[WIP] Partial bot skeleton, non-functional though

vendor/github.com/cockroachdb/pebble/internal/manifest/level.go

@@ -0,0 +1,46 @@
+// Copyright 2020 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 manifest
+
+import "fmt"
+
+const (
+	// 3 bits are necessary to represent level values from 0-6.
+	levelBits = 3
+	levelMask = (1 << levelBits) - 1
+	// invalidSublevel denotes an invalid or non-applicable sublevel.
+	invalidSublevel = -1
+)
+
+// Level encodes a level and optional sublevel for use in log and error
+// messages. The encoding has the property that Level(0) ==
+// L0Sublevel(invalidSublevel).
+type Level uint32
+
+func makeLevel(level, sublevel int) Level {
+	return Level(((sublevel + 1) << levelBits) | level)
+}
+
+// LevelToInt returns the int representation of a Level
+func LevelToInt(l Level) int {
+	return int(l) & levelMask
+}
+
+// L0Sublevel returns a Level representing the specified L0 sublevel.
+func L0Sublevel(sublevel int) Level {
+	if sublevel < 0 {
+		panic(fmt.Sprintf("invalid L0 sublevel: %d", sublevel))
+	}
+	return makeLevel(0, sublevel)
+}
+
+func (l Level) String() string {
+	level := int(l) & levelMask
+	sublevel := (int(l) >> levelBits) - 1
+	if sublevel != invalidSublevel {
+		return fmt.Sprintf("L%d.%d", level, sublevel)
+	}
+	return fmt.Sprintf("L%d", level)
+}

vendor/github.com/cockroachdb/pebble/internal/manifest/level_metadata.go

@@ -0,0 +1,681 @@
+// Copyright 2020 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 manifest
+
+import (
+	"bytes"
+	"fmt"
+
+	"github.com/cockroachdb/pebble/internal/base"
+	"github.com/cockroachdb/pebble/internal/invariants"
+)
+
+// LevelMetadata contains metadata for all of the files within
+// a level of the LSM.
+type LevelMetadata struct {
+	level int
+	tree  btree
+}
+
+// clone makes a copy of the level metadata, implicitly increasing the ref
+// count of every file contained within lm.
+func (lm *LevelMetadata) clone() LevelMetadata {
+	return LevelMetadata{
+		level: lm.level,
+		tree:  lm.tree.Clone(),
+	}
+}
+
+func (lm *LevelMetadata) release() (obsolete []*FileMetadata) {
+	return lm.tree.Release()
+}
+
+func makeLevelMetadata(cmp Compare, level int, files []*FileMetadata) LevelMetadata {
+	bcmp := btreeCmpSeqNum
+	if level > 0 {
+		bcmp = btreeCmpSmallestKey(cmp)
+	}
+	var lm LevelMetadata
+	lm.level = level
+	lm.tree, _ = makeBTree(bcmp, files)
+	return lm
+}
+
+func makeBTree(cmp btreeCmp, files []*FileMetadata) (btree, LevelSlice) {
+	var t btree
+	t.cmp = cmp
+	for _, f := range files {
+		t.Insert(f)
+	}
+	return t, newLevelSlice(t.Iter())
+}
+
+// Empty indicates whether there are any files in the level.
+func (lm *LevelMetadata) Empty() bool {
+	return lm.tree.Count() == 0
+}
+
+// Len returns the number of files within the level.
+func (lm *LevelMetadata) Len() int {
+	return lm.tree.Count()
+}
+
+// Iter constructs a LevelIterator over the entire level.
+func (lm *LevelMetadata) Iter() LevelIterator {
+	return LevelIterator{iter: lm.tree.Iter()}
+}
+
+// Slice constructs a slice containing the entire level.
+func (lm *LevelMetadata) Slice() LevelSlice {
+	return newLevelSlice(lm.tree.Iter())
+}
+
+// Find finds the provided file in the level if it exists.
+func (lm *LevelMetadata) Find(cmp base.Compare, m *FileMetadata) *LevelFile {
+	iter := lm.Iter()
+	if lm.level != 0 {
+		// If lm holds files for levels >0, we can narrow our search by binary
+		// searching by bounds.
+		o := overlaps(iter, cmp, m.Smallest.UserKey,
+			m.Largest.UserKey, m.Largest.IsExclusiveSentinel())
+		iter = o.Iter()
+	}
+	for f := iter.First(); f != nil; f = iter.Next() {
+		if f == m {
+			lf := iter.Take()
+			return &lf
+		}
+	}
+	return nil
+}
+
+// Annotation lazily calculates and returns the annotation defined by
+// Annotator. The Annotator is used as the key for pre-calculated
+// values, so equal Annotators must be used to avoid duplicate computations
+// and cached annotations. Annotation must not be called concurrently, and in
+// practice this is achieved by requiring callers to hold DB.mu.
+func (lm *LevelMetadata) Annotation(annotator Annotator) interface{} {
+	if lm.Empty() {
+		return annotator.Zero(nil)
+	}
+	v, _ := lm.tree.root.Annotation(annotator)
+	return v
+}
+
+// InvalidateAnnotation clears any cached annotations defined by Annotator. The
+// Annotator is used as the key for pre-calculated values, so equal Annotators
+// must be used to clear the appropriate cached annotation. InvalidateAnnotation
+// must not be called concurrently, and in practice this is achieved by
+// requiring callers to hold DB.mu.
+func (lm *LevelMetadata) InvalidateAnnotation(annotator Annotator) {
+	if lm.Empty() {
+		return
+	}
+	lm.tree.root.InvalidateAnnotation(annotator)
+}
+
+// LevelFile holds a file's metadata along with its position
+// within a level of the LSM.
+type LevelFile struct {
+	*FileMetadata
+	slice LevelSlice
+}
+
+// Slice constructs a LevelSlice containing only this file.
+func (lf LevelFile) Slice() LevelSlice {
+	return lf.slice
+}
+
+// NewLevelSliceSeqSorted constructs a LevelSlice over the provided files,
+// sorted by the L0 sequence number sort order.
+// TODO(jackson): Can we improve this interface or avoid needing to export
+// a slice constructor like this?
+func NewLevelSliceSeqSorted(files []*FileMetadata) LevelSlice {
+	tr, slice := makeBTree(btreeCmpSeqNum, files)
+	tr.Release()
+	slice.verifyInvariants()
+	return slice
+}
+
+// NewLevelSliceKeySorted constructs a LevelSlice over the provided files,
+// sorted by the files smallest keys.
+// TODO(jackson): Can we improve this interface or avoid needing to export
+// a slice constructor like this?
+func NewLevelSliceKeySorted(cmp base.Compare, files []*FileMetadata) LevelSlice {
+	tr, slice := makeBTree(btreeCmpSmallestKey(cmp), files)
+	tr.Release()
+	slice.verifyInvariants()
+	return slice
+}
+
+// NewLevelSliceSpecificOrder constructs a LevelSlice over the provided files,
+// ordering the files by their order in the provided slice. It's used in
+// tests.
+// TODO(jackson): Update tests to avoid requiring this and remove it.
