Initial commit

vendor/github.com/cockroachdb/pebble/sstable/block_property.go

@@ -0,0 +1,602 @@
+// Copyright 2021 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 sstable
+
+import (
+	"encoding/binary"
+	"fmt"
+	"math"
+	"sync"
+
+	"github.com/cockroachdb/pebble/internal/base"
+	"github.com/cockroachdb/pebble/internal/rangekey"
+)
+
+// Block properties are an optional user-facing feature that can be used to
+// filter data blocks (and whole sstables) from an Iterator before they are
+// loaded. They do not apply to range delete blocks. These are expected to
+// very concisely represent a set of some attribute value contained within the
+// key or value, such that the set includes all the attribute values in the
+// block. This has some similarities with OLAP pruning approaches that
+// maintain min-max attribute values for some column (which concisely
+// represent a set), that is then used to prune at query time. In Pebble's
+// case, data blocks are small, typically 25-50KB, so these properties should
+// reduce their precision in order to be concise -- a good rule of thumb is to
+// not consume more than 50-100 bytes across all properties maintained for a
+// block, i.e., a 500x reduction compared to loading the data block.
+//
+// A block property must be assigned a unique name, which is encoded and
+// stored in the sstable. This name must be unique among all user-properties
+// encoded in an sstable.
+//
+// A property is represented as a []byte. A nil value or empty byte slice are
+// considered semantically identical. The caller is free to choose the
+// semantics of an empty byte slice e.g. they could use it to represent the
+// empty set or the universal set, whichever they think is more common and
+// therefore better to encode more concisely. The serialization of the
+// property for the various Finish*() calls in a BlockPropertyCollector
+// implementation should be identical, since the corresponding
+// BlockPropertyFilter implementation is not told the context in which it is
+// deserializing the property.
+//
+// Block properties are more general than table properties and should be
+// preferred over using table properties. A BlockPropertyCollector can achieve
+// identical behavior to table properties by returning the nil slice from
+// FinishDataBlock and FinishIndexBlock, and interpret them as the universal
+// set in BlockPropertyFilter, and return a non-universal set in FinishTable.
+//
+// Block property filtering is nondeterministic because the separation of keys
+// into blocks is nondeterministic. Clients use block-property filters to
+// implement efficient application of a filter F that applies to key-value pairs
+// (abbreviated as kv-filter). Consider correctness defined as surfacing exactly
+// the same key-value pairs that would be surfaced if one applied the filter F
+// above normal iteration. With this correctness definition, block property
+// filtering may introduce two kinds of errors:
+//
+//   a) Block property filtering that uses a kv-filter may produce additional
+//      key-value pairs that don't satisfy the filter because of the separation
+//      of keys into blocks. Clients may remove these extra key-value pairs by
+//      re-applying the kv filter while reading results back from Pebble.
+//
+//   b) Block property filtering may surface deleted key-value pairs if the
+//      the kv filter is not a strict function of the key's user key. A block
+//      containing k.DEL may be filtered, while a block containing the deleted
+//      key k.SET may not be filtered, if the kv filter applies to one but not
+//      the other.
+//
+//      This error may be avoided trivially by using a kv filter that is a pure
+//      function of the the user key. A filter that examines values or key kinds
+//      requires care to ensure F(k.SET, <value>) = F(k.DEL) = F(k.SINGLEDEL).
+//
+// The combination of range deletions and filtering by table-level properties
+// add another opportunity for deleted point keys to be surfaced. The pebble
+// Iterator stack takes care to correctly apply filtered tables' range deletions
+// to lower tables, preventing this form of nondeterministic error.
+
+// BlockPropertyCollector is used when writing a sstable.
+// - All calls to Add are included in the next FinishDataBlock, after which
+//   the next data block is expected to start.
+//
+// - The index entry generated for the data block, which contains the return
+//   value from FinishDataBlock, is not immediately included in the current
+//   index block. It is included when AddPrevDataBlockToIndexBlock is called.
+//   An alternative would be to return an opaque handle from FinishDataBlock
+//   and pass it to a new AddToIndexBlock method, which requires more
+//   plumbing, and passing of an interface{} results in a undesirable heap
+//   allocation. AddPrevDataBlockToIndexBlock must be called before keys are
+//   added to the new data block.
+type BlockPropertyCollector interface {
+	// Name returns the name of the block property collector.
+	Name() string
+	// Add is called with each new entry added to a data block in the sstable.
+	// The callee can assume that these are in sorted order.
+	Add(key InternalKey, value []byte) error
+	// FinishDataBlock is called when all the entries have been added to a
+	// data block. Subsequent Add calls will be for the next data block. It
+	// returns the property value for the finished block.
+	FinishDataBlock(buf []byte) ([]byte, error)
+	// AddPrevDataBlockToIndexBlock adds the entry corresponding to the
+	// previous FinishDataBlock to the current index block.
