ingest.go

// Copyright 2018 The LevelDB-Go and Pebble Authors. All rights reserved. Use
// of this source code is governed by a BSD-style license that can be found in
// the LICENSE file.

package pebble

import (
	"context"
	"sort"
	"sync/atomic"
	"time"

	"github.com/cockroachdb/errors"
	"github.com/cockroachdb/pebble/internal/base"
	"github.com/cockroachdb/pebble/internal/keyspan"
	"github.com/cockroachdb/pebble/internal/manifest"
	"github.com/cockroachdb/pebble/internal/private"
	"github.com/cockroachdb/pebble/objstorage"
	"github.com/cockroachdb/pebble/sstable"
)

func sstableKeyCompare(userCmp Compare, a, b InternalKey) int {
	c := userCmp(a.UserKey, b.UserKey)
	if c != 0 {
		return c
	}
	if a.IsExclusiveSentinel() {
		if !b.IsExclusiveSentinel() {
			return -1
		}
	} else if b.IsExclusiveSentinel() {
		return +1
	}
	return 0
}

func ingestValidateKey(opts *Options, key *InternalKey) error {
	if key.Kind() == InternalKeyKindInvalid {
		return base.CorruptionErrorf("pebble: external sstable has corrupted key: %s",
			key.Pretty(opts.Comparer.FormatKey))
	}
	if key.SeqNum() != 0 {
		return base.CorruptionErrorf("pebble: external sstable has non-zero seqnum: %s",
			key.Pretty(opts.Comparer.FormatKey))
	}
	return nil
}

func ingestLoad1(
	opts *Options, fmv FormatMajorVersion, path string, cacheID uint64, fileNum FileNum,
) (*fileMetadata, error) {
	f, err := opts.FS.Open(path)
	if err != nil {
		return nil, err
	}

	readable, err := sstable.NewSimpleReadable(f)
	if err != nil {
		return nil, err
	}

	cacheOpts := private.SSTableCacheOpts(cacheID, fileNum).(sstable.ReaderOption)
	r, err := sstable.NewReader(readable, opts.MakeReaderOptions(), cacheOpts)
	if err != nil {
		return nil, err
	}
	defer r.Close()

	// Avoid ingesting tables with format versions this DB doesn't support.
	tf, err := r.TableFormat()
	if err != nil {
		return nil, err
	}
	if tf < fmv.MinTableFormat() || tf > fmv.MaxTableFormat() {
		return nil, errors.Newf(
			"pebble: table format %s is not within range supported at DB format major version %d, (%s,%s)",
			tf, fmv, fmv.MinTableFormat(), fmv.MaxTableFormat(),
		)
	}

	meta := &fileMetadata{}
	meta.FileNum = fileNum
	meta.Size = uint64(readable.Size())
	meta.CreationTime = time.Now().Unix()

	// Avoid loading into the table cache for collecting stats if we
	// don't need to. If there are no range deletions, we have all the
	// information to compute the stats here.
	//
	// This is helpful in tests for avoiding awkwardness around deletion of
	// ingested files from MemFS. MemFS implements the Windows semantics of
	// disallowing removal of an open file. Under MemFS, if we don't populate
	// meta.Stats here, the file will be loaded into the table cache for
	// calculating stats before we can remove the original link.
	maybeSetStatsFromProperties(meta, &r.Properties)

	{
		iter, err := r.NewIter(nil /* lower */, nil /* upper */)
		if err != nil {
			return nil, err
		}
		defer iter.Close()
		var smallest InternalKey
		if key, _ := iter.First(); key != nil {
			if err := ingestValidateKey(opts, key); err != nil {
				return nil, err
			}
			smallest = (*key).Clone()
		}
		if err := iter.Error(); err != nil {
			return nil, err
		}
		if key, _ := iter.Last(); key != nil {
			if err := ingestValidateKey(opts, key); err != nil {
				return nil, err
			}
			meta.ExtendPointKeyBounds(opts.Comparer.Compare, smallest, key.Clone())
		}
		if err := iter.Error(); err != nil {
			return nil, err
		}
	}

	iter, err := r.NewRawRangeDelIter()
	if err != nil {
		return nil, err
	}
	if iter != nil {
		defer iter.Close()
		var smallest InternalKey
		if s := iter.First(); s != nil {
			key := s.SmallestKey()
			if err := ingestValidateKey(opts, &key); err != nil {
				return nil, err
			}
			smallest = key.Clone()
		}
		if err := iter.Error(); err != nil {
			return nil, err
		}
		if s := iter.Last(); s != nil {
			k := s.SmallestKey()
			if err := ingestValidateKey(opts, &k); err != nil {
				return nil, err
			}
			largest := s.LargestKey().Clone()
			meta.ExtendPointKeyBounds(opts.Comparer.Compare, smallest, largest)
		}
	}