+func NewLevelSliceSpecificOrder(files []*FileMetadata) LevelSlice {
+	tr, slice := makeBTree(btreeCmpSpecificOrder(files), files)
+	tr.Release()
+	slice.verifyInvariants()
+	return slice
+}
+
+// newLevelSlice constructs a new LevelSlice backed by iter.
+func newLevelSlice(iter iterator) LevelSlice {
+	s := LevelSlice{iter: iter}
+	if iter.r != nil {
+		s.length = iter.r.subtreeCount
+	}
+	s.verifyInvariants()
+	return s
+}
+
+// newBoundedLevelSlice constructs a new LevelSlice backed by iter and bounded
+// by the provided start and end bounds. The provided startBound and endBound
+// iterators must be iterators over the same B-Tree. Both start and end bounds
+// are inclusive.
+func newBoundedLevelSlice(iter iterator, startBound, endBound *iterator) LevelSlice {
+	s := LevelSlice{
+		iter:  iter,
+		start: startBound,
+		end:   endBound,
+	}
+	if iter.valid() {
+		s.length = endBound.countLeft() - startBound.countLeft()
+		// NB: The +1 is a consequence of the end bound being inclusive.
+		if endBound.valid() {
+			s.length++
+		}
+		// NB: A slice that's empty due to its bounds may have an endBound
+		// positioned before the startBound due to the inclusive bounds.
+		// TODO(jackson): Consider refactoring the end boundary to be exclusive;
+		// it would simplify some areas (eg, here) and complicate others (eg,
+		// Reslice-ing to grow compactions).
+		if s.length < 0 {
+			s.length = 0
+		}
+	}
+	s.verifyInvariants()
+	return s
+}
+
+// LevelSlice contains a slice of the files within a level of the LSM.
+// A LevelSlice is immutable once created, but may be used to construct a
+// mutable LevelIterator over the slice's files.
+//
+// LevelSlices should be constructed through one of the existing constructors,
+// not manually initialized.
+type LevelSlice struct {
+	iter   iterator
+	length int
+	// start and end form the inclusive bounds of a slice of files within a
+	// level of the LSM. They may be nil if the entire B-Tree backing iter is
+	// accessible.
+	start *iterator
+	end   *iterator
+}
+
+func (ls LevelSlice) verifyInvariants() {
+	if invariants.Enabled {
+		i := ls.Iter()
+		var length int
+		for f := i.First(); f != nil; f = i.Next() {
+			length++
+		}
+		if ls.length != length {
+			panic(fmt.Sprintf("LevelSlice %s has length %d value; actual length is %d", ls, ls.length, length))
+		}
+	}
+}
+
+// Each invokes fn for each element in the slice.
+func (ls LevelSlice) Each(fn func(*FileMetadata)) {
+	iter := ls.Iter()
+	for f := iter.First(); f != nil; f = iter.Next() {
+		fn(f)
+	}
+}
+
+// String implements fmt.Stringer.
+func (ls LevelSlice) String() string {
+	var buf bytes.Buffer
+	fmt.Fprintf(&buf, "%d files: ", ls.length)
+	ls.Each(func(f *FileMetadata) {
+		if buf.Len() > 0 {
+			fmt.Fprintf(&buf, " ")
+		}
+		fmt.Fprint(&buf, f)
+	})
+	return buf.String()
+}
+
+// Empty indicates whether the slice contains any files.
+func (ls *LevelSlice) Empty() bool {
+	return emptyWithBounds(ls.iter, ls.start, ls.end)
+}
+
+// Iter constructs a LevelIterator that iterates over the slice.
+func (ls *LevelSlice) Iter() LevelIterator {
+	return LevelIterator{
+		start: ls.start,
+		end:   ls.end,
+		iter:  ls.iter.clone(),
+	}
+}
+
+// Len returns the number of files in the slice. Its runtime is constant.
+func (ls *LevelSlice) Len() int {
+	return ls.length
+}
+
+// SizeSum sums the size of all files in the slice. Its runtime is linear in
+// the length of the slice.
+func (ls *LevelSlice) SizeSum() uint64 {
+	var sum uint64
+	iter := ls.Iter()
+	for f := iter.First(); f != nil; f = iter.Next() {
+		sum += f.Size
+	}
+	return sum
+}
+
+// Reslice constructs a new slice backed by the same underlying level, with
+// new start and end positions. Reslice invokes the provided function, passing
+// two LevelIterators: one positioned to i's inclusive start and one
+// positioned to i's inclusive end. The resliceFunc may move either iterator
+// forward or backwards, including beyond the callee's original bounds to
+// capture additional files from the underlying level. Reslice constructs and
+// returns a new LevelSlice with the final bounds of the iterators after
+// calling resliceFunc.
+func (ls LevelSlice) Reslice(resliceFunc func(start, end *LevelIterator)) LevelSlice {
+	if ls.iter.r == nil {
+		return ls
+	}
+	var start, end LevelIterator
+	if ls.start == nil {
+		start.iter = ls.iter.clone()
+		start.iter.first()
+	} else {
+		start.iter = ls.start.clone()
+	}
+	if ls.end == nil {
+		end.iter = ls.iter.clone()
+		end.iter.last()
+	} else {
+		end.iter = ls.end.clone()
+	}
+	resliceFunc(&start, &end)
+	return newBoundedLevelSlice(start.iter.clone(), &start.iter, &end.iter)
+}
+
+// KeyType is used to specify the type of keys we're looking for in
+// LevelIterator positioning operations. Files not containing any keys of the
+// desired type are skipped.
+type KeyType int8
+
+const (
+	// KeyTypePointAndRange denotes a search among the entire keyspace, including
+	// both point keys and range keys. No sstables are skipped.
+	KeyTypePointAndRange KeyType = iota
+	// KeyTypePoint denotes a search among the point keyspace. SSTables with no
+	// point keys will be skipped. Note that the point keyspace includes rangedels.
+	KeyTypePoint
+	// KeyTypeRange denotes a search among the range keyspace. SSTables with no
+	// range keys will be skipped.
+	KeyTypeRange
+)
+
+type keyTypeAnnotator struct{}
+
+var _ Annotator = keyTypeAnnotator{}
+
+func (k keyTypeAnnotator) Zero(dst interface{}) interface{} {
+	var val *KeyType
+	if dst != nil {
+		val = dst.(*KeyType)
+	} else {
+		val = new(KeyType)
+	}
+	*val = KeyTypePoint
+	return val
+}
+
+func (k keyTypeAnnotator) Accumulate(m *FileMetadata, dst interface{}) (interface{}, bool) {
+	v := dst.(*KeyType)
+	switch *v {
+	case KeyTypePoint:
+		if m.HasRangeKeys {
+			*v = KeyTypePointAndRange
+		}
+	case KeyTypePointAndRange:
+		// Do nothing.
+	default:
+		panic("unexpected key type")
+	}
+	return v, true
+}
+
+func (k keyTypeAnnotator) Merge(src interface{}, dst interface{}) interface{} {
+	v := dst.(*KeyType)
+	srcVal := src.(*KeyType)
+	switch *v {
+	case KeyTypePoint:
+		if *srcVal == KeyTypePointAndRange {
+			*v = KeyTypePointAndRange
+		}
+	case KeyTypePointAndRange:
+		// Do nothing.