+	AddPrevDataBlockToIndexBlock()
+	// FinishIndexBlock is called when an index block, containing all the
+	// key-value pairs since the last FinishIndexBlock, will no longer see new
+	// entries. It returns the property value for the index block.
+	FinishIndexBlock(buf []byte) ([]byte, error)
+	// FinishTable is called when the sstable is finished, and returns the
+	// property value for the sstable.
+	FinishTable(buf []byte) ([]byte, error)
+}
+
+// SuffixReplaceableBlockCollector is an extension to the BlockPropertyCollector
+// interface that allows a block property collector to indicate the it supports
+// being *updated* during suffix replacement, i.e. when an existing SST in which
+// all keys have the same key suffix is updated to have a new suffix.
+//
+// A collector which supports being updated in such cases must be able to derive
+// its updated value from its old value and the change being made to the suffix,
+// without needing to be passed each updated K/V.
+//
+// For example, a collector that only inspects values would can simply copy its
+// previously computed property as-is, since key-suffix replacement does not
+// change values, while a collector that depends only on key suffixes, like one
+// which collected mvcc-timestamp bounds from timestamp-suffixed keys, can just
+// set its new bounds from the new suffix, as it is common to all keys, without
+// needing to recompute it from every key.
+//
+// An implementation of DataBlockIntervalCollector can also implement this
+// interface, in which case the BlockPropertyCollector returned by passing it to
+// NewBlockIntervalCollector will also implement this interface automatically.
+type SuffixReplaceableBlockCollector interface {
+	// UpdateKeySuffixes is called when a block is updated to change the suffix of
+	// all keys in the block, and is passed the old value for that prop, if any,
+	// for that block as well as the old and new suffix.
+	UpdateKeySuffixes(oldProp []byte, oldSuffix, newSuffix []byte) error
+}
+
+// BlockPropertyFilter is used in an Iterator to filter sstables and blocks
+// within the sstable. It should not maintain any per-sstable state, and must
+// be thread-safe.
+type BlockPropertyFilter = base.BlockPropertyFilter
+
+// BlockIntervalCollector is a helper implementation of BlockPropertyCollector
+// for users who want to represent a set of the form [lower,upper) where both
+// lower and upper are uint64, and lower <= upper.
+//
+// The set is encoded as:
+// - Two varint integers, (lower,upper-lower), when upper-lower > 0
+// - Nil, when upper-lower=0
+//
+// Users must not expect this to preserve differences between empty sets --
+// they will all get turned into the semantically equivalent [0,0).
+//
+// A BlockIntervalCollector that collects over point and range keys needs to
+// have both the point and range DataBlockIntervalCollector specified, since
+// point and range keys are fed to the BlockIntervalCollector in an interleaved
+// fashion, independently of one another. This also implies that the
+// DataBlockIntervalCollectors for point and range keys should be references to
+// independent instances, rather than references to the same collector, as point
+// and range keys are tracked independently.
+type BlockIntervalCollector struct {
+	name   string
+	points DataBlockIntervalCollector
+	ranges DataBlockIntervalCollector
+
+	blockInterval interval
+	indexInterval interval
+	tableInterval interval
+}
+
+var _ BlockPropertyCollector = &BlockIntervalCollector{}
+
+// DataBlockIntervalCollector is the interface used by BlockIntervalCollector
+// that contains the actual logic pertaining to the property. It only
+// maintains state for the current data block, and resets that state in
+// FinishDataBlock. This interface can be used to reduce parsing costs.
+type DataBlockIntervalCollector interface {
+	// Add is called with each new entry added to a data block in the sstable.
+	// The callee can assume that these are in sorted order.
+	Add(key InternalKey, value []byte) error
+	// FinishDataBlock is called when all the entries have been added to a
+	// data block. Subsequent Add calls will be for the next data block. It
+	// returns the [lower, upper) for the finished block.
+	FinishDataBlock() (lower uint64, upper uint64, err error)
+}
+
+// NewBlockIntervalCollector constructs a BlockIntervalCollector with the given
+// name. The BlockIntervalCollector makes use of the given point and range key
+// DataBlockIntervalCollectors when encountering point and range keys,
+// respectively.
+//
+// The caller may pass a nil DataBlockIntervalCollector for one of the point or
+// range key collectors, in which case keys of those types will be ignored. This
+// allows for flexible construction of BlockIntervalCollectors that operate on
+// just point keys, just range keys, or both point and range keys.