	// Update the range-key bounds for the table.
	{
		iter, err := r.NewRawRangeKeyIter()
		if err != nil {
			return nil, err
		}
		if iter != nil {
			defer iter.Close()
			var smallest InternalKey
			if s := iter.First(); s != nil {
				key := s.SmallestKey()
				if err := ingestValidateKey(opts, &key); err != nil {
					return nil, err
				}
				smallest = key.Clone()
			}
			if err := iter.Error(); err != nil {
				return nil, err
			}
			if s := iter.Last(); s != nil {
				k := s.SmallestKey()
				if err := ingestValidateKey(opts, &k); err != nil {
					return nil, err
				}
				// As range keys are fragmented, the end key of the last range key in
				// the table provides the upper bound for the table.
				largest := s.LargestKey().Clone()
				meta.ExtendRangeKeyBounds(opts.Comparer.Compare, smallest, largest)
			}
			if err := iter.Error(); err != nil {
				return nil, err
			}
		}
	}

	if !meta.HasPointKeys && !meta.HasRangeKeys {
		return nil, nil
	}

	// Sanity check that the various bounds on the file were set consistently.
	if err := meta.Validate(opts.Comparer.Compare, opts.Comparer.FormatKey); err != nil {
		return nil, err
	}

	return meta, nil
}

func ingestLoad(
	opts *Options, fmv FormatMajorVersion, paths []string, cacheID uint64, pending []FileNum,
) ([]*fileMetadata, []string, error) {
	meta := make([]*fileMetadata, 0, len(paths))
	newPaths := make([]string, 0, len(paths))
	for i := range paths {
		m, err := ingestLoad1(opts, fmv, paths[i], cacheID, pending[i])
		if err != nil {
			return nil, nil, err
		}
		if m != nil {
			meta = append(meta, m)
			newPaths = append(newPaths, paths[i])
		}
	}
	return meta, newPaths, nil
}

// Struct for sorting metadatas by smallest user keys, while ensuring the
// matching path also gets swapped to the same index. For use in
// ingestSortAndVerify.
type metaAndPaths struct {
	meta  []*fileMetadata
	paths []string
	cmp   Compare
}

func (m metaAndPaths) Len() int {
	return len(m.meta)
}

func (m metaAndPaths) Less(i, j int) bool {
	return m.cmp(m.meta[i].Smallest.UserKey, m.meta[j].Smallest.UserKey) < 0
}

func (m metaAndPaths) Swap(i, j int) {
	m.meta[i], m.meta[j] = m.meta[j], m.meta[i]
	m.paths[i], m.paths[j] = m.paths[j], m.paths[i]
}

func ingestSortAndVerify(cmp Compare, meta []*fileMetadata, paths []string) error {
	if len(meta) <= 1 {
		return nil
	}

	sort.Sort(&metaAndPaths{
		meta:  meta,
		paths: paths,
		cmp:   cmp,
	})

	for i := 1; i < len(meta); i++ {
		if sstableKeyCompare(cmp, meta[i-1].Largest, meta[i].Smallest) >= 0 {
			return errors.New("pebble: external sstables have overlapping ranges")
		}
	}
	return nil
}

func ingestCleanup(objProvider *objstorage.Provider, meta []*fileMetadata) error {
	var firstErr error
	for i := range meta {
		if err := objProvider.Remove(fileTypeTable, meta[i].FileNum); err != nil {
			firstErr = firstError(firstErr, err)
		}
	}
	return firstErr
}

// ingestLink creates new objects which are backed by either hardlinks to or
// copies of the ingested files.
func ingestLink(
	jobID int, opts *Options, objProvider *objstorage.Provider, paths []string, meta []*fileMetadata,
) error {
	for i := range paths {
		objMeta, err := objProvider.LinkOrCopyFromLocal(opts.FS, paths[i], fileTypeTable, meta[i].FileNum)
		if err != nil {
			if err2 := ingestCleanup(objProvider, meta[:i]); err2 != nil {
				opts.Logger.Infof("ingest cleanup failed: %v", err2)
			}
			return err
		}
		if opts.EventListener.TableCreated != nil {
			opts.EventListener.TableCreated(TableCreateInfo{
				JobID:   jobID,
				Reason:  "ingesting",
				Path:    objProvider.Path(objMeta),
				FileNum: meta[i].FileNum,
			})
		}
	}

	return nil
}

func ingestMemtableOverlaps(cmp Compare, mem flushable, meta []*fileMetadata) bool {
	iter := mem.newIter(nil)
	rangeDelIter := mem.newRangeDelIter(nil)
	rkeyIter := mem.newRangeKeyIter(nil)

	closeIters := func() error {
		err := iter.Close()
		if rangeDelIter != nil {
			err = firstError(err, rangeDelIter.Close())
		}
		if rkeyIter != nil {
			err = firstError(err, rkeyIter.Close())
		}
		return err
	}

	for _, m := range meta {
		if overlapWithIterator(iter, &rangeDelIter, rkeyIter, m, cmp) {
			closeIters()
			return true
		}
	}

	// Assume overlap if any iterator errored out.
	return closeIters() != nil
}

func ingestUpdateSeqNum(
	cmp Compare, format base.FormatKey, seqNum uint64, meta []*fileMetadata,
) error {
	setSeqFn := func(k base.InternalKey) base.InternalKey {
		return base.MakeInternalKey(k.UserKey, seqNum, k.Kind())
	}
	for _, m := range meta {
		// NB: we set the fields directly here, rather than via their Extend*
		// methods, as we are updating sequence numbers.
		if m.HasPointKeys {
			m.SmallestPointKey = setSeqFn(m.SmallestPointKey)
		}
		if m.HasRangeKeys {
			m.SmallestRangeKey = setSeqFn(m.SmallestRangeKey)
		}
		m.Smallest = setSeqFn(m.Smallest)
		// Only update the seqnum for the largest key if that key is not an
		// "exclusive sentinel" (i.e. a range deletion sentinel or a range key
		// boundary), as doing so effectively drops the exclusive sentinel (by
		// lowering the seqnum from the max value), and extends the bounds of the
		// table.
		// NB: as the largest range key is always an exclusive sentinel, it is never
		// updated.
		if m.HasPointKeys && !m.LargestPointKey.IsExclusiveSentinel() {
			m.LargestPointKey = setSeqFn(m.LargestPointKey)
		}
		if !m.Largest.IsExclusiveSentinel() {
			m.Largest = setSeqFn(m.Largest)
		}
		// Setting smallestSeqNum == largestSeqNum triggers the setting of
		// Properties.GlobalSeqNum when an sstable is loaded.
		m.SmallestSeqNum = seqNum
		m.LargestSeqNum = seqNum
		// Ensure the new bounds are consistent.
		if err := m.Validate(cmp, format); err != nil {
			return err
		}
		seqNum++
	}
	return nil
}