+	default:
+		panic("unexpected key type")
+	}
+	return v
+}
+
+// LevelIterator iterates over a set of files' metadata. Its zero value is an
+// empty iterator.
+type LevelIterator struct {
+	iter   iterator
+	start  *iterator
+	end    *iterator
+	filter KeyType
+}
+
+func (i LevelIterator) String() string {
+	var buf bytes.Buffer
+	iter := i.iter.clone()
+	iter.first()
+	iter.prev()
+	if i.iter.pos == -1 {
+		fmt.Fprint(&buf, "(<start>)*")
+	}
+	iter.next()
+	for ; iter.valid(); iter.next() {
+		if buf.Len() > 0 {
+			fmt.Fprint(&buf, "   ")
+		}
+
+		if i.start != nil && cmpIter(iter, *i.start) == 0 {
+			fmt.Fprintf(&buf, " [ ")
+		}
+		isCurrentPos := cmpIter(iter, i.iter) == 0
+		if isCurrentPos {
+			fmt.Fprint(&buf, " ( ")
+		}
+		fmt.Fprint(&buf, iter.cur().String())
+		if isCurrentPos {
+			fmt.Fprint(&buf, " )*")
+		}
+		if i.end != nil && cmpIter(iter, *i.end) == 0 {
+			fmt.Fprintf(&buf, " ]")
+		}
+	}
+	if i.iter.n != nil && i.iter.pos >= i.iter.n.count {
+		if buf.Len() > 0 {
+			fmt.Fprint(&buf, "   ")
+		}
+		fmt.Fprint(&buf, "(<end>)*")
+	}
+	return buf.String()
+}
+
+// Clone copies the iterator, returning an independent iterator at the same
+// position.
+func (i *LevelIterator) Clone() LevelIterator {
+	if i.iter.r == nil {
+		return *i
+	}
+	// The start and end iterators are not cloned and are treated as
+	// immutable.
+	return LevelIterator{
+		iter:   i.iter.clone(),
+		start:  i.start,
+		end:    i.end,
+		filter: i.filter,
+	}
+}
+
+// Current returns the item at the current iterator position.
+//
+// Current is deprecated. Callers should instead use the return value of a
+// positioning operation.
+func (i *LevelIterator) Current() *FileMetadata {
+	if !i.iter.valid() ||
+		(i.end != nil && cmpIter(i.iter, *i.end) > 0) ||
+		(i.start != nil && cmpIter(i.iter, *i.start) < 0) {
+		return nil
+	}
+	return i.iter.cur()
+}
+
+func (i *LevelIterator) empty() bool {
+	return emptyWithBounds(i.iter, i.start, i.end)
+}
+
+// Filter clones the iterator and sets the desired KeyType as the key to filter
+// files on.
+func (i *LevelIterator) Filter(keyType KeyType) LevelIterator {
+	l := i.Clone()
+	l.filter = keyType
+	return l
+}
+
+func emptyWithBounds(i iterator, start, end *iterator) bool {
+	// If i.r is nil, the iterator was constructed from an empty btree.
+	// If the end bound is before the start bound, the bounds represent an
+	// empty slice of the B-Tree.
+	return i.r == nil || (start != nil && end != nil && cmpIter(*end, *start) < 0)
+}
+
+// First seeks to the first file in the iterator and returns it.
+func (i *LevelIterator) First() *FileMetadata {
+	if i.empty() {
+		return nil
+	}
+	if i.start != nil {
+		i.iter = i.start.clone()
+	} else {
+		i.iter.first()
+	}
+	if !i.iter.valid() {
+		return nil
+	}
+	return i.skipFilteredForward(i.iter.cur())
+}
+
+// Last seeks to the last file in the iterator and returns it.
+func (i *LevelIterator) Last() *FileMetadata {
+	if i.empty() {
+		return nil
+	}
+	if i.end != nil {
+		i.iter = i.end.clone()
+	} else {
+		i.iter.last()
+	}
+	if !i.iter.valid() {
+		return nil
+	}
+	return i.skipFilteredBackward(i.iter.cur())
+}
+
+// Next advances the iterator to the next file and returns it.
+func (i *LevelIterator) Next() *FileMetadata {
+	if i.iter.r == nil {
+		return nil
+	}
+	if invariants.Enabled && (i.iter.pos >= i.iter.n.count || (i.end != nil && cmpIter(i.iter, *i.end) > 0)) {
+		panic("pebble: cannot next forward-exhausted iterator")
+	}
+	i.iter.next()
+	if !i.iter.valid() {
+		return nil
+	}
+	return i.skipFilteredForward(i.iter.cur())
+}
+
+// Prev moves the iterator the previous file and returns it.
+func (i *LevelIterator) Prev() *FileMetadata {
+	if i.iter.r == nil {
+		return nil
+	}
+	if invariants.Enabled && (i.iter.pos < 0 || (i.start != nil && cmpIter(i.iter, *i.start) < 0)) {
+		panic("pebble: cannot prev backward-exhausted iterator")
+	}
+	i.iter.prev()
+	if !i.iter.valid() {
+		return nil
+	}
+	return i.skipFilteredBackward(i.iter.cur())
+}
+
+// SeekGE seeks to the first file in the iterator's file set with a largest
+// user key greater than or equal to the provided user key. The iterator must
+// have been constructed from L1+, because it requires the underlying files to
+// be sorted by user keys and non-overlapping.
+func (i *LevelIterator) SeekGE(cmp Compare, userKey []byte) *FileMetadata {
+	// TODO(jackson): Assert that i.iter.cmp == btreeCmpSmallestKey.
+	if i.iter.r == nil {
+		return nil
+	}
+	m := i.seek(func(m *FileMetadata) bool {
+		return cmp(m.Largest.UserKey, userKey) >= 0
+	})
+	if i.filter != KeyTypePointAndRange && m != nil {
+		b, ok := m.LargestBound(i.filter)
+		if !ok {
+			m = i.Next()
+		} else if c := cmp(b.UserKey, userKey); c < 0 || c == 0 && b.IsExclusiveSentinel() {
+			// This file does not contain any keys of the type ≥ lower. It
+			// should be filtered, even though it does contain point keys.
+			m = i.Next()
+		}
+	}
+	return i.skipFilteredForward(m)
+}
+
+// SeekLT seeks to the last file in the iterator's file set with a smallest
+// user key less than the provided user key. The iterator must have been
+// constructed from L1+, because it requires the underlying files to be sorted
+// by user keys and non-overlapping.
+func (i *LevelIterator) SeekLT(cmp Compare, userKey []byte) *FileMetadata {
+	// TODO(jackson): Assert that i.iter.cmp == btreeCmpSmallestKey.
+	if i.iter.r == nil {
+		return nil
+	}
+	i.seek(func(m *FileMetadata) bool {
+		return cmp(m.Smallest.UserKey, userKey) >= 0
+	})
+	m := i.Prev()
+	// Although i.Prev() guarantees that the current file contains keys of the
+	// relevant type, it doesn't guarantee that the keys of the relevant type
+	// are < userKey.
