github.com/cockroachdb/pebble@v1.1.2/internal/keyspan/merging_iter.go

// Copyright 2022 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 keyspan

import (
	"bytes"
	"fmt"
	"sort"

	"github.com/cockroachdb/pebble/internal/base"
	"github.com/cockroachdb/pebble/internal/invariants"
	"github.com/cockroachdb/pebble/internal/manifest"
)

// TODO(jackson): Consider implementing an optimization to seek lower levels
// past higher levels' RANGEKEYDELs. This would be analaogous to the
// optimization pebble.mergingIter performs for RANGEDELs during point key
// seeks. It may not be worth it, because range keys are rare and cascading
// seeks would require introducing key comparisons to switchTo{Min,Max}Heap
// where there currently are none.

// TODO(jackson): There are several opportunities to use base.Equal in the
// MergingIter implementation, but will require a bit of plumbing to thread the
// Equal function.

// MergingIter merges spans across levels of the LSM, exposing an iterator over
// spans that yields sets of spans fragmented at unique user key boundaries.
//
// A MergingIter is initialized with an arbitrary number of child iterators over
// fragmented spans. Each child iterator exposes fragmented key spans, such that
// overlapping keys are surfaced in a single Span. Key spans from one child
// iterator may overlap key spans from another child iterator arbitrarily.
//
// The spans combined by MergingIter will return spans with keys sorted by
// trailer descending. If the MergingIter is configured with a Transformer, it's
// permitted to modify the ordering of the spans' keys returned by MergingIter.
//
// # Algorithm
//
// The merging iterator wraps child iterators, merging and fragmenting spans
// across levels. The high-level algorithm is:
//
//  1. Initialize the heap with bound keys from child iterators' spans.
//  2. Find the next [or previous] two unique user keys' from bounds.
//  3. Consider the span formed between the two unique user keys a candidate
//     span.
//  4. Determine if any of the child iterators' spans overlap the candidate
//     span.
//     4a. If any of the child iterator's current bounds are end keys
//     (during forward iteration) or start keys (during reverse
//     iteration), then all the spans with that bound overlap the
//     candidate span.
//     4b. Apply the configured transform, which may remove keys.
//     4c. If no spans overlap, forget the smallest (forward iteration)
//     or largest (reverse iteration) unique user key and advance
//     the iterators to the next unique user key. Start again from 3.
//
// # Detailed algorithm
//
// Each level (i0, i1, ...) has a user-provided input FragmentIterator. The
// merging iterator steps through individual boundaries of the underlying
// spans separately. If the underlying FragmentIterator has fragments
// [a,b){#2,#1} [b,c){#1} the mergingIterLevel.{next,prev} step through:
//
//	(a, start), (b, end), (b, start), (c, end)
//
// Note that (a, start) and (b, end) are observed ONCE each, despite two keys
// sharing those bounds. Also note that (b, end) and (b, start) are two distinct
// iterator positions of a mergingIterLevel.
//
// The merging iterator maintains a heap (min during forward iteration, max
// during reverse iteration) containing the boundKeys. Each boundKey is a
// 3-tuple holding the bound user key, whether the bound is a start or end key
// and the set of keys from that level that have that bound. The heap orders
// based on the boundKey's user key only.
//
// The merging iterator is responsible for merging spans across levels to
// determine which span is next, but it's also responsible for fragmenting
// overlapping spans. Consider the example:
//
//	       i0:     b---d e-----h
//	       i1:   a---c         h-----k
//	       i2:   a------------------------------p
//
//	fragments:   a-b-c-d-e-----h-----k----------p
//
// None of the individual child iterators contain a span with the exact bounds
// [c,d), but the merging iterator must produce a span [c,d). To accomplish
// this, the merging iterator visits every span between unique boundary user
// keys. In the above example, this is:
//
//	[a,b), [b,c), [c,d), [d,e), [e, h), [h, k), [k, p)
//
// The merging iterator first initializes the heap to prepare for iteration.
// The description below discusses the mechanics of forward iteration after a
// call to First, but the mechanics are similar for reverse iteration and
// other positioning methods.
//
// During a call to First, the heap is initialized by seeking every
// mergingIterLevel to the first bound of the first fragment. In the above
// example, this seeks the child iterators to:
//
//	i0: (b, boundKindFragmentStart, [ [b,d) ])
//	i1: (a, boundKindFragmentStart, [ [a,c) ])
//	i2: (a, boundKindFragmentStart, [ [a,p) ])
//
// After fixing up the heap, the root of the heap is a boundKey with the
// smallest user key ('a' in the example). Once the heap is setup for iteration
// in the appropriate direction and location, the merging iterator uses
// find{Next,Prev}FragmentSet to find the next/previous span bounds.
//
// During forward iteration, the root of the heap's user key is the start key
// key of next merged span. findNextFragmentSet sets m.start to this user
// key. The heap may contain other boundKeys with the same user key if another
// level has a fragment starting or ending at the same key, so the
// findNextFragmentSet method pulls from the heap until it finds the first key
// greater than m.start. This key is used as the end key.
//
// In the above example, this results in m.start = 'a', m.end = 'b' and child
// iterators in the following positions:
//
//	i0: (b, boundKindFragmentStart, [ [b,d) ])
//	i1: (c, boundKindFragmentEnd,   [ [a,c) ])
//	i2: (p, boundKindFragmentEnd,   [ [a,p) ])
//
// With the user key bounds of the next merged span established,
// findNextFragmentSet must determine which, if any, fragments overlap the span.
// During forward iteration any child iterator that is now positioned at an end
// boundary has an overlapping span. (Justification: The child iterator's end
// boundary is ≥ m.end. The corresponding start boundary must be ≤ m.start since
// there were no other user keys between m.start and m.end. So the fragments
// associated with the iterator's current end boundary have start and end bounds
// such that start ≤ m.start < m.end ≤ end).
//
// findNextFragmentSet iterates over the levels, collecting keys from any child
// iterators positioned at end boundaries. In the above example, i1 and i2 are
// positioned at end boundaries, so findNextFragmentSet collects the keys of
// [a,c) and [a,p). These spans contain the merging iterator's [m.start, m.end)
// span, but they may also extend beyond the m.start and m.end. The merging
// iterator returns the keys with the merging iter's m.start and m.end bounds,
// preserving the underlying keys' sequence numbers, key kinds and values.
//
// A MergingIter is configured with a Transform that's applied to the span
// before surfacing it to the iterator user. A Transform may remove keys
// arbitrarily, but it may not modify the values themselves.
//
// It may be the case that findNextFragmentSet finds no levels positioned at end
// boundaries, or that there are no spans remaining after applying a transform,
// in which case the span [m.start, m.end) overlaps with nothing. In this case
// findNextFragmentSet loops, repeating the above process again until it finds a
// span that does contain keys.
//
// # Memory safety
//
// The FragmentIterator interface only guarantees stability of a Span and its
// associated slices until the next positioning method is called. Adjacent Spans
// may be contained in different sstables, requring the FragmentIterator
// implementation to close one sstable, releasing its memory, before opening the
// next. Most of the state used by the MergingIter is derived from spans at
// current child iterator positions only, ensuring state is stable. The one
// exception is the start bound during forward iteration and the end bound
// during reverse iteration.
//
// If the heap root originates from an end boundary when findNextFragmentSet
// begins, a Next on the heap root level may invalidate the end boundary. To
// accommodate this, find{Next,Prev}FragmentSet copy the initial boundary if the
// subsequent Next/Prev would move to the next span.
type MergingIter struct {
	*MergingBuffers
	// start and end hold the bounds for the span currently under the
	// iterator position.
	//
	// Invariant: None of the levels' iterators contain spans with a bound
	// between start and end. For all bounds b, b ≤ start || b ≥ end.
	start, end []byte