func overlapWithIterator(
	iter internalIterator,
	rangeDelIter *keyspan.FragmentIterator,
	rkeyIter keyspan.FragmentIterator,
	meta *fileMetadata,
	cmp Compare,
) bool {
	// Check overlap with point operations.
	//
	// When using levelIter, it seeks to the SST whose boundaries
	// contain meta.Smallest.UserKey(S).
	// It then tries to find a point in that SST that is >= S.
	// If there's no such point it means the SST ends in a tombstone in which case
	// levelIter.SeekGE generates a boundary range del sentinel.
	// The comparison of this boundary with meta.Largest(L) below
	// is subtle but maintains correctness.
	// 1) boundary < L,
	//    since boundary is also > S (initial seek),
	//    whatever the boundary's start key may be, we're always overlapping.
	// 2) boundary > L,
	//    overlap with boundary cannot be determined since we don't know boundary's start key.
	//    We require checking for overlap with rangeDelIter.
	// 3) boundary == L and L is not sentinel,
	//    means boundary < L and hence is similar to 1).
	// 4) boundary == L and L is sentinel,
	//    we'll always overlap since for any values of i,j ranges [i, k) and [j, k) always overlap.
	key, _ := iter.SeekGE(meta.Smallest.UserKey, base.SeekGEFlagsNone)
	if key != nil {
		c := sstableKeyCompare(cmp, *key, meta.Largest)
		if c <= 0 {
			return true
		}
	}
	// Assume overlap if iterator errored.
	if err := iter.Error(); err != nil {
		return true
	}

	computeOverlapWithSpans := func(rIter keyspan.FragmentIterator) bool {
		// NB: The spans surfaced by the fragment iterator are non-overlapping.
		span := rIter.SeekLT(meta.Smallest.UserKey)
		if span == nil {
			span = rIter.Next()
		}
		for ; span != nil; span = rIter.Next() {
			if span.Empty() {
				continue
			}
			key := span.SmallestKey()
			c := sstableKeyCompare(cmp, key, meta.Largest)
			if c > 0 {
				// The start of the span is after the largest key in the
				// ingested table.
				return false
			}
			if cmp(span.End, meta.Smallest.UserKey) > 0 {
				// The end of the span is greater than the smallest in the
				// table. Note that the span end key is exclusive, thus ">0"
				// instead of ">=0".
				return true
			}
		}
		// Assume overlap if iterator errored.
		if err := rIter.Error(); err != nil {
			return true
		}
		return false
	}

	// rkeyIter is either a range key level iter, or a range key iterator
	// over a single file.
	if rkeyIter != nil {
		if computeOverlapWithSpans(rkeyIter) {
			return true
		}
	}

	// Check overlap with range deletions.
	if rangeDelIter == nil || *rangeDelIter == nil {
		return false
	}
	return computeOverlapWithSpans(*rangeDelIter)
}

func ingestTargetLevel(
	newIters tableNewIters,
	newRangeKeyIter keyspan.TableNewSpanIter,
	iterOps IterOptions,
	cmp Compare,
	v *version,
	baseLevel int,
	compactions map[*compaction]struct{},
	meta *fileMetadata,
) (int, error) {
	// Find the lowest level which does not have any files which overlap meta. We
	// search from L0 to L6 looking for whether there are any files in the level
	// which overlap meta. We want the "lowest" level (where lower means
	// increasing level number) in order to reduce write amplification.
	//
	// There are 2 kinds of overlap we need to check for: file boundary overlap
	// and data overlap. Data overlap implies file boundary overlap. Note that it
	// is always possible to ingest into L0.
	//
	// To place meta at level i where i > 0:
	// - there must not be any data overlap with levels <= i, since that will
	//   violate the sequence number invariant.
	// - no file boundary overlap with level i, since that will violate the
	//   invariant that files do not overlap in levels i > 0.
	//
	// The file boundary overlap check is simpler to conceptualize. Consider the
	// following example, in which the ingested file lies completely before or
	// after the file being considered.
	//
	//   |--|           |--|  ingested file: [a,b] or [f,g]
	//         |-----|        existing file: [c,e]
	//  _____________________
	//   a  b  c  d  e  f  g
	//
	// In both cases the ingested file can move to considering the next level.
	//
	// File boundary overlap does not necessarily imply data overlap. The check
	// for data overlap is a little more nuanced. Consider the following examples:
	//
	//  1. No data overlap:
	//
	//          |-|   |--|    ingested file: [cc-d] or [ee-ff]
	//  |*--*--*----*------*| existing file: [a-g], points: [a, b, c, dd, g]
	//  _____________________
	//   a  b  c  d  e  f  g
	//
	// In this case the ingested files can "fall through" this level. The checks
	// continue at the next level.
	//
	//  2. Data overlap:
	//
	//            |--|        ingested file: [d-e]
	//  |*--*--*----*------*| existing file: [a-g], points: [a, b, c, dd, g]
	//  _____________________
	//   a  b  c  d  e  f  g
	//
	// In this case the file cannot be ingested into this level as the point 'dd'
	// is in the way.
	//
	// It is worth noting that the check for data overlap is only approximate. In
	// the previous example, the ingested table [d-e] could contain only the
	// points 'd' and 'e', in which case the table would be eligible for
	// considering lower levels. However, such a fine-grained check would need to
	// be exhaustive (comparing points and ranges in both the ingested existing
	// tables) and such a check is prohibitively expensive. Thus Pebble treats any
	// existing point that falls within the ingested table bounds as being "data
	// overlap".