+	if i.filter != KeyTypePointAndRange && m != nil {
+		b, ok := m.SmallestBound(i.filter)
+		if !ok {
+			panic("unreachable")
+		}
+		if c := cmp(b.UserKey, userKey); c >= 0 {
+			// This file does not contain any keys of the type ≥ lower. It
+			// should be filtered, even though it does contain point keys.
+			m = i.Prev()
+		}
+	}
+	return i.skipFilteredBackward(m)
+}
+
+// skipFilteredForward takes the file metadata at the iterator's current
+// position, and skips forward if the current key-type filter (i.filter)
+// excludes the file. It skips until it finds an unfiltered file or exhausts the
+// level. If lower is != nil, skipFilteredForward skips any files that do not
+// contain keys with the provided key-type ≥ lower.
+//
+// skipFilteredForward also enforces the upper bound, returning nil if at any
+// point the upper bound is exceeded.
+func (i *LevelIterator) skipFilteredForward(meta *FileMetadata) *FileMetadata {
+	for meta != nil && !meta.ContainsKeyType(i.filter) {
+		i.iter.next()
+		if !i.iter.valid() {
+			meta = nil
+		} else {
+			meta = i.iter.cur()
+		}
+	}
+	if meta != nil && i.end != nil && cmpIter(i.iter, *i.end) > 0 {
+		// Exceeded upper bound.
+		meta = nil
+	}
+	return meta
+}
+
+// skipFilteredBackward takes the file metadata at the iterator's current
+// position, and skips backward if the current key-type filter (i.filter)
+// excludes the file. It skips until it finds an unfiltered file or exhausts the
+// level. If upper is != nil, skipFilteredBackward skips any files that do not
+// contain keys with the provided key-type < upper.
+//
+// skipFilteredBackward also enforces the lower bound, returning nil if at any
+// point the lower bound is exceeded.
+func (i *LevelIterator) skipFilteredBackward(meta *FileMetadata) *FileMetadata {
+	for meta != nil && !meta.ContainsKeyType(i.filter) {
+		i.iter.prev()
+		if !i.iter.valid() {
+			meta = nil
+		} else {
+			meta = i.iter.cur()
+		}
+	}
+	if meta != nil && i.start != nil && cmpIter(i.iter, *i.start) < 0 {
+		// Exceeded lower bound.
+		meta = nil
+	}
+	return meta
+}
+
+func (i *LevelIterator) seek(fn func(*FileMetadata) bool) *FileMetadata {
+	i.iter.seek(fn)
+
+	// i.iter.seek seeked in the unbounded underlying B-Tree. If the iterator
+	// has start or end bounds, we may have exceeded them. Reset to the bounds
+	// if necessary.
+	//
+	// NB: The LevelIterator and LevelSlice semantics require that a bounded
+	// LevelIterator/LevelSlice containing files x0, x1, ..., xn behave
+	// identically to an unbounded LevelIterator/LevelSlice of a B-Tree
+	// containing x0, x1, ..., xn. In other words, any files outside the
+	// LevelIterator's bounds should not influence the iterator's behavior.
+	// When seeking, this means a SeekGE that seeks beyond the end bound,
+	// followed by a Prev should return the last element within bounds.
+	if i.end != nil && cmpIter(i.iter, *i.end) > 0 {
+		i.iter = i.end.clone()
+		// Since seek(fn) positioned beyond i.end, we know there is nothing to
+		// return within bounds.
+		i.iter.next()
+		return nil
+	} else if i.start != nil && cmpIter(i.iter, *i.start) < 0 {
+		i.iter = i.start.clone()
+	}
+	if !i.iter.valid() {
+		return nil
+	}
+	return i.iter.cur()
+}
+
+// Take constructs a LevelFile containing the file at the iterator's current
+// position. Take panics if the iterator is not currently positioned over a
+// file.
+func (i *LevelIterator) Take() LevelFile {
+	m := i.Current()
+	if m == nil {
+		panic("Take called on invalid LevelIterator")
+	}
+	// LevelSlice's start and end fields are immutable and are positioned to
+	// the same position for a LevelFile because they're inclusive, so we can
+	// share one iterator stack between the two bounds.
+	boundsIter := i.iter.clone()
+	s := newBoundedLevelSlice(i.iter.clone(), &boundsIter, &boundsIter)
+	return LevelFile{
+		FileMetadata: m,
+		slice:        s,
+	}
+}

vendor/github.com/cockroachdb/pebble/internal/manifest/version_edit.go

@@ -0,0 +1,842 @@
+// Copyright 2012 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 manifest
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"io"
+	"sort"
+	"sync/atomic"
+	"time"
+
+	"github.com/cockroachdb/errors"
+	"github.com/cockroachdb/pebble/internal/base"
+)
+
+// TODO(peter): describe the MANIFEST file format, independently of the C++
+// project.
+
+var errCorruptManifest = base.CorruptionErrorf("pebble: corrupt manifest")
+
+type byteReader interface {
+	io.ByteReader
+	io.Reader
+}
+
+// Tags for the versionEdit disk format.
+// Tag 8 is no longer used.
+const (
+	// LevelDB tags.
+	tagComparator     = 1
+	tagLogNumber      = 2
+	tagNextFileNumber = 3
+	tagLastSequence   = 4
+	tagCompactPointer = 5
+	tagDeletedFile    = 6
+	tagNewFile        = 7
+	tagPrevLogNumber  = 9
+
+	// RocksDB tags.
+	tagNewFile2         = 100
+	tagNewFile3         = 102
+	tagNewFile4         = 103
+	tagColumnFamily     = 200
+	tagColumnFamilyAdd  = 201
+	tagColumnFamilyDrop = 202
+	tagMaxColumnFamily  = 203
+
+	// Pebble tags.
+	tagNewFile5 = 104 // Range keys.
+
+	// The custom tags sub-format used by tagNewFile4 and above.
+	customTagTerminate         = 1
+	customTagNeedsCompaction   = 2
+	customTagCreationTime      = 6
+	customTagPathID            = 65
+	customTagNonSafeIgnoreMask = 1 << 6
+)
+
+// DeletedFileEntry holds the state for a file deletion from a level. The file
+// itself might still be referenced by another level.
+type DeletedFileEntry struct {
+	Level   int
+	FileNum base.FileNum
+}
+
+// NewFileEntry holds the state for a new file or one moved from a different
+// level.
+type NewFileEntry struct {
+	Level int
+	Meta  *FileMetadata
+}
+
+// VersionEdit holds the state for an edit to a Version along with other
+// on-disk state (log numbers, next file number, and the last sequence number).
+type VersionEdit struct {
+	// ComparerName is the value of Options.Comparer.Name. This is only set in
+	// the first VersionEdit in a manifest (either when the DB is created, or
+	// when a new manifest is created) and is used to verify that the comparer
+	// specified at Open matches the comparer that was previously used.
+	ComparerName string
+
+	// MinUnflushedLogNum is the smallest WAL log file number corresponding to
+	// mutations that have not been flushed to an sstable.