	// transformer defines a transformation to be applied to a span before it's
	// yielded to the user. Transforming may filter individual keys contained
	// within the span.
	transformer Transformer
	// span holds the iterator's current span. This span is used as the
	// destination for transforms. Every tranformed span overwrites the
	// previous.
	span Span
	err  error
	dir  int8

	// alloc preallocates mergingIterLevel and mergingIterItems for use by the
	// merging iterator. As long as the merging iterator is used with
	// manifest.NumLevels+3 and fewer fragment iterators, the merging iterator
	// will not need to allocate upon initialization. The value NumLevels+3
	// mirrors the preallocated levels in iterAlloc used for point iterators.
	// Invariant: cap(levels) == cap(items)
	alloc struct {
		levels [manifest.NumLevels + 3]mergingIterLevel
		items  [manifest.NumLevels + 3]mergingIterItem
	}
}

// MergingBuffers holds buffers used while merging keyspans.
type MergingBuffers struct {
	// keys holds all of the keys across all levels that overlap the key span
	// [start, end), sorted by Trailer descending. This slice is reconstituted
	// in synthesizeKeys from each mergingIterLevel's keys every time the
	// [start, end) bounds change.
	//
	// Each element points into a child iterator's memory, so the keys may not
	// be directly modified.
	keys keysBySeqNumKind
	// levels holds levels allocated by MergingIter.init. The MergingIter will
	// prefer use of its `manifest.NumLevels+3` array, so this slice will be
	// longer if set.
	levels []mergingIterLevel
	// heap holds a slice for the merging iterator heap allocated by
	// MergingIter.init. The MergingIter will prefer use of its
	// `manifest.NumLevels+3` items array, so this slice will be longer if set.
	heap mergingIterHeap
	// buf is a buffer used to save [start, end) boundary keys.
	buf []byte
}

// PrepareForReuse discards any excessively large buffers.
func (bufs *MergingBuffers) PrepareForReuse() {
	if cap(bufs.buf) > bufferReuseMaxCapacity {
		bufs.buf = nil
	}
}

// MergingIter implements the FragmentIterator interface.
var _ FragmentIterator = (*MergingIter)(nil)

type mergingIterLevel struct {
	iter FragmentIterator

	// heapKey holds the current key at this level for use within the heap.
	heapKey boundKey
}

func (l *mergingIterLevel) next() {
	if l.heapKey.kind == boundKindFragmentStart {
		l.heapKey = boundKey{
			kind: boundKindFragmentEnd,
			key:  l.heapKey.span.End,
			span: l.heapKey.span,
		}
		return
	}
	if s := l.iter.Next(); s == nil {
		l.heapKey = boundKey{kind: boundKindInvalid}
	} else {
		l.heapKey = boundKey{
			kind: boundKindFragmentStart,
			key:  s.Start,
			span: s,
		}
	}
}

func (l *mergingIterLevel) prev() {
	if l.heapKey.kind == boundKindFragmentEnd {
		l.heapKey = boundKey{
			kind: boundKindFragmentStart,
			key:  l.heapKey.span.Start,
			span: l.heapKey.span,
		}
		return
	}
	if s := l.iter.Prev(); s == nil {
		l.heapKey = boundKey{kind: boundKindInvalid}
	} else {
		l.heapKey = boundKey{
			kind: boundKindFragmentEnd,
			key:  s.End,
			span: s,
		}
	}
}