	targetLevel := 0

	// Do we overlap with keys in L0?
	// TODO(bananabrick): Use sublevels to compute overlap.
	iter := v.Levels[0].Iter()
	for meta0 := iter.First(); meta0 != nil; meta0 = iter.Next() {
		c1 := sstableKeyCompare(cmp, meta.Smallest, meta0.Largest)
		c2 := sstableKeyCompare(cmp, meta.Largest, meta0.Smallest)
		if c1 > 0 || c2 < 0 {
			continue
		}

		// TODO(sumeer): ingest is a user-facing operation, so we should accept a
		// context and plumb it through, for tracing.
		iter, rangeDelIter, err := newIters(context.Background(), meta0, nil, internalIterOpts{})
		if err != nil {
			return 0, err
		}
		rkeyIter, err := newRangeKeyIter(meta0, nil)
		if err != nil {
			return 0, err
		}
		overlap := overlapWithIterator(iter, &rangeDelIter, rkeyIter, meta, cmp)
		err = firstError(err, iter.Close())
		if rangeDelIter != nil {
			err = firstError(err, rangeDelIter.Close())
		}
		if rkeyIter != nil {
			err = firstError(err, rkeyIter.Close())
		}
		if err != nil {
			return 0, err
		}
		if overlap {
			return targetLevel, nil
		}
	}

	level := baseLevel
	for ; level < numLevels; level++ {
		levelIter := newLevelIter(iterOps, cmp, nil /* split */, newIters,
			v.Levels[level].Iter(), manifest.Level(level), nil)
		var rangeDelIter keyspan.FragmentIterator
		// Pass in a non-nil pointer to rangeDelIter so that levelIter.findFileGE
		// sets it up for the target file.
		levelIter.initRangeDel(&rangeDelIter)

		rkeyLevelIter := &keyspan.LevelIter{}
		rkeyLevelIter.Init(
			keyspan.SpanIterOptions{}, cmp, newRangeKeyIter,
			v.Levels[level].Iter(), manifest.Level(level), manifest.KeyTypeRange,
		)

		overlap := overlapWithIterator(levelIter, &rangeDelIter, rkeyLevelIter, meta, cmp)
		err := levelIter.Close() // Closes range del iter as well.
		err = firstError(err, rkeyLevelIter.Close())
		if err != nil {
			return 0, err
		}
		if overlap {
			return targetLevel, nil
		}

		// Check boundary overlap.
		boundaryOverlaps := v.Overlaps(level, cmp, meta.Smallest.UserKey,
			meta.Largest.UserKey, meta.Largest.IsExclusiveSentinel())
		if !boundaryOverlaps.Empty() {
			continue
		}

		// Check boundary overlap with any ongoing compactions.
		//
		// We cannot check for data overlap with the new SSTs compaction will
		// produce since compaction hasn't been done yet. However, there's no need
		// to check since all keys in them will either be from c.startLevel or
		// c.outputLevel, both levels having their data overlap already tested
		// negative (else we'd have returned earlier).
		overlaps := false
		for c := range compactions {
			if c.outputLevel == nil || level != c.outputLevel.level {
				continue
			}
			if cmp(meta.Smallest.UserKey, c.largest.UserKey) <= 0 &&
				cmp(meta.Largest.UserKey, c.smallest.UserKey) >= 0 {
				overlaps = true
				break
			}
		}
		if !overlaps {
			targetLevel = level
		}
	}
	return targetLevel, nil
}

// Ingest ingests a set of sstables into the DB. Ingestion of the files is
// atomic and semantically equivalent to creating a single batch containing all
// of the mutations in the sstables. Ingestion may require the memtable to be
// flushed. The ingested sstable files are moved into the DB and must reside on
// the same filesystem as the DB. Sstables can be created for ingestion using
// sstable.Writer. On success, Ingest removes the input paths.
//
// All sstables *must* be Sync()'d by the caller after all bytes are written
// and before its file handle is closed; failure to do so could violate
// durability or lead to corrupted on-disk state. This method cannot, in a
// platform-and-FS-agnostic way, ensure that all sstables in the input are
// properly synced to disk. Opening new file handles and Sync()-ing them
// does not always guarantee durability; see the discussion here on that:
// https://github.com/cockroachdb/pebble/pull/835#issuecomment-663075379
//
// Ingestion loads each sstable into the lowest level of the LSM which it
// doesn't overlap (see ingestTargetLevel). If an sstable overlaps a memtable,
// ingestion forces the memtable to flush, and then waits for the flush to
// occur.
//
// The steps for ingestion are:
//
//  1. Allocate file numbers for every sstable being ingested.
//  2. Load the metadata for all sstables being ingest.
//  3. Sort the sstables by smallest key, verifying non overlap.
//  4. Hard link (or copy) the sstables into the DB directory.
//  5. Allocate a sequence number to use for all of the entries in the
//     sstables. This is the step where overlap with memtables is
//     determined. If there is overlap, we remember the most recent memtable
//     that overlaps.
//  6. Update the sequence number in the ingested sstables.
//  7. Wait for the most recent memtable that overlaps to flush (if any).
//  8. Add the ingested sstables to the version (DB.ingestApply).
//  9. Publish the ingestion sequence number.
//
// Note that if the mutable memtable overlaps with ingestion, a flush of the
// memtable is forced equivalent to DB.Flush. Additionally, subsequent
// mutations that get sequence numbers larger than the ingestion sequence
// number get queued up behind the ingestion waiting for it to complete. This
// can produce a noticeable hiccup in performance. See
// https://github.com/cockroachdb/pebble/issues/25 for an idea for how to fix
// this hiccup.
func (d *DB) Ingest(paths []string) error {
	if err := d.closed.Load(); err != nil {
		panic(err)
	}
	if d.opts.ReadOnly {
		return ErrReadOnly
	}
	_, err := d.ingest(paths, ingestTargetLevel)
	return err
}