+	//
+	// This is an optional field, and 0 represents it is not set.
+	MinUnflushedLogNum base.FileNum
+
+	// ObsoletePrevLogNum is a historic artifact from LevelDB that is not used by
+	// Pebble, RocksDB, or even LevelDB. Its use in LevelDB was deprecated in
+	// 6/2011. We keep it around purely for informational purposes when
+	// displaying MANIFEST contents.
+	ObsoletePrevLogNum uint64
+
+	// The next file number. A single counter is used to assign file numbers
+	// for the WAL, MANIFEST, sstable, and OPTIONS files.
+	NextFileNum base.FileNum
+
+	// LastSeqNum is an upper bound on the sequence numbers that have been
+	// assigned in flushed WALs. Unflushed WALs (that will be replayed during
+	// recovery) may contain sequence numbers greater than this value.
+	LastSeqNum uint64
+
+	// A file num may be present in both deleted files and new files when it
+	// is moved from a lower level to a higher level (when the compaction
+	// found that there was no overlapping file at the higher level).
+	DeletedFiles map[DeletedFileEntry]*FileMetadata
+	NewFiles     []NewFileEntry
+}
+
+// Decode decodes an edit from the specified reader.
+func (v *VersionEdit) Decode(r io.Reader) error {
+	br, ok := r.(byteReader)
+	if !ok {
+		br = bufio.NewReader(r)
+	}
+	d := versionEditDecoder{br}
+	for {
+		tag, err := binary.ReadUvarint(br)
+		if err == io.EOF {
+			break
+		}
+		if err != nil {
+			return err
+		}
+		switch tag {
+		case tagComparator:
+			s, err := d.readBytes()
+			if err != nil {
+				return err
+			}
+			v.ComparerName = string(s)
+
+		case tagLogNumber:
+			n, err := d.readFileNum()
+			if err != nil {
+				return err
+			}
+			v.MinUnflushedLogNum = n
+
+		case tagNextFileNumber:
+			n, err := d.readFileNum()
+			if err != nil {
+				return err
+			}
+			v.NextFileNum = n
+
+		case tagLastSequence:
+			n, err := d.readUvarint()
+			if err != nil {
+				return err
+			}
+			v.LastSeqNum = n
+
+		case tagCompactPointer:
+			if _, err := d.readLevel(); err != nil {
+				return err
+			}
+			if _, err := d.readBytes(); err != nil {
+				return err
+			}
+			// NB: RocksDB does not use compaction pointers anymore.
+
+		case tagDeletedFile:
+			level, err := d.readLevel()
+			if err != nil {
+				return err
+			}
+			fileNum, err := d.readFileNum()
+			if err != nil {
+				return err
+			}
+			if v.DeletedFiles == nil {
+				v.DeletedFiles = make(map[DeletedFileEntry]*FileMetadata)
+			}
+			v.DeletedFiles[DeletedFileEntry{level, fileNum}] = nil
+
+		case tagNewFile, tagNewFile2, tagNewFile3, tagNewFile4, tagNewFile5:
+			level, err := d.readLevel()
+			if err != nil {
+				return err
+			}
+			fileNum, err := d.readFileNum()
+			if err != nil {
+				return err
+			}
+			if tag == tagNewFile3 {
+				// The pathID field appears unused in RocksDB.
+				_ /* pathID */, err := d.readUvarint()
+				if err != nil {
+					return err
+				}
+			}
+			size, err := d.readUvarint()
+			if err != nil {
+				return err
+			}
+			// We read the smallest / largest key bounds differently depending on
+			// whether we have point, range or both types of keys present in the
+			// table.
+			var (
+				smallestPointKey, largestPointKey []byte
+				smallestRangeKey, largestRangeKey []byte
+				parsedPointBounds                 bool
+				boundsMarker                      byte
+			)
+			if tag != tagNewFile5 {
+				// Range keys not present in the table. Parse the point key bounds.
+				smallestPointKey, err = d.readBytes()
+				if err != nil {
+					return err
+				}
+				largestPointKey, err = d.readBytes()
+				if err != nil {
+					return err
+				}
+			} else {
+				// Range keys are present in the table. Determine whether we have point
+				// keys to parse, in addition to the bounds.
+				boundsMarker, err = d.ReadByte()
+				if err != nil {
+					return err
+				}
+				// Parse point key bounds, if present.
+				if boundsMarker&maskContainsPointKeys > 0 {
+					smallestPointKey, err = d.readBytes()
+					if err != nil {
+						return err
+					}
+					largestPointKey, err = d.readBytes()
+					if err != nil {
+						return err
+					}
+					parsedPointBounds = true
+				} else {
+					// The table does not have point keys.
+					// Sanity check: the bounds must be range keys.
+					if boundsMarker&maskSmallest != 0 || boundsMarker&maskLargest != 0 {
+						return base.CorruptionErrorf(
+							"new-file-4-range-keys: table without point keys has point key bounds: marker=%x",
+							boundsMarker,
+						)
+					}
+				}
+				// Parse range key bounds.
+				smallestRangeKey, err = d.readBytes()
+				if err != nil {
+					return err
+				}
+				largestRangeKey, err = d.readBytes()
+				if err != nil {
+					return err
+				}
+			}
+			var smallestSeqNum uint64
+			var largestSeqNum uint64
+			if tag != tagNewFile {
+				smallestSeqNum, err = d.readUvarint()
+				if err != nil {
+					return err
+				}
+				largestSeqNum, err = d.readUvarint()
+				if err != nil {
+					return err
+				}
+			}
+			var markedForCompaction bool
+			var creationTime uint64
+			if tag == tagNewFile4 || tag == tagNewFile5 {
+				for {
+					customTag, err := d.readUvarint()
+					if err != nil {
+						return err
+					}
+					if customTag == customTagTerminate {
+						break
+					}
+					field, err := d.readBytes()
+					if err != nil {
+						return err
+					}
+					switch customTag {
+					case customTagNeedsCompaction:
+						if len(field) != 1 {
+							return base.CorruptionErrorf("new-file4: need-compaction field wrong size")
+						}
+						markedForCompaction = (field[0] == 1)
+
+					case customTagCreationTime:
+						var n int
+						creationTime, n = binary.Uvarint(field)
+						if n != len(field) {
+							return base.CorruptionErrorf("new-file4: invalid file creation time")
+						}
+
+					case customTagPathID:
+						return base.CorruptionErrorf("new-file4: path-id field not supported")
+
+					default:
+						if (customTag & customTagNonSafeIgnoreMask) != 0 {
+							return base.CorruptionErrorf("new-file4: custom field not supported: %d", customTag)
+						}
+					}
+				}
+			}
+			m := &FileMetadata{
+				FileNum:             fileNum,
+				Size:                size,
+				CreationTime:        int64(creationTime),
+				SmallestSeqNum:      smallestSeqNum,
+				LargestSeqNum:       largestSeqNum,
+				MarkedForCompaction: markedForCompaction,
+			}
+			if tag != tagNewFile5 { // no range keys present
+				m.SmallestPointKey = base.DecodeInternalKey(smallestPointKey)
+				m.LargestPointKey = base.DecodeInternalKey(largestPointKey)
+				m.HasPointKeys = true
+				m.Smallest, m.Largest = m.SmallestPointKey, m.LargestPointKey
+				m.boundTypeSmallest, m.boundTypeLargest = boundTypePointKey, boundTypePointKey
+			} else { // range keys present
+				// Set point key bounds, if parsed.