// Init initializes the merging iterator with the provided fragment iterators.
func (m *MergingIter) Init(
	cmp base.Compare, transformer Transformer, bufs *MergingBuffers, iters ...FragmentIterator,
) {
	*m = MergingIter{
		MergingBuffers: bufs,
		transformer:    transformer,
	}
	m.heap.cmp = cmp
	levels, items := m.levels, m.heap.items

	// Invariant: cap(levels) >= cap(items)
	// Invariant: cap(alloc.levels) == cap(alloc.items)
	if len(iters) <= len(m.alloc.levels) {
		// The slices allocated on the MergingIter struct are large enough.
		m.levels = m.alloc.levels[:len(iters)]
		m.heap.items = m.alloc.items[:0]
	} else if len(iters) <= cap(levels) {
		// The existing heap-allocated slices are large enough, so reuse them.
		m.levels = levels[:len(iters)]
		m.heap.items = items[:0]
	} else {
		// Heap allocate new slices.
		m.levels = make([]mergingIterLevel, len(iters))
		m.heap.items = make([]mergingIterItem, 0, len(iters))
	}
	for i := range m.levels {
		m.levels[i] = mergingIterLevel{iter: iters[i]}
	}
}

// AddLevel adds a new level to the bottom of the merging iterator. AddLevel
// must be called after Init and before any other method.
func (m *MergingIter) AddLevel(iter FragmentIterator) {
	m.levels = append(m.levels, mergingIterLevel{iter: iter})
}

// SeekGE moves the iterator to the first span covering a key greater than
// or equal to the given key. This is equivalent to seeking to the first
// span with an end key greater than the given key.
func (m *MergingIter) SeekGE(key []byte) *Span {
	m.invalidate() // clear state about current position

	// SeekGE(k) seeks to the first span with an end key greater than the given
	// key. The merged span M that we're searching for might straddle the seek
	// `key`. In this case, the M.Start may be a key ≤ the seek key.
	//
	// Consider a SeekGE(dog) in the following example.
	//
	//            i0:     b---d e-----h
	//            i1:   a---c         h-----k
	//            i2:   a------------------------------p
	//        merged:   a-b-c-d-e-----h-----k----------p
	//
	// The merged span M containing 'dog' is [d,e). The 'd' of the merged span
	// comes from i0's [b,d)'s end boundary. The [b,d) span does not cover any
	// key >= dog, so we cannot find the span by positioning the child iterators
	// using a SeekGE(dog).
	//
	// Instead, if we take all the child iterators' spans bounds:
	//                  a b c d e     h     k          p
	// We want to partition them into keys ≤ `key` and keys > `key`.
	//                        dog
	//                         │
	//                  a b c d│e     h     k          p
	//                         │
	// The largest key on the left of the partition forms the merged span's
	// start key, and the smallest key on the right of the partition forms the
	// merged span's end key. Recharacterized:
	//
	//   M.Start: the largest boundary ≤ k of any child span
	//   M.End:   the smallest boundary > k of any child span
	//
	// The FragmentIterator interface doesn't implement seeking by all bounds,
	// it implements seeking by containment. A SeekGE(k) will ensure we observe
	// all start boundaries ≥ k and all end boundaries > k but does not ensure
	// we observe end boundaries = k or any boundaries < k.  A SeekLT(k) will
	// ensure we observe all start boundaries < k and all end boundaries ≤ k but
	// does not ensure we observe any start boundaries = k or any boundaries >
	// k. This forces us to seek in one direction and step in the other.
	//
	// In a SeekGE, we want to end up oriented in the forward direction when
	// complete, so we begin with searching for M.Start by SeekLT-ing every
	// child iterator to `k`.  For every child span found, we determine the
	// largest bound ≤ `k` and use it to initialize our max heap. The resulting
	// root of the max heap is a preliminary value for `M.Start`.
	for i := range m.levels {
		l := &m.levels[i]
		s := l.iter.SeekLT(key)
		if s == nil {
			l.heapKey = boundKey{kind: boundKindInvalid}
		} else if m.cmp(s.End, key) <= 0 {
			l.heapKey = boundKey{
				kind: boundKindFragmentEnd,
				key:  s.End,
				span: s,
			}
		} else {
			// s.End > key && s.Start < key
			// We need to use this span's start bound, since that's the largest
			// bound ≤ key.
			l.heapKey = boundKey{
				kind: boundKindFragmentStart,
				key:  s.Start,
				span: s,
			}
		}
	}
	m.initMaxHeap()
	if m.err != nil {
		return nil
	} else if len(m.heap.items) == 0 {
		// There are no spans covering any key < `key`. There is no span that
		// straddles the seek key. Reorient the heap into a min heap and return
		// the first span we find in the forward direction.
		m.switchToMinHeap()
		return m.findNextFragmentSet()
	}

	// The heap root is now the largest boundary key b such that:
	//   1. b < k
	//   2. b = k, and b is an end boundary
	// There's a third case that we will need to consider later, after we've
	// switched to a min heap:
	//   3. there exists a start boundary key b such that b = k.
	// A start boundary key equal to k would not be surfaced when we seeked all
	// the levels using SeekLT(k), since no key <k would be covered within a
	// span within an inclusive `k` start boundary.
	//
	// Assume that the tightest boundary ≤ k is the current heap root (cases 1 &
	// 2). After we switch to a min heap, we'll check for the third case and
	// adjust the start boundary if necessary.
	m.start = m.heap.items[0].boundKey.key

