	d := sb.lastEnd.Sub(sb.lastStart)

	statsKeyWithReason := append(sb.statsKey, string(reason))
	stops.Add(statsKeyWithReason, 1)

	lastFailoverDurationMs.Set(sb.statsKey, int64(d/time.Millisecond))
	failoverDurationSumMs.Add(sb.statsKey, int64(d/time.Millisecond))
	if sb.mode == bufferModeDryRun {
		utilDryRunMax := int64(
			float64(lastRequestsDryRunMax.Counts()[sb.statsKeyJoined]) / float64(sb.buf.config.Size) * 100.0)
		utilizationDryRunSum.Add(sb.statsKey, utilDryRunMax)
	} else {
		utilMax := int64(
			float64(lastRequestsInFlightMax.Counts()[sb.statsKeyJoined]) / float64(sb.buf.config.Size) * 100.0)
		utilizationSum.Add(sb.statsKey, utilMax)
	}

	sb.logErrorIfStateNotLocked(stateBuffering)
	sb.state = stateDraining
	q := sb.queue
	// Clear the queue such that remove(), oldestEntry() and evictOldestEntry()
	// will not work on obsolete data.
	sb.queue = nil

	msg := "Stopping buffering"
	if sb.mode == bufferModeDryRun {
		msg = "Dry-run: Would have stopped buffering"
	}
	log.Infof("%v for shard: %s after: %.1f seconds due to: %v. Draining %d buffered requests now.",
		msg, topoproto.KeyspaceShardString(sb.keyspace, sb.shard), d.Seconds(), details, len(q))

	var clientEntryError error
	if reason == stopShardMissing {
		clientEntryError = ShardMissingError
	}

	// Start the drain. (Use a new Go routine to release the lock.)
	sb.wg.Add(1)
	go sb.drain(q, clientEntryError)
}

// parallelRangeIndex uses counter to return a unique idx value up to the
// passed max and ok will be set to false if the counter exceeds the max
func parallelRangeIndex(counter *atomic.Int64, max int) (idx int, ok bool) {
	next := counter.Add(1)
	if next-1 > int64(max) {
		return -1, false
	}
	// if this is a 32-bit platform, max won't exceed the 32-bit integer limit
	// so a cast from a too-large 64-bit int to a 32-bit int will never happen
	return int(next) - 1, true
}

func (sb *shardBuffer) drain(q []*entry, err error) {
	defer sb.wg.Done()

	// stop must be called outside of the lock because the thread may access
	// shardBuffer as well e.g. to get the current oldest entry.
	sb.timeoutThread.stop()

	start := sb.timeNow()

	entryCount := len(q)
	parallelism := min(sb.buf.config.DrainConcurrency, entryCount)

	var wg sync.WaitGroup
	var rangeCounter atomic.Int64

	wg.Add(parallelism)
	for i := 0; i < parallelism; i++ {
		go func() {
			defer wg.Done()
			for {
				idx, ok := parallelRangeIndex(&rangeCounter, entryCount-1)
				if !ok {
					break
				}
				// Shared access to the q slice is concurrency-safe because each goroutine receives
				// a unique set of slice indices from parallelRangeIndex above and the slice remains
				// immutable for the lifetime of this operation.
				sb.unblockAndWait(q[idx], err, true /* releaseSlot */, true /* blockingWait */)
			}
		}()
	}

	wg.Wait()

	d := sb.timeNow().Sub(start)
	log.Infof("Draining finished for shard: %s Took: %v for: %d requests.", topoproto.KeyspaceShardString(sb.keyspace, sb.shard), d, len(q))
	requestsDrained.Add(sb.statsKey, int64(len(q)))

	// Draining is done. Change state from "draining" to "idle".
	sb.mu.Lock()
	defer sb.mu.Unlock()
	sb.logErrorIfStateNotLocked(stateDraining)
	sb.state = stateIdle
	sb.timeoutThread = nil
}

func (sb *shardBuffer) shutdown() {
	sb.mu.Lock()
	sb.stopBufferingLocked(stopShutdown, "shutdown")
	sb.mu.Unlock()
}

func (sb *shardBuffer) waitForShutdown() {
	sb.wg.Wait()
}

// sizeForTesting is used by the unit test only to find out the current number
// of buffered requests.
// TODO(mberlin): Remove this if we add a more general statistics reporting.
func (sb *shardBuffer) testGetSize() int {
	sb.mu.RLock()
	defer sb.mu.RUnlock()
	return len(sb.queue)
}

// stateForTesting is used by unit tests only to probe the current state.
func (sb *shardBuffer) testGetState() bufferState {
	sb.mu.RLock()
	defer sb.mu.RUnlock()
	return sb.state
}