// IngestOperationStats provides some information about where in the LSM the
// bytes were ingested.
type IngestOperationStats struct {
	// Bytes is the total bytes in the ingested sstables.
	Bytes uint64
	// ApproxIngestedIntoL0Bytes is the approximate number of bytes ingested
	// into L0. This value is approximate when flushable ingests are active and
	// an ingest overlaps an entry in the flushable queue. Currently, this
	// approximation is very rough, only including tables that overlapped the
	// memtable. This estimate may be improved with #2112.
	ApproxIngestedIntoL0Bytes uint64
	// MemtableOverlappingFiles is the count of ingested sstables
	// that overlapped keys in the memtables.
	MemtableOverlappingFiles int
}

// IngestWithStats does the same as Ingest, and additionally returns
// IngestOperationStats.
func (d *DB) IngestWithStats(paths []string) (IngestOperationStats, error) {
	if err := d.closed.Load(); err != nil {
		panic(err)
	}
	if d.opts.ReadOnly {
		return IngestOperationStats{}, ErrReadOnly
	}
	return d.ingest(paths, ingestTargetLevel)
}

// Both DB.mu and commitPipeline.mu must be held while this is called.
func (d *DB) newIngestedFlushableEntry(
	meta []*fileMetadata, seqNum uint64, logNum FileNum,
) (*flushableEntry, error) {
	// Update the sequence number for all of the sstables in the
	// metadata. Writing the metadata to the manifest when the
	// version edit is applied is the mechanism that persists the
	// sequence number. The sstables themselves are left unmodified.
	// In this case, a version edit will only be written to the manifest
	// when the flushable is eventually flushed. If Pebble restarts in that
	// time, then we'll lose the ingest sequence number information. But this
	// information will also be reconstructed on node restart.
	if err := ingestUpdateSeqNum(
		d.cmp, d.opts.Comparer.FormatKey, seqNum, meta,
	); err != nil {
		return nil, err
	}

	f := newIngestedFlushable(meta, d.cmp, d.split, d.newIters, d.tableNewRangeKeyIter)

	// NB: The logNum/seqNum are the WAL number which we're writing this entry
	// to and the sequence number within the WAL which we'll write this entry
	// to.
	entry := d.newFlushableEntry(f, logNum, seqNum)
	// The flushable entry starts off with a single reader ref, so increment
	// the FileMetadata.Refs.
	for _, file := range f.files {
		atomic.AddInt32(&file.Refs, 1)
	}
	entry.unrefFiles = func() []*fileMetadata {
		var obsolete []*fileMetadata
		for _, file := range f.files {
			if val := atomic.AddInt32(&file.Refs, -1); val == 0 {
				obsolete = append(obsolete, file)
			}
		}
		return obsolete
	}

	entry.flushForced = true
	entry.releaseMemAccounting = func() {}
	return entry, nil
}

// Both DB.mu and commitPipeline.mu must be held while this is called. Since
// we're holding both locks, the order in which we rotate the memtable or
// recycle the WAL in this function is irrelevant as long as the correct log
// numbers are assigned to the appropriate flushable.
func (d *DB) handleIngestAsFlushable(meta []*fileMetadata, seqNum uint64) error {
	b := d.NewBatch()
	for _, m := range meta {
		b.ingestSST(m.FileNum)
	}
	b.setSeqNum(seqNum)

	// If the WAL is disabled, then the logNum used to create the flushable
	// entry doesn't matter. We just use the logNum assigned to the current
	// mutable memtable. If the WAL is enabled, then this logNum will be
	// overwritten by the logNum of the log which will contain the log entry
	// for the ingestedFlushable.
	logNum := d.mu.mem.queue[len(d.mu.mem.queue)-1].logNum
	if !d.opts.DisableWAL {
		// We create a new WAL for the flushable instead of reusing the end of
		// the previous WAL. This simplifies the increment of the minimum
		// unflushed log number, and also simplifies WAL replay.
		logNum, _ = d.recycleWAL()
		d.mu.Unlock()
		err := d.commit.directWrite(b)
		if err != nil {
			d.opts.Logger.Fatalf("%v", err)
		}
		d.mu.Lock()
	}

	entry, err := d.newIngestedFlushableEntry(meta, seqNum, logNum)
	if err != nil {
		return err
	}
	nextSeqNum := seqNum + uint64(b.Count())