	// Before switching the direction of the heap, save a copy of the start
	// boundary if it's the end boundary of some child span. Next-ing the child
	// iterator might switch files and invalidate the memory of the bound.
	if m.heap.items[0].boundKey.kind == boundKindFragmentEnd {
		m.buf = append(m.buf[:0], m.start...)
		m.start = m.buf
	}

	// Switch to a min heap. This will move each level to the next bound in
	// every level, and then establish a min heap. This allows us to obtain the
	// smallest boundary key > `key`, which will serve as our candidate end
	// bound.
	m.switchToMinHeap()
	if m.err != nil {
		return nil
	} else if len(m.heap.items) == 0 {
		return nil
	}

	// Check for the case 3 described above. It's possible that when we switch
	// heap directions, we discover a start boundary of some child span that is
	// equal to the seek key `key`. In this case, we want this key to be our
	// start boundary.
	if m.heap.items[0].boundKey.kind == boundKindFragmentStart &&
		m.cmp(m.heap.items[0].boundKey.key, key) == 0 {
		// Call findNextFragmentSet, which will set m.start to the heap root and
		// proceed forward.
		return m.findNextFragmentSet()
	}

	m.end = m.heap.items[0].boundKey.key
	if found, s := m.synthesizeKeys(+1); found && s != nil {
		return s
	}
	return m.findNextFragmentSet()

}

// SeekLT moves the iterator to the last span covering a key less than the
// given key. This is equivalent to seeking to the last span with a start
// key less than the given key.
func (m *MergingIter) SeekLT(key []byte) *Span {
	m.invalidate() // clear state about current position

	// SeekLT(k) seeks to the last span with a start key less than the given
	// key. The merged span M that we're searching for might straddle the seek
	// `key`. In this case, the M.End may be a key ≥ the seek key.
	//
	// Consider a SeekLT(dog) in the following example.
	//
	//            i0:     b---d e-----h
	//            i1:   a---c         h-----k
	//            i2:   a------------------------------p
	//        merged:   a-b-c-d-e-----h-----k----------p
	//
	// The merged span M containing the largest key <'dog' is [d,e). The 'e' of
	// the merged span comes from i0's [e,h)'s start boundary. The [e,h) span
	// does not cover any key < dog, so we cannot find the span by positioning
	// the child iterators using a SeekLT(dog).
	//
	// Instead, if we take all the child iterators' spans bounds:
	//                  a b c d e     h     k          p
	// We want to partition them into keys < `key` and keys ≥ `key`.
	//                        dog
	//                         │
	//                  a b c d│e     h     k          p
	//                         │
	// The largest key on the left of the partition forms the merged span's
	// start key, and the smallest key on the right of the partition forms the
	// merged span's end key. Recharacterized:
	//
	//   M.Start: the largest boundary < k of any child span
	//   M.End:   the smallest boundary ≥ k of any child span
	//
	// The FragmentIterator interface doesn't implement seeking by all bounds,
	// it implements seeking by containment. A SeekGE(k) will ensure we observe
	// all start boundaries ≥ k and all end boundaries > k but does not ensure
	// we observe end boundaries = k or any boundaries < k.  A SeekLT(k) will
	// ensure we observe all start boundaries < k and all end boundaries ≤ k but
	// does not ensure we observe any start boundaries = k or any boundaries >
	// k. This forces us to seek in one direction and step in the other.
	//
	// In a SeekLT, we want to end up oriented in the backward direction when
	// complete, so we begin with searching for M.End by SeekGE-ing every
	// child iterator to `k`. For every child span found, we determine the
	// smallest bound ≥ `k` and use it to initialize our min heap. The resulting
	// root of the min heap is a preliminary value for `M.End`.
	for i := range m.levels {
		l := &m.levels[i]
		s := l.iter.SeekGE(key)
		if s == nil {
			l.heapKey = boundKey{kind: boundKindInvalid}
		} else if m.cmp(s.Start, key) >= 0 {
			l.heapKey = boundKey{
				kind: boundKindFragmentStart,
				key:  s.Start,
				span: s,
			}
		} else {
			// s.Start < key
			// We need to use this span's end bound, since that's the smallest
			// bound > key.
			l.heapKey = boundKey{
				kind: boundKindFragmentEnd,
				key:  s.End,
				span: s,
			}
		}
	}
	m.initMinHeap()
	if m.err != nil {
		return nil
	} else if len(m.heap.items) == 0 {
		// There are no spans covering any key ≥ `key`. There is no span that
		// straddles the seek key. Reorient the heap into a max heap and return
		// the first span we find in the reverse direction.
		m.switchToMaxHeap()
		return m.findPrevFragmentSet()
	}

	// The heap root is now the smallest boundary key b such that:
	//   1. b > k
	//   2. b = k, and b is a start boundary
	// There's a third case that we will need to consider later, after we've
	// switched to a max heap:
	//   3. there exists an end boundary key b such that b = k.
	// An end boundary key equal to k would not be surfaced when we seeked all
	// the levels using SeekGE(k), since k would not be contained within the
	// exclusive end boundary.
	//
	// Assume that the tightest boundary ≥ k is the current heap root (cases 1 &
	// 2). After we switch to a max heap, we'll check for the third case and
	// adjust the end boundary if necessary.
	m.end = m.heap.items[0].boundKey.key