+				if parsedPointBounds {
+					m.SmallestPointKey = base.DecodeInternalKey(smallestPointKey)
+					m.LargestPointKey = base.DecodeInternalKey(largestPointKey)
+					m.HasPointKeys = true
+				}
+				// Set range key bounds.
+				m.SmallestRangeKey = base.DecodeInternalKey(smallestRangeKey)
+				m.LargestRangeKey = base.DecodeInternalKey(largestRangeKey)
+				m.HasRangeKeys = true
+				// Set overall bounds (by default assume range keys).
+				m.Smallest, m.Largest = m.SmallestRangeKey, m.LargestRangeKey
+				m.boundTypeSmallest, m.boundTypeLargest = boundTypeRangeKey, boundTypeRangeKey
+				if boundsMarker&maskSmallest == maskSmallest {
+					m.Smallest = m.SmallestPointKey
+					m.boundTypeSmallest = boundTypePointKey
+				}
+				if boundsMarker&maskLargest == maskLargest {
+					m.Largest = m.LargestPointKey
+					m.boundTypeLargest = boundTypePointKey
+				}
+			}
+			m.boundsSet = true
+			v.NewFiles = append(v.NewFiles, NewFileEntry{
+				Level: level,
+				Meta:  m,
+			})
+
+		case tagPrevLogNumber:
+			n, err := d.readUvarint()
+			if err != nil {
+				return err
+			}
+			v.ObsoletePrevLogNum = n
+
+		case tagColumnFamily, tagColumnFamilyAdd, tagColumnFamilyDrop, tagMaxColumnFamily:
+			return base.CorruptionErrorf("column families are not supported")
+
+		default:
+			return errCorruptManifest
+		}
+	}
+	return nil
+}
+
+// String implements fmt.Stringer for a VersionEdit.
+func (v *VersionEdit) String() string {
+	var buf bytes.Buffer
+	if v.ComparerName != "" {
+		fmt.Fprintf(&buf, "  comparer:     %s", v.ComparerName)
+	}
+	if v.MinUnflushedLogNum != 0 {
+		fmt.Fprintf(&buf, "  log-num:       %d\n", v.MinUnflushedLogNum)
+	}
+	if v.ObsoletePrevLogNum != 0 {
+		fmt.Fprintf(&buf, "  prev-log-num:  %d\n", v.ObsoletePrevLogNum)
+	}
+	if v.NextFileNum != 0 {
+		fmt.Fprintf(&buf, "  next-file-num: %d\n", v.NextFileNum)
+	}
+	if v.LastSeqNum != 0 {
+		fmt.Fprintf(&buf, "  last-seq-num:  %d\n", v.LastSeqNum)
+	}
+	entries := make([]DeletedFileEntry, 0, len(v.DeletedFiles))
+	for df := range v.DeletedFiles {
+		entries = append(entries, df)
+	}
+	sort.Slice(entries, func(i, j int) bool {
+		if entries[i].Level != entries[j].Level {
+			return entries[i].Level < entries[j].Level
+		}
+		return entries[i].FileNum < entries[j].FileNum
+	})
+	for _, df := range entries {
+		fmt.Fprintf(&buf, "  deleted:       L%d %s\n", df.Level, df.FileNum)
+	}
+	for _, nf := range v.NewFiles {
+		fmt.Fprintf(&buf, "  added:         L%d %s", nf.Level, nf.Meta.String())
+		if nf.Meta.CreationTime != 0 {
+			fmt.Fprintf(&buf, " (%s)",
+				time.Unix(nf.Meta.CreationTime, 0).UTC().Format(time.RFC3339))
+		}
+		fmt.Fprintln(&buf)
+	}
+	return buf.String()
+}
+
+// Encode encodes an edit to the specified writer.
+func (v *VersionEdit) Encode(w io.Writer) error {
+	e := versionEditEncoder{new(bytes.Buffer)}
+
+	if v.ComparerName != "" {
+		e.writeUvarint(tagComparator)
+		e.writeString(v.ComparerName)
+	}
+	if v.MinUnflushedLogNum != 0 {
+		e.writeUvarint(tagLogNumber)
+		e.writeUvarint(uint64(v.MinUnflushedLogNum))
+	}
+	if v.ObsoletePrevLogNum != 0 {
+		e.writeUvarint(tagPrevLogNumber)
+		e.writeUvarint(v.ObsoletePrevLogNum)
+	}
+	if v.NextFileNum != 0 {
+		e.writeUvarint(tagNextFileNumber)
+		e.writeUvarint(uint64(v.NextFileNum))
+	}
+	// RocksDB requires LastSeqNum to be encoded for the first MANIFEST entry,
+	// even though its value is zero. We detect this by encoding LastSeqNum when
+	// ComparerName is set.
+	if v.LastSeqNum != 0 || v.ComparerName != "" {
+		e.writeUvarint(tagLastSequence)
+		e.writeUvarint(v.LastSeqNum)
+	}
+	for x := range v.DeletedFiles {
+		e.writeUvarint(tagDeletedFile)
+		e.writeUvarint(uint64(x.Level))
+		e.writeUvarint(uint64(x.FileNum))
+	}
+	for _, x := range v.NewFiles {
+		customFields := x.Meta.MarkedForCompaction || x.Meta.CreationTime != 0
+		var tag uint64
+		switch {
+		case x.Meta.HasRangeKeys:
+			tag = tagNewFile5
+		case customFields:
+			tag = tagNewFile4
+		default:
+			tag = tagNewFile2
+		}
+		e.writeUvarint(tag)
+		e.writeUvarint(uint64(x.Level))
+		e.writeUvarint(uint64(x.Meta.FileNum))
+		e.writeUvarint(x.Meta.Size)
+		if !x.Meta.HasRangeKeys {
+			// If we have no range keys, preserve the original format and write the
+			// smallest and largest point keys.
+			e.writeKey(x.Meta.SmallestPointKey)
+			e.writeKey(x.Meta.LargestPointKey)
+		} else {
+			// When range keys are present, we first write a marker byte that
+			// indicates if the table also contains point keys, in addition to how the
+			// overall bounds for the table should be reconstructed. This byte is
+			// followed by the keys themselves.
+			b, err := x.Meta.boundsMarker()
+			if err != nil {
+				return err
+			}
+			if err = e.WriteByte(b); err != nil {
+				return err
+			}
+			// Write point key bounds (if present).
+			if x.Meta.HasPointKeys {
+				e.writeKey(x.Meta.SmallestPointKey)
+				e.writeKey(x.Meta.LargestPointKey)
+			}
+			// Write range key bounds.