	// Set newLogNum to the logNum of the previous flushable. This value is
	// irrelevant if the WAL is disabled. If the WAL is enabled, then we set
	// the appropriate value below.
	newLogNum := d.mu.mem.queue[len(d.mu.mem.queue)-1].logNum
	if !d.opts.DisableWAL {
		// This is WAL num of the next mutable memtable which comes after the
		// ingestedFlushable in the flushable queue. The mutable memtable
		// will be created below.
		newLogNum, _ = d.recycleWAL()
		if err != nil {
			return err
		}
	}

	currMem := d.mu.mem.mutable
	// NB: Placing ingested sstables above the current memtables
	// requires rotating of the existing memtables/WAL. There is
	// some concern of churning through tiny memtables due to
	// ingested sstables being placed on top of them, but those
	// memtables would have to be flushed anyways.
	d.mu.mem.queue = append(d.mu.mem.queue, entry)
	d.rotateMemtable(newLogNum, nextSeqNum, currMem)
	d.updateReadStateLocked(d.opts.DebugCheck)
	d.maybeScheduleFlush()
	return nil
}

func (d *DB) ingest(
	paths []string, targetLevelFunc ingestTargetLevelFunc,
) (IngestOperationStats, error) {
	// Allocate file numbers for all of the files being ingested and mark them as
	// pending in order to prevent them from being deleted. Note that this causes
	// the file number ordering to be out of alignment with sequence number
	// ordering. The sorting of L0 tables by sequence number avoids relying on
	// that (busted) invariant.
	d.mu.Lock()
	pendingOutputs := make([]FileNum, len(paths))
	for i := range paths {
		pendingOutputs[i] = d.mu.versions.getNextFileNum()
	}
	jobID := d.mu.nextJobID
	d.mu.nextJobID++
	d.mu.Unlock()

	// Load the metadata for all of the files being ingested. This step detects
	// and elides empty sstables.
	meta, paths, err := ingestLoad(d.opts, d.FormatMajorVersion(), paths, d.cacheID, pendingOutputs)
	if err != nil {
		return IngestOperationStats{}, err
	}
	if len(meta) == 0 {
		// All of the sstables to be ingested were empty. Nothing to do.
		return IngestOperationStats{}, nil
	}

	// Verify the sstables do not overlap.
	if err := ingestSortAndVerify(d.cmp, meta, paths); err != nil {
		return IngestOperationStats{}, err
	}

	// Hard link the sstables into the DB directory. Since the sstables aren't
	// referenced by a version, they won't be used. If the hard linking fails
	// (e.g. because the files reside on a different filesystem), ingestLink will
	// fall back to copying, and if that fails we undo our work and return an
	// error.
	if err := ingestLink(jobID, d.opts, d.objProvider, paths, meta); err != nil {
		return IngestOperationStats{}, err
	}
	// Make the new tables durable. We need to do this at some point before we
	// update the MANIFEST (via logAndApply), otherwise a crash can have the
	// tables referenced in the MANIFEST, but not present in the provider.
	if err := d.objProvider.Sync(); err != nil {
		return IngestOperationStats{}, err
	}

	// metaFlushableOverlaps is a slice parallel to meta indicating which of the
	// ingested sstables overlap some table in the flushable queue. It's used to
	// approximate ingest-into-L0 stats when using flushable ingests.
	metaFlushableOverlaps := make([]bool, len(meta))
	var mem *flushableEntry
	var mut *memTable
	// asFlushable indicates whether the sstable was ingested as a flushable.
	var asFlushable bool
	prepare := func(seqNum uint64) {
		// Note that d.commit.mu is held by commitPipeline when calling prepare.

		d.mu.Lock()
		defer d.mu.Unlock()

		// Check to see if any files overlap with any of the memtables. The queue
		// is ordered from oldest to newest with the mutable memtable being the
		// last element in the slice. We want to wait for the newest table that
		// overlaps.

		for i := len(d.mu.mem.queue) - 1; i >= 0; i-- {
			m := d.mu.mem.queue[i]
			iter := m.newIter(nil)
			rangeDelIter := m.newRangeDelIter(nil)
			rkeyIter := m.newRangeKeyIter(nil)
			for i := range meta {
				if metaFlushableOverlaps[i] {
					// This table already overlapped a more recent flushable.
					continue
				}
				if overlapWithIterator(iter, &rangeDelIter, rkeyIter, meta[i], d.cmp) {
					// If this is the first table to overlap a flushable, save
					// the flushable. This ingest must be ingested or flushed
					// after it.
					if mem == nil {
						mem = m
					}
					metaFlushableOverlaps[i] = true
				}
			}
			err := iter.Close()
			if rangeDelIter != nil {
				err = firstError(err, rangeDelIter.Close())
			}
			if rkeyIter != nil {
				err = firstError(err, rkeyIter.Close())
			}
			if err != nil {
				d.opts.Logger.Infof("ingest error reading flushable for log %s: %s", m.logNum, err)
			}
		}

		if mem == nil {
			// No overlap with any of the queued flushables, so no need to queue
			// after them.