	// Before switching the direction of the heap, save a copy of the end
	// boundary if it's the start boundary of some child span. Prev-ing the
	// child iterator might switch files and invalidate the memory of the bound.
	if m.heap.items[0].boundKey.kind == boundKindFragmentStart {
		m.buf = append(m.buf[:0], m.end...)
		m.end = m.buf
	}

	// Switch to a max heap. This will move each level to the previous bound in
	// every level, and then establish a max heap. This allows us to obtain the
	// largest boundary key < `key`, which will serve as our candidate start
	// bound.
	m.switchToMaxHeap()
	if m.err != nil {
		return nil
	} else if len(m.heap.items) == 0 {
		return nil
	}
	// Check for the case 3 described above. It's possible that when we switch
	// heap directions, we discover an end boundary of some child span that is
	// equal to the seek key `key`. In this case, we want this key to be our end
	// boundary.
	if m.heap.items[0].boundKey.kind == boundKindFragmentEnd &&
		m.cmp(m.heap.items[0].boundKey.key, key) == 0 {
		// Call findPrevFragmentSet, which will set m.end to the heap root and
		// proceed backwards.
		return m.findPrevFragmentSet()
	}

	m.start = m.heap.items[0].boundKey.key
	if found, s := m.synthesizeKeys(-1); found && s != nil {
		return s
	}
	return m.findPrevFragmentSet()
}

// First seeks the iterator to the first span.
func (m *MergingIter) First() *Span {
	m.invalidate() // clear state about current position
	for i := range m.levels {
		if s := m.levels[i].iter.First(); s == nil {
			m.levels[i].heapKey = boundKey{kind: boundKindInvalid}
		} else {
			m.levels[i].heapKey = boundKey{
				kind: boundKindFragmentStart,
				key:  s.Start,
				span: s,
			}
		}
	}
	m.initMinHeap()
	return m.findNextFragmentSet()
}

// Last seeks the iterator to the last span.
func (m *MergingIter) Last() *Span {
	m.invalidate() // clear state about current position
	for i := range m.levels {
		if s := m.levels[i].iter.Last(); s == nil {
			m.levels[i].heapKey = boundKey{kind: boundKindInvalid}
		} else {
			m.levels[i].heapKey = boundKey{
				kind: boundKindFragmentEnd,
				key:  s.End,
				span: s,
			}
		}
	}
	m.initMaxHeap()
	return m.findPrevFragmentSet()
}

// Next advances the iterator to the next span.
func (m *MergingIter) Next() *Span {
	if m.err != nil {
		return nil
	}
	if m.dir == +1 && (m.end == nil || m.start == nil) {
		return nil
	}
	if m.dir != +1 {
		m.switchToMinHeap()
	}
	return m.findNextFragmentSet()
}

// Prev advances the iterator to the previous span.
func (m *MergingIter) Prev() *Span {
	if m.err != nil {
		return nil
	}
	if m.dir == -1 && (m.end == nil || m.start == nil) {
		return nil
	}
	if m.dir != -1 {
		m.switchToMaxHeap()
	}
	return m.findPrevFragmentSet()
}

// Error returns any accumulated error.
func (m *MergingIter) Error() error {
	if m.heap.len() == 0 || m.err != nil {
		return m.err
	}
	return m.levels[m.heap.items[0].index].iter.Error()
}

// Close closes the iterator, releasing all acquired resources.
func (m *MergingIter) Close() error {
	for i := range m.levels {
		if err := m.levels[i].iter.Close(); err != nil && m.err == nil {
			m.err = err
		}
	}
	m.levels = nil
	m.heap.items = m.heap.items[:0]
	return m.err
}

// String implements fmt.Stringer.
func (m *MergingIter) String() string {
	return "merging-keyspan"
}

func (m *MergingIter) initMinHeap() {
	m.dir = +1
	m.heap.reverse = false
	m.initHeap()
}

func (m *MergingIter) initMaxHeap() {
	m.dir = -1
	m.heap.reverse = true
	m.initHeap()
}

func (m *MergingIter) initHeap() {
	m.heap.items = m.heap.items[:0]
	for i := range m.levels {
		if l := &m.levels[i]; l.heapKey.kind != boundKindInvalid {
			m.heap.items = append(m.heap.items, mergingIterItem{
				index:    i,
				boundKey: &l.heapKey,
			})
		} else {
			m.err = firstError(m.err, l.iter.Error())
			if m.err != nil {
				return
			}
		}
	}
	m.heap.init()
}

func (m *MergingIter) switchToMinHeap() {
	// switchToMinHeap reorients the heap for forward iteration, without moving
	// the current MergingIter position.

	// The iterator is currently positioned at the span [m.start, m.end),
	// oriented in the reverse direction, so each level's iterator is positioned
	// to the largest key ≤ m.start. To reorient in the forward direction, we
	// must advance each level's iterator to the smallest key ≥ m.end. Consider
	// this three-level example.
	//
	//         i0:     b---d e-----h
	//         i1:   a---c         h-----k
	//         i2:   a------------------------------p
	//
	//     merged:   a-b-c-d-e-----h-----k----------p
	//
	// If currently positioned at the merged span [c,d), then the level
	// iterators' heap keys are:
	//
	//    i0: (b, [b, d))   i1: (c, [a,c))   i2: (a, [a,p))
	//
	// Reversing the heap should not move the merging iterator and should not
	// change the current [m.start, m.end) bounds. It should only prepare for
	// forward iteration by updating the child iterators' heap keys to:
	//
	//    i0: (d, [b, d))   i1: (h, [h,k))   i2: (p, [a,p))
	//
	// In every level the first key ≥ m.end is the next in the iterator.
	// Justification: Suppose not and a level iterator's next key was some key k
	// such that k < m.end. The max-heap invariant dictates that the current
	// iterator position is the largest entry with a user key ≥ m.start. This
	// means k > m.start. We started with the assumption that k < m.end, so
	// m.start < k < m.end. But then k is between our current span bounds,
	// and reverse iteration would have constructed the current interval to be
	// [k, m.end) not [m.start, m.end).