+			e.writeKey(x.Meta.SmallestRangeKey)
+			e.writeKey(x.Meta.LargestRangeKey)
+		}
+		e.writeUvarint(x.Meta.SmallestSeqNum)
+		e.writeUvarint(x.Meta.LargestSeqNum)
+		if customFields {
+			if x.Meta.CreationTime != 0 {
+				e.writeUvarint(customTagCreationTime)
+				var buf [binary.MaxVarintLen64]byte
+				n := binary.PutUvarint(buf[:], uint64(x.Meta.CreationTime))
+				e.writeBytes(buf[:n])
+			}
+			if x.Meta.MarkedForCompaction {
+				e.writeUvarint(customTagNeedsCompaction)
+				e.writeBytes([]byte{1})
+			}
+			e.writeUvarint(customTagTerminate)
+		}
+	}
+	_, err := w.Write(e.Bytes())
+	return err
+}
+
+type versionEditDecoder struct {
+	byteReader
+}
+
+func (d versionEditDecoder) readBytes() ([]byte, error) {
+	n, err := d.readUvarint()
+	if err != nil {
+		return nil, err
+	}
+	s := make([]byte, n)
+	_, err = io.ReadFull(d, s)
+	if err != nil {
+		if err == io.ErrUnexpectedEOF {
+			return nil, errCorruptManifest
+		}
+		return nil, err
+	}
+	return s, nil
+}
+
+func (d versionEditDecoder) readLevel() (int, error) {
+	u, err := d.readUvarint()
+	if err != nil {
+		return 0, err
+	}
+	if u >= NumLevels {
+		return 0, errCorruptManifest
+	}
+	return int(u), nil
+}
+
+func (d versionEditDecoder) readFileNum() (base.FileNum, error) {
+	u, err := d.readUvarint()
+	if err != nil {
+		return 0, err
+	}
+	return base.FileNum(u), nil
+}
+
+func (d versionEditDecoder) readUvarint() (uint64, error) {
+	u, err := binary.ReadUvarint(d)
+	if err != nil {
+		if err == io.EOF {
+			return 0, errCorruptManifest
+		}
+		return 0, err
+	}
+	return u, nil
+}
+
+type versionEditEncoder struct {
+	*bytes.Buffer
+}
+
+func (e versionEditEncoder) writeBytes(p []byte) {
+	e.writeUvarint(uint64(len(p)))
+	e.Write(p)
+}
+
+func (e versionEditEncoder) writeKey(k InternalKey) {
+	e.writeUvarint(uint64(k.Size()))
+	e.Write(k.UserKey)
+	buf := k.EncodeTrailer()
+	e.Write(buf[:])
+}
+
+func (e versionEditEncoder) writeString(s string) {
+	e.writeUvarint(uint64(len(s)))
+	e.WriteString(s)
+}
+
+func (e versionEditEncoder) writeUvarint(u uint64) {
+	var buf [binary.MaxVarintLen64]byte
+	n := binary.PutUvarint(buf[:], u)
+	e.Write(buf[:n])
+}
+
+// BulkVersionEdit summarizes the files added and deleted from a set of version
+// edits.
+type BulkVersionEdit struct {
+	Added   [NumLevels][]*FileMetadata
+	Deleted [NumLevels]map[base.FileNum]*FileMetadata
+
+	// AddedByFileNum maps file number to file metadata for all added files
+	// from accumulated version edits. AddedByFileNum is only populated if set
+	// to non-nil by a caller. It must be set to non-nil when replaying
+	// version edits read from a MANIFEST (as opposed to VersionEdits
+	// constructed in-memory).  While replaying a MANIFEST file,
+	// VersionEdit.DeletedFiles map entries have nil values, because the
+	// on-disk deletion record encodes only the file number. Accumulate
+	// uses AddedByFileNum to correctly populate the BulkVersionEdit's Deleted
+	// field with non-nil *FileMetadata.
+	AddedByFileNum map[base.FileNum]*FileMetadata
+
+	// MarkedForCompactionCountDiff holds the aggregated count of files
+	// marked for compaction added or removed.
+	MarkedForCompactionCountDiff int
+}
+
+// Accumulate adds the file addition and deletions in the specified version
+// edit to the bulk edit's internal state.
+func (b *BulkVersionEdit) Accumulate(ve *VersionEdit) error {
+	for df, m := range ve.DeletedFiles {
+		dmap := b.Deleted[df.Level]
+		if dmap == nil {
+			dmap = make(map[base.FileNum]*FileMetadata)
+			b.Deleted[df.Level] = dmap
+		}
+
+		if m == nil {
+			// m is nil only when replaying a MANIFEST.
+			if b.AddedByFileNum == nil {
+				return errors.Errorf("deleted file L%d.%s's metadata is absent and bve.AddedByFileNum is nil", df.Level, df.FileNum)
+			}
+			m = b.AddedByFileNum[df.FileNum]
+			if m == nil {
+				return base.CorruptionErrorf("pebble: file deleted L%d.%s before it was inserted", df.Level, df.FileNum)
+			}
+		}
+		if m.MarkedForCompaction {
+			b.MarkedForCompactionCountDiff--
+		}
+		dmap[df.FileNum] = m
+	}
+
+	for _, nf := range ve.NewFiles {
+		// A new file should not have been deleted in this or a preceding
+		// VersionEdit at the same level (though files can move across levels).
+		if dmap := b.Deleted[nf.Level]; dmap != nil {
+			if _, ok := dmap[nf.Meta.FileNum]; ok {
+				return base.CorruptionErrorf("pebble: file deleted L%d.%s before it was inserted", nf.Level, nf.Meta.FileNum)
+			}
+		}
+		b.Added[nf.Level] = append(b.Added[nf.Level], nf.Meta)
+		if b.AddedByFileNum != nil {
+			b.AddedByFileNum[nf.Meta.FileNum] = nf.Meta
+		}
+		if nf.Meta.MarkedForCompaction {
+			b.MarkedForCompactionCountDiff++
+		}
+	}
+	return nil
+}
+
+// Apply applies the delta b to the current version to produce a new
+// version. The new version is consistent with respect to the comparer cmp.
+//
+// curr may be nil, which is equivalent to a pointer to a zero version.
+//
+// On success, a map of zombie files containing the file numbers and sizes of
+// deleted files is returned. These files are considered zombies because they
+// are no longer referenced by the returned Version, but cannot be deleted from
+// disk as they are still in use by the incoming Version.
+func (b *BulkVersionEdit) Apply(
+	curr *Version,
+	cmp Compare,
+	formatKey base.FormatKey,
+	flushSplitBytes int64,
+	readCompactionRate int64,
+) (_ *Version, zombies map[base.FileNum]uint64, _ error) {
+	addZombie := func(fileNum base.FileNum, size uint64) {
+		if zombies == nil {
+			zombies = make(map[base.FileNum]uint64)
+		}
+		zombies[fileNum] = size
+	}
+	// The remove zombie function is used to handle tables that are moved from
+	// one level to another during a version edit (i.e. a "move" compaction).
+	removeZombie := func(fileNum base.FileNum) {
+		if zombies != nil {
+			delete(zombies, fileNum)
+		}
+	}
+
+	v := new(Version)
+
+	// Adjust the count of files marked for compaction.