			// New writes with higher sequence numbers may be concurrently
			// committed. We must ensure they don't flush before this ingest
			// completes. To do that, we ref the mutable memtable as a writer,
			// preventing its flushing (and the flushing of all subsequent
			// flushables in the queue). Once we've acquired the manifest lock
			// to add the ingested sstables to the LSM, we can unref as we're
			// guaranteed that the flush won't edit the LSM before this ingest.
			mut = d.mu.mem.mutable
			mut.writerRef()
			return
		}
		// The ingestion overlaps with some entry in the flushable queue.
		if d.FormatMajorVersion() < FormatFlushableIngest ||
			d.opts.Experimental.DisableIngestAsFlushable() ||
			(len(d.mu.mem.queue) > d.opts.MemTableStopWritesThreshold-1) {
			// We're not able to ingest as a flushable,
			// so we must synchronously flush.
			if mem.flushable == d.mu.mem.mutable {
				err = d.makeRoomForWrite(nil)
			}
			// New writes with higher sequence numbers may be concurrently
			// committed. We must ensure they don't flush before this ingest
			// completes. To do that, we ref the mutable memtable as a writer,
			// preventing its flushing (and the flushing of all subsequent
			// flushables in the queue). Once we've acquired the manifest lock
			// to add the ingested sstables to the LSM, we can unref as we're
			// guaranteed that the flush won't edit the LSM before this ingest.
			mut = d.mu.mem.mutable
			mut.writerRef()
			mem.flushForced = true
			d.maybeScheduleFlush()
			return
		}
		// Since there aren't too many memtables already queued up, we can
		// slide the ingested sstables on top of the existing memtables.
		asFlushable = true
		err = d.handleIngestAsFlushable(meta, seqNum)
	}

	var ve *versionEdit
	apply := func(seqNum uint64) {
		if err != nil || asFlushable {
			// An error occurred during prepare.
			if mut != nil {
				if mut.writerUnref() {
					d.mu.Lock()
					d.maybeScheduleFlush()
					d.mu.Unlock()
				}
			}
			return
		}

		// Update the sequence number for all of the sstables in the
		// metadata. Writing the metadata to the manifest when the
		// version edit is applied is the mechanism that persists the
		// sequence number. The sstables themselves are left unmodified.
		if err = ingestUpdateSeqNum(
			d.cmp, d.opts.Comparer.FormatKey, seqNum, meta,
		); err != nil {
			if mut != nil {
				if mut.writerUnref() {
					d.mu.Lock()
					d.maybeScheduleFlush()
					d.mu.Unlock()
				}
			}
			return
		}

		// If we overlapped with a memtable in prepare wait for the flush to
		// finish.
		if mem != nil {
			<-mem.flushed
		}

		// Assign the sstables to the correct level in the LSM and apply the
		// version edit.
		ve, err = d.ingestApply(jobID, meta, targetLevelFunc, mut)
	}

	d.commit.AllocateSeqNum(len(meta), prepare, apply)

	if err != nil {
		if err2 := ingestCleanup(d.objProvider, meta); err2 != nil {
			d.opts.Logger.Infof("ingest cleanup failed: %v", err2)
		}
	} else {
		// Since we either created a hard link to the ingesting files, or copied
		// them over, it is safe to remove the originals paths.
		for _, path := range paths {
			if err2 := d.opts.FS.Remove(path); err2 != nil {
				d.opts.Logger.Infof("ingest failed to remove original file: %s", err2)
			}
		}
	}

	info := TableIngestInfo{
		JobID:        jobID,
		GlobalSeqNum: meta[0].SmallestSeqNum,
		Err:          err,
		flushable:    asFlushable,
	}
	var stats IngestOperationStats
	if ve != nil {
		info.Tables = make([]struct {
			TableInfo
			Level int
		}, len(ve.NewFiles))
		for i := range ve.NewFiles {
			e := &ve.NewFiles[i]
			info.Tables[i].Level = e.Level
			info.Tables[i].TableInfo = e.Meta.TableInfo()
			stats.Bytes += e.Meta.Size
			if e.Level == 0 {
				stats.ApproxIngestedIntoL0Bytes += e.Meta.Size
			}
			if metaFlushableOverlaps[i] {
				stats.MemtableOverlappingFiles++
			}
		}
	} else if asFlushable {
		info.Tables = make([]struct {
			TableInfo
			Level int
		}, len(meta))
		for i, f := range meta {
			info.Tables[i].Level = -1
			info.Tables[i].TableInfo = f.TableInfo()
			stats.Bytes += f.Size
			// We don't have exact stats on which files will be ingested into
			// L0, because actual ingestion into the LSM has been deferred until
			// flush time. Instead, we infer based on memtable overlap.
			//
			// TODO(jackson): If we optimistically compute data overlap (#2112)
			// before entering the commit pipeline, we can use that overlap to
			// improve our approximation by incorporating overlap with L0, not
			// just memtables.
			if metaFlushableOverlaps[i] {
				stats.ApproxIngestedIntoL0Bytes += f.Size
				stats.MemtableOverlappingFiles++
			}
		}
	}
	d.opts.EventListener.TableIngested(info)

	return stats, err
}

type ingestTargetLevelFunc func(
	newIters tableNewIters,
	newRangeKeyIter keyspan.TableNewSpanIter,
	iterOps IterOptions,
	cmp Compare,
	v *version,
	baseLevel int,
	compactions map[*compaction]struct{},
	meta *fileMetadata,
) (int, error)

func (d *DB) ingestApply(
	jobID int, meta []*fileMetadata, findTargetLevel ingestTargetLevelFunc, mut *memTable,
) (*versionEdit, error) {
	d.mu.Lock()
	defer d.mu.Unlock()

	ve := &versionEdit{
		NewFiles: make([]newFileEntry, len(meta)),
	}
	metrics := make(map[int]*LevelMetrics)