	if invariants.Enabled {
		for i := range m.levels {
			l := &m.levels[i]
			if l.heapKey.kind != boundKindInvalid && m.cmp(l.heapKey.key, m.start) > 0 {
				panic("pebble: invariant violation: max-heap key > m.start")
			}
		}
	}

	for i := range m.levels {
		m.levels[i].next()
	}
	m.initMinHeap()
}

func (m *MergingIter) switchToMaxHeap() {
	// switchToMaxHeap reorients the heap for reverse iteration, without moving
	// the current MergingIter position.

	// The iterator is currently positioned at the span [m.start, m.end),
	// oriented in the forward direction. Each level's iterator is positioned at
	// the smallest bound ≥ m.end. To reorient in the reverse direction, we must
	// move each level's iterator to the largest key ≤ m.start. Consider this
	// three-level example.
	//
	//         i0:     b---d e-----h
	//         i1:   a---c         h-----k
	//         i2:   a------------------------------p
	//
	//     merged:   a-b-c-d-e-----h-----k----------p
	//
	// If currently positioned at the merged span [c,d), then the level
	// iterators' heap keys are:
	//
	//    i0: (d, [b, d))   i1: (h, [h,k))   i2: (p, [a,p))
	//
	// Reversing the heap should not move the merging iterator and should not
	// change the current [m.start, m.end) bounds. It should only prepare for
	// reverse iteration by updating the child iterators' heap keys to:
	//
	//    i0: (b, [b, d))   i1: (c, [a,c))   i2: (a, [a,p))
	//
	// In every level the largest key ≤ m.start is the prev in the iterator.
	// Justification: Suppose not and a level iterator's prev key was some key k
	// such that k > m.start. The min-heap invariant dictates that the current
	// iterator position is the smallest entry with a user key ≥ m.end. This
	// means k < m.end, otherwise the iterator would be positioned at k. We
	// started with the assumption that k > m.start, so m.start < k < m.end. But
	// then k is between our current span bounds, and reverse iteration
	// would have constructed the current interval to be [m.start, k) not
	// [m.start, m.end).

	if invariants.Enabled {
		for i := range m.levels {
			l := &m.levels[i]
			if l.heapKey.kind != boundKindInvalid && m.cmp(l.heapKey.key, m.end) < 0 {
				panic("pebble: invariant violation: min-heap key < m.end")
			}
		}
	}

	for i := range m.levels {
		m.levels[i].prev()
	}
	m.initMaxHeap()
}

func (m *MergingIter) cmp(a, b []byte) int {
	return m.heap.cmp(a, b)
}

func (m *MergingIter) findNextFragmentSet() *Span {
	// Each iteration of this loop considers a new merged span between unique
	// user keys. An iteration may find that there exists no overlap for a given
	// span, (eg, if the spans [a,b), [d, e) exist within level iterators, the
	// below loop will still consider [b,d) before continuing to [d, e)). It
	// returns when it finds a span that is covered by at least one key.

	for m.heap.len() > 0 && m.err == nil {
		// Initialize the next span's start bound. SeekGE and First prepare the
		// heap without advancing. Next leaves the heap in a state such that the
		// root is the smallest bound key equal to the returned span's end key,
		// so the heap is already positioned at the next merged span's start key.

		// NB: m.heapRoot() might be either an end boundary OR a start boundary
		// of a level's span. Both end and start boundaries may still be a start
		// key of a span in the set of fragmented spans returned by MergingIter.
		// Consider the scenario:
		//       a----------l      #1
		//         b-----------m   #2
		//
		// The merged, fully-fragmented spans that MergingIter exposes to the caller
		// have bounds:
		//        a-b              #1
		//          b--------l     #1
		//          b--------l     #2
		//                   l-m   #2
		//
		// When advancing to l-m#2, we must set m.start to 'l', which originated
		// from [a,l)#1's end boundary.
		m.start = m.heap.items[0].boundKey.key

		// Before calling nextEntry, consider whether it might invalidate our
		// start boundary. If the start boundary key originated from an end
		// boundary, then we need to copy the start key before advancing the
		// underlying iterator to the next Span.
		if m.heap.items[0].boundKey.kind == boundKindFragmentEnd {
			m.buf = append(m.buf[:0], m.start...)
			m.start = m.buf
		}

		// There may be many entries all with the same user key. Spans in other
		// levels may also start or end at this same user key. For eg:
		// L1:   [a, c) [c, d)
		// L2:          [c, e)
		// If we're positioned at L1's end(c) end boundary, we want to advance
		// to the first bound > c.
		m.nextEntry()
		for len(m.heap.items) > 0 && m.err == nil && m.cmp(m.heapRoot(), m.start) == 0 {
			m.nextEntry()
		}
		if len(m.heap.items) == 0 || m.err != nil {
			break
		}

		// The current entry at the top of the heap is the first key > m.start.
		// It must become the end bound for the span we will return to the user.
		// In the above example, the root of the heap is L1's end(d).
		m.end = m.heap.items[0].boundKey.key