+	if curr != nil {
+		v.Stats.MarkedForCompaction = curr.Stats.MarkedForCompaction
+	}
+	v.Stats.MarkedForCompaction += b.MarkedForCompactionCountDiff
+	if v.Stats.MarkedForCompaction < 0 {
+		return nil, nil, base.CorruptionErrorf("pebble: version marked for compaction count negative")
+	}
+
+	for level := range v.Levels {
+		if curr == nil || curr.Levels[level].tree.root == nil {
+			v.Levels[level] = makeLevelMetadata(cmp, level, nil /* files */)
+		} else {
+			v.Levels[level] = curr.Levels[level].clone()
+		}
+		if curr == nil || curr.RangeKeyLevels[level].tree.root == nil {
+			v.RangeKeyLevels[level] = makeLevelMetadata(cmp, level, nil /* files */)
+		} else {
+			v.RangeKeyLevels[level] = curr.RangeKeyLevels[level].clone()
+		}
+
+		if len(b.Added[level]) == 0 && len(b.Deleted[level]) == 0 {
+			// There are no edits on this level.
+			if level == 0 {
+				// Initialize L0Sublevels.
+				if curr == nil || curr.L0Sublevels == nil {
+					if err := v.InitL0Sublevels(cmp, formatKey, flushSplitBytes); err != nil {
+						return nil, nil, errors.Wrap(err, "pebble: internal error")
+					}
+				} else {
+					v.L0Sublevels = curr.L0Sublevels
+					v.L0SublevelFiles = v.L0Sublevels.Levels
+				}
+			}
+			continue
+		}
+
+		// Some edits on this level.
+		lm := &v.Levels[level]
+		lmRange := &v.RangeKeyLevels[level]
+		addedFiles := b.Added[level]
+		deletedMap := b.Deleted[level]
+		if n := v.Levels[level].Len() + len(addedFiles); n == 0 {
+			return nil, nil, base.CorruptionErrorf(
+				"pebble: internal error: No current or added files but have deleted files: %d",
+				errors.Safe(len(deletedMap)))
+		}
+
+		// NB: addedFiles may be empty and it also is not necessarily
+		// internally consistent: it does not reflect deletions in deletedMap.
+
+		for _, f := range deletedMap {
+			addZombie(f.FileNum, f.Size)
+			if obsolete := v.Levels[level].tree.Delete(f); obsolete {
+				// Deleting a file from the B-Tree may decrement its
+				// reference count. However, because we cloned the
+				// previous level's B-Tree, this should never result in a
+				// file's reference count dropping to zero.
+				err := errors.Errorf("pebble: internal error: file L%d.%s obsolete during B-Tree removal", level, f.FileNum)
+				return nil, nil, err
+			}
+			if f.HasRangeKeys {
+				if obsolete := v.RangeKeyLevels[level].tree.Delete(f); obsolete {
+					// Deleting a file from the B-Tree may decrement its
+					// reference count. However, because we cloned the
+					// previous level's B-Tree, this should never result in a
+					// file's reference count dropping to zero.
+					err := errors.Errorf("pebble: internal error: file L%d.%s obsolete during range-key B-Tree removal", level, f.FileNum)
+					return nil, nil, err
+				}
+			}
+		}
+
+		var sm, la *FileMetadata
+		for _, f := range addedFiles {
+			if _, ok := deletedMap[f.FileNum]; ok {
+				// Already called addZombie on this file in the preceding
+				// loop, so we don't need to do it here.
+				continue
+			}
+
+			// NB: allowedSeeks is used for read triggered compactions. It is set using
+			// Options.Experimental.ReadCompactionRate which defaults to 32KB.
+			var allowedSeeks int64
+			if readCompactionRate != 0 {
+				allowedSeeks = int64(f.Size) / readCompactionRate
+			}
+			if allowedSeeks < 100 {
+				allowedSeeks = 100
+			}
+			atomic.StoreInt64(&f.Atomic.AllowedSeeks, allowedSeeks)
+			f.InitAllowedSeeks = allowedSeeks
+
+			err := lm.tree.Insert(f)
+			if err != nil {
+				return nil, nil, errors.Wrap(err, "pebble")
+			}
+			if f.HasRangeKeys {
+				err = lmRange.tree.Insert(f)
+				if err != nil {
+					return nil, nil, errors.Wrap(err, "pebble")
+				}
+			}
+			removeZombie(f.FileNum)
+			// Track the keys with the smallest and largest keys, so that we can
+			// check consistency of the modified span.
+			if sm == nil || base.InternalCompare(cmp, sm.Smallest, f.Smallest) > 0 {
+				sm = f
+			}
+			if la == nil || base.InternalCompare(cmp, la.Largest, f.Largest) < 0 {
+				la = f
+			}
+		}
+
+		if level == 0 {
+			if curr != nil && curr.L0Sublevels != nil && len(deletedMap) == 0 {
+				// Flushes and ingestions that do not delete any L0 files do not require
+				// a regeneration of L0Sublevels from scratch. We can instead generate
+				// it incrementally.
+				var err error
+				// AddL0Files requires addedFiles to be sorted in seqnum order.
+				addedFiles = append([]*FileMetadata(nil), addedFiles...)
+				SortBySeqNum(addedFiles)
+				v.L0Sublevels, err = curr.L0Sublevels.AddL0Files(addedFiles, flushSplitBytes, &v.Levels[0])
+				if errors.Is(err, errInvalidL0SublevelsOpt) {
+					err = v.InitL0Sublevels(cmp, formatKey, flushSplitBytes)
+				}
+				if err != nil {
+					return nil, nil, errors.Wrap(err, "pebble: internal error")
+				}
+				v.L0SublevelFiles = v.L0Sublevels.Levels
+			} else if err := v.InitL0Sublevels(cmp, formatKey, flushSplitBytes); err != nil {
+				return nil, nil, errors.Wrap(err, "pebble: internal error")
+			}
+			if err := CheckOrdering(cmp, formatKey, Level(0), v.Levels[level].Iter()); err != nil {
+				return nil, nil, errors.Wrap(err, "pebble: internal error")
+			}
+			continue
+		}
+
+		// Check consistency of the level in the vicinity of our edits.
+		if sm != nil && la != nil {
+			overlap := overlaps(v.Levels[level].Iter(), cmp, sm.Smallest.UserKey,
+				la.Largest.UserKey, la.Largest.IsExclusiveSentinel())
+			// overlap contains all of the added files. We want to ensure that
+			// the added files are consistent with neighboring existing files
+			// too, so reslice the overlap to pull in a neighbor on each side.
+			check := overlap.Reslice(func(start, end *LevelIterator) {
+				if m := start.Prev(); m == nil {
+					start.Next()
+				}
+				if m := end.Next(); m == nil {
+					end.Prev()
+				}
+			})
+			if err := CheckOrdering(cmp, formatKey, Level(level), check.Iter()); err != nil {
+				return nil, nil, errors.Wrap(err, "pebble: internal error")
+			}
+		}
+	}
+	return v, zombies, nil
+}