	// Lock the manifest for writing before we use the current version to
	// determine the target level. This prevents two concurrent ingestion jobs
	// from using the same version to determine the target level, and also
	// provides serialization with concurrent compaction and flush jobs.
	// logAndApply unconditionally releases the manifest lock, but any earlier
	// returns must unlock the manifest.
	d.mu.versions.logLock()

	scheduleFlush := false
	if mut != nil {
		// Unref the mutable memtable to allows its flush to proceed. Now that we've
		// acquired the manifest lock, we can be certain that if the mutable
		// memtable has received more recent conflicting writes, the flush won't
		// beat us to applying to the manifest resulting in sequence number
		// inversion.
		scheduleFlush = mut.writerUnref()
	}

	current := d.mu.versions.currentVersion()
	baseLevel := d.mu.versions.picker.getBaseLevel()
	iterOps := IterOptions{logger: d.opts.Logger}
	for i := range meta {
		// Determine the lowest level in the LSM for which the sstable doesn't
		// overlap any existing files in the level.
		m := meta[i]
		f := &ve.NewFiles[i]
		var err error
		f.Level, err = findTargetLevel(d.newIters, d.tableNewRangeKeyIter, iterOps, d.cmp, current, baseLevel, d.mu.compact.inProgress, m)
		if err != nil {
			d.mu.versions.logUnlock()
			return nil, err
		}
		f.Meta = m
		levelMetrics := metrics[f.Level]
		if levelMetrics == nil {
			levelMetrics = &LevelMetrics{}
			metrics[f.Level] = levelMetrics
		}
		levelMetrics.NumFiles++
		levelMetrics.Size += int64(m.Size)
		levelMetrics.BytesIngested += m.Size
		levelMetrics.TablesIngested++
	}
	if err := d.mu.versions.logAndApply(jobID, ve, metrics, false /* forceRotation */, func() []compactionInfo {
		return d.getInProgressCompactionInfoLocked(nil)
	}); err != nil {
		return nil, err
	}
	d.updateReadStateLocked(d.opts.DebugCheck)
	d.updateTableStatsLocked(ve.NewFiles)
	d.deleteObsoleteFiles(jobID, false /* waitForOngoing */)
	// The ingestion may have pushed a level over the threshold for compaction,
	// so check to see if one is necessary and schedule it.
	d.maybeScheduleCompaction()
	if scheduleFlush {
		d.maybeScheduleFlush()
	}
	d.maybeValidateSSTablesLocked(ve.NewFiles)
	return ve, nil
}

// maybeValidateSSTablesLocked adds the slice of newFileEntrys to the pending
// queue of files to be validated, when the feature is enabled.
// DB.mu must be locked when calling.
func (d *DB) maybeValidateSSTablesLocked(newFiles []newFileEntry) {
	// Only add to the validation queue when the feature is enabled.
	if !d.opts.Experimental.ValidateOnIngest {
		return
	}

	d.mu.tableValidation.pending = append(d.mu.tableValidation.pending, newFiles...)
	if d.shouldValidateSSTablesLocked() {
		go d.validateSSTables()
	}
}

// shouldValidateSSTablesLocked returns true if SSTable validation should run.
// DB.mu must be locked when calling.
func (d *DB) shouldValidateSSTablesLocked() bool {
	return !d.mu.tableValidation.validating &&
		d.closed.Load() == nil &&
		d.opts.Experimental.ValidateOnIngest &&
		len(d.mu.tableValidation.pending) > 0
}

// validateSSTables runs a round of validation on the tables in the pending
// queue.
func (d *DB) validateSSTables() {
	d.mu.Lock()
	if !d.shouldValidateSSTablesLocked() {
		d.mu.Unlock()
		return
	}

	pending := d.mu.tableValidation.pending
	d.mu.tableValidation.pending = nil
	d.mu.tableValidation.validating = true
	jobID := d.mu.nextJobID
	d.mu.nextJobID++
	rs := d.loadReadState()

	// Drop DB.mu before performing IO.
	d.mu.Unlock()

	// Validate all tables in the pending queue. This could lead to a situation
	// where we are starving IO from other tasks due to having to page through
	// all the blocks in all the sstables in the queue.
	// TODO(travers): Add some form of pacing to avoid IO starvation.
	for _, f := range pending {
		// The file may have been moved or deleted since it was ingested, in
		// which case we skip.
		if !rs.current.Contains(f.Level, d.cmp, f.Meta) {
			// Assume the file was moved to a lower level. It is rare enough
			// that a table is moved or deleted between the time it was ingested
			// and the time the validation routine runs that the overall cost of
			// this inner loop is tolerably low, when amortized over all
			// ingested tables.
			found := false
			for i := f.Level + 1; i < numLevels; i++ {
				if rs.current.Contains(i, d.cmp, f.Meta) {
					found = true
					break
				}
			}
			if !found {
				continue
			}
		}

		err := d.tableCache.withReader(f.Meta, func(r *sstable.Reader) error {
			return r.ValidateBlockChecksums()
		})
		if err != nil {
			// TODO(travers): Hook into the corruption reporting pipeline, once
			// available. See pebble#1192.
			d.opts.Logger.Fatalf("pebble: encountered corruption during ingestion: %s", err)
		}

		d.opts.EventListener.TableValidated(TableValidatedInfo{
			JobID: jobID,
			Meta:  f.Meta,
		})
	}
	rs.unref()

	d.mu.Lock()
	defer d.mu.Unlock()
	d.mu.tableValidation.validating = false
	d.mu.tableValidation.cond.Broadcast()
	if d.shouldValidateSSTablesLocked() {
		go d.validateSSTables()
	}
}