		// Each level within m.levels may have a span that overlaps the
		// fragmented key span [m.start, m.end). Update m.keys to point to them
		// and sort them by kind, sequence number. There may not be any keys
		// defined over [m.start, m.end) if we're between the end of one span
		// and the start of the next, OR if the configured transform filters any
		// keys out. We allow empty spans that were emitted by child iterators, but
		// we elide empty spans created by the mergingIter itself that don't overlap
		// with any child iterator returned spans (i.e. empty spans that bridge two
		// distinct child-iterator-defined spans).
		if found, s := m.synthesizeKeys(+1); found && s != nil {
			return s
		}
	}
	// Exhausted.
	m.clear()
	return nil
}

func (m *MergingIter) findPrevFragmentSet() *Span {
	// Each iteration of this loop considers a new merged span between unique
	// user keys. An iteration may find that there exists no overlap for a given
	// span, (eg, if the spans [a,b), [d, e) exist within level iterators, the
	// below loop will still consider [b,d) before continuing to [a, b)). It
	// returns when it finds a span that is covered by at least one key.

	for m.heap.len() > 0 && m.err == nil {
		// Initialize the next span's end bound. SeekLT and Last prepare the
		// heap without advancing. Prev leaves the heap in a state such that the
		// root is the largest bound key equal to the returned span's start key,
		// so the heap is already positioned at the next merged span's end key.

		// NB: m.heapRoot() might be either an end boundary OR a start boundary
		// of a level's span. Both end and start boundaries may still be a start
		// key of a span returned by MergingIter. Consider the scenario:
		//       a----------l      #2
		//         b-----------m   #1
		//
		// The merged, fully-fragmented spans that MergingIter exposes to the caller
		// have bounds:
		//        a-b              #2
		//          b--------l     #2
		//          b--------l     #1
		//                   l-m   #1
		//
		// When Preving to a-b#2, we must set m.end to 'b', which originated
		// from [b,m)#1's start boundary.
		m.end = m.heap.items[0].boundKey.key

		// Before calling prevEntry, consider whether it might invalidate our
		// end boundary. If the end boundary key originated from a start
		// boundary, then we need to copy the end key before advancing the
		// underlying iterator to the previous Span.
		if m.heap.items[0].boundKey.kind == boundKindFragmentStart {
			m.buf = append(m.buf[:0], m.end...)
			m.end = m.buf
		}

		// There may be many entries all with the same user key. Spans in other
		// levels may also start or end at this same user key. For eg:
		// L1:   [a, c) [c, d)
		// L2:          [c, e)
		// If we're positioned at L1's start(c) start boundary, we want to prev
		// to move to the first bound < c.
		m.prevEntry()
		for len(m.heap.items) > 0 && m.err == nil && m.cmp(m.heapRoot(), m.end) == 0 {
			m.prevEntry()
		}
		if len(m.heap.items) == 0 || m.err != nil {
			break
		}

		// The current entry at the top of the heap is the first key < m.end.
		// It must become the start bound for the span we will return to the
		// user. In the above example, the root of the heap is L1's start(a).
		m.start = m.heap.items[0].boundKey.key

		// Each level within m.levels may have a set of keys that overlap the
		// fragmented key span [m.start, m.end). Update m.keys to point to them
		// and sort them by kind, sequence number. There may not be any keys
		// spanning [m.start, m.end) if we're between the end of one span and
		// the start of the next, OR if the configured transform filters any
		// keys out.  We allow empty spans that were emitted by child iterators, but
		// we elide empty spans created by the mergingIter itself that don't overlap
		// with any child iterator returned spans (i.e. empty spans that bridge two
		// distinct child-iterator-defined spans).
		if found, s := m.synthesizeKeys(-1); found && s != nil {
			return s
		}
	}
	// Exhausted.
	m.clear()
	return nil
}

func (m *MergingIter) heapRoot() []byte {
	return m.heap.items[0].boundKey.key
}

// synthesizeKeys is called by find{Next,Prev}FragmentSet to populate and
// sort the set of keys overlapping [m.start, m.end).
//
// During forward iteration, if the current heap item is a fragment end,
// then the fragment's start must be ≤ m.start and the fragment overlaps the
// current iterator position of [m.start, m.end).
//
// During reverse iteration, if the current heap item is a fragment start,
// then the fragment's end must be ≥ m.end and the fragment overlaps the
// current iteration position of [m.start, m.end).
//
// The boolean return value, `found`, is true if the returned span overlaps
// with a span returned by a child iterator.
func (m *MergingIter) synthesizeKeys(dir int8) (bool, *Span) {
	if invariants.Enabled {
		if m.cmp(m.start, m.end) >= 0 {
			panic(fmt.Sprintf("pebble: invariant violation: span start ≥ end: %s >= %s", m.start, m.end))
		}
	}

	m.keys = m.keys[:0]
	found := false
	for i := range m.levels {
		if dir == +1 && m.levels[i].heapKey.kind == boundKindFragmentEnd ||
			dir == -1 && m.levels[i].heapKey.kind == boundKindFragmentStart {
			m.keys = append(m.keys, m.levels[i].heapKey.span.Keys...)
			found = true
		}
	}
	// TODO(jackson): We should be able to remove this sort and instead
	// guarantee that we'll return keys in the order of the levels they're from.
	// With careful iterator construction, this would  guarantee that they're
	// sorted by trailer descending for the range key iteration use case.
	sort.Sort(&m.keys)

	// Apply the configured transform. See VisibleTransform.
	m.span = Span{
		Start:     m.start,
		End:       m.end,
		Keys:      m.keys,
		KeysOrder: ByTrailerDesc,
	}
	// NB: m.heap.cmp is a base.Compare, whereas m.cmp is a method on
	// MergingIter.
	if err := m.transformer.Transform(m.heap.cmp, m.span, &m.span); err != nil {
		m.err = err
		return false, nil
	}
	return found, &m.span
}

func (m *MergingIter) invalidate() {
	m.err = nil
}

func (m *MergingIter) clear() {
	for fi := range m.keys {
		m.keys[fi] = Key{}
	}
	m.keys = m.keys[:0]
}

// nextEntry steps to the next entry.
func (m *MergingIter) nextEntry() {
	l := &m.levels[m.heap.items[0].index]
	l.next()
	if !l.heapKey.valid() {
		// l.iter is exhausted.
		m.err = l.iter.Error()
		if m.err == nil {
			m.heap.pop()
		}
		return
	}

	if m.heap.len() > 1 {
		m.heap.fix(0)
	}
}

// prevEntry steps to the previous entry.
func (m *MergingIter) prevEntry() {
	l := &m.levels[m.heap.items[0].index]
	l.prev()
	if !l.heapKey.valid() {
		// l.iter is exhausted.
		m.err = l.iter.Error()
		if m.err == nil {
			m.heap.pop()
		}
		return
	}

	if m.heap.len() > 1 {
		m.heap.fix(0)
	}
}

// DebugString returns a string representing the current internal state of the
// merging iterator and its heap for debugging purposes.
func (m *MergingIter) DebugString() string {
	var buf bytes.Buffer
	fmt.Fprintf(&buf, "Current bounds: [%q, %q)\n", m.start, m.end)
	for i := range m.levels {
		fmt.Fprintf(&buf, "%d: heap key %s\n", i, m.levels[i].heapKey)
	}
	return buf.String()
}

type mergingIterItem struct {
	// boundKey points to the corresponding mergingIterLevel's `iterKey`.
	*boundKey
	// index is the index of this level within the MergingIter's levels field.
	index int
}

// mergingIterHeap is copied from mergingIterHeap defined in the root pebble
// package for use with point keys.

type mergingIterHeap struct {
	cmp     base.Compare
	reverse bool
	items   []mergingIterItem
}

func (h *mergingIterHeap) len() int {
	return len(h.items)
}

func (h *mergingIterHeap) less(i, j int) bool {
	// This key comparison only uses the user key and not the boundKind. Bound
	// kind doesn't matter because when stepping over a user key,
	// findNextFragmentSet and findPrevFragmentSet skip past all heap items with
	// that user key, and makes no assumptions on ordering. All other heap
	// examinations only consider the user key.
	ik, jk := h.items[i].key, h.items[j].key
	c := h.cmp(ik, jk)
	if h.reverse {
		return c > 0
	}
	return c < 0
}

func (h *mergingIterHeap) swap(i, j int) {
	h.items[i], h.items[j] = h.items[j], h.items[i]
}

// init, fix, up and down are copied from the go stdlib.
func (h *mergingIterHeap) init() {
	// heapify
	n := h.len()
	for i := n/2 - 1; i >= 0; i-- {
		h.down(i, n)
	}
}

func (h *mergingIterHeap) fix(i int) {
	if !h.down(i, h.len()) {
		h.up(i)
	}
}

func (h *mergingIterHeap) pop() *mergingIterItem {
	n := h.len() - 1
	h.swap(0, n)
	h.down(0, n)
	item := &h.items[n]
	h.items = h.items[:n]
	return item
}

func (h *mergingIterHeap) up(j int) {
	for {
		i := (j - 1) / 2 // parent
		if i == j || !h.less(j, i) {
			break
		}
		h.swap(i, j)
		j = i
	}
}

func (h *mergingIterHeap) down(i0, n int) bool {
	i := i0
	for {
		j1 := 2*i + 1
		if j1 >= n || j1 < 0 { // j1 < 0 after int overflow
			break
		}
		j := j1 // left child
		if j2 := j1 + 1; j2 < n && h.less(j2, j1) {
			j = j2 // = 2*i + 2  // right child
		}
		if !h.less(j, i) {
			break
		}
		h.swap(i, j)
		i = j
	}
	return i > i0
}

type boundKind int8

const (
	boundKindInvalid boundKind = iota
	boundKindFragmentStart
	boundKindFragmentEnd
)

type boundKey struct {
	kind boundKind
	key  []byte
	// span holds the span the bound key comes from.
	//
	// If kind is boundKindFragmentStart, then key is span.Start. If kind is
	// boundKindFragmentEnd, then key is span.End.
	span *Span
}

func (k boundKey) valid() bool {
	return k.kind != boundKindInvalid
}

func (k boundKey) String() string {
	var buf bytes.Buffer
	switch k.kind {
	case boundKindInvalid:
		fmt.Fprint(&buf, "invalid")
	case boundKindFragmentStart:
		fmt.Fprint(&buf, "fragment-start")
	case boundKindFragmentEnd:
		fmt.Fprint(&buf, "fragment-end  ")
	default:
		fmt.Fprintf(&buf, "unknown-kind(%d)", k.kind)
	}
	fmt.Fprintf(&buf, " %s [", k.key)
	fmt.Fprintf(&buf, "%s", k.span)
	fmt.Fprint(&buf, "]")
	return buf.String()
}