+//
+// If both point and range keys are to be tracked, two independent collectors
+// should be provided, rather than the same collector passed in twice (see the
+// comment on BlockIntervalCollector for more detail)
+func NewBlockIntervalCollector(
+	name string, pointCollector, rangeCollector DataBlockIntervalCollector,
+) BlockPropertyCollector {
+	if pointCollector == nil && rangeCollector == nil {
+		panic("sstable: at least one interval collector must be provided")
+	}
+	bic := BlockIntervalCollector{
+		name:   name,
+		points: pointCollector,
+		ranges: rangeCollector,
+	}
+	if _, ok := pointCollector.(SuffixReplaceableBlockCollector); ok {
+		return &suffixReplacementBlockCollectorWrapper{bic}
+	}
+	return &bic
+}
+
+// Name implements the BlockPropertyCollector interface.
+func (b *BlockIntervalCollector) Name() string {
+	return b.name
+}
+
+// Add implements the BlockPropertyCollector interface.
+func (b *BlockIntervalCollector) Add(key InternalKey, value []byte) error {
+	if rangekey.IsRangeKey(key.Kind()) {
+		if b.ranges != nil {
+			return b.ranges.Add(key, value)
+		}
+	} else if b.points != nil {
+		return b.points.Add(key, value)
+	}
+	return nil
+}
+
+// FinishDataBlock implements the BlockPropertyCollector interface.
+func (b *BlockIntervalCollector) FinishDataBlock(buf []byte) ([]byte, error) {
+	if b.points == nil {
+		return buf, nil
+	}
+	var err error
+	b.blockInterval.lower, b.blockInterval.upper, err = b.points.FinishDataBlock()
+	if err != nil {
+		return buf, err
+	}
+	buf = b.blockInterval.encode(buf)
+	b.tableInterval.union(b.blockInterval)
+	return buf, nil
+}
+
+// AddPrevDataBlockToIndexBlock implements the BlockPropertyCollector
+// interface.
+func (b *BlockIntervalCollector) AddPrevDataBlockToIndexBlock() {
+	b.indexInterval.union(b.blockInterval)
+	b.blockInterval = interval{}
+}
+
+// FinishIndexBlock implements the BlockPropertyCollector interface.
+func (b *BlockIntervalCollector) FinishIndexBlock(buf []byte) ([]byte, error) {
+	buf = b.indexInterval.encode(buf)
+	b.indexInterval = interval{}
+	return buf, nil
+}
+
+// FinishTable implements the BlockPropertyCollector interface.
+func (b *BlockIntervalCollector) FinishTable(buf []byte) ([]byte, error) {
+	// If the collector is tracking range keys, the range key interval is union-ed
+	// with the point key interval for the table.
+	if b.ranges != nil {
+		var rangeInterval interval
+		var err error
+		rangeInterval.lower, rangeInterval.upper, err = b.ranges.FinishDataBlock()
+		if err != nil {
+			return buf, err
+		}
+		b.tableInterval.union(rangeInterval)
+	}
+	return b.tableInterval.encode(buf), nil
+}
+
+type interval struct {
+	lower uint64
+	upper uint64
+}
+
+func (i interval) encode(buf []byte) []byte {
+	if i.lower < i.upper {
+		var encoded [binary.MaxVarintLen64 * 2]byte
+		n := binary.PutUvarint(encoded[:], i.lower)
+		n += binary.PutUvarint(encoded[n:], i.upper-i.lower)
+		buf = append(buf, encoded[:n]...)
+	}
+	return buf
+}
+
+func (i *interval) decode(buf []byte) error {
+	if len(buf) == 0 {
+		*i = interval{}
+		return nil
+	}
+	var n int
+	i.lower, n = binary.Uvarint(buf)
+	if n <= 0 || n >= len(buf) {
+		return base.CorruptionErrorf("cannot decode interval from buf %x", buf)
+	}
+	pos := n
+	i.upper, n = binary.Uvarint(buf[pos:])
+	pos += n
+	if pos != len(buf) || n <= 0 {
+		return base.CorruptionErrorf("cannot decode interval from buf %x", buf)
+	}
+	// Delta decode.
+	i.upper += i.lower
+	if i.upper < i.lower {
+		return base.CorruptionErrorf("unexpected overflow, upper %d < lower %d", i.upper, i.lower)
+	}
+	return nil
+}
+
+func (i *interval) union(x interval) {
+	if x.lower >= x.upper {
+		// x is the empty set.
+		return
+	}
+	if i.lower >= i.upper {
+		// i is the empty set.
+		*i = x
+		return
+	}
+	// Both sets are non-empty.
+	if x.lower < i.lower {
+		i.lower = x.lower
+	}
+	if x.upper > i.upper {
+		i.upper = x.upper
+	}
+}
+
+func (i interval) intersects(x interval) bool {
+	if i.lower >= i.upper || x.lower >= x.upper {
+		// At least one of the sets is empty.
+		return false
+	}
+	// Neither set is empty.
+	return i.upper > x.lower && i.lower < x.upper
+}
+
+type suffixReplacementBlockCollectorWrapper struct {
+	BlockIntervalCollector
+}
+
+// UpdateKeySuffixes implements the SuffixReplaceableBlockCollector interface.
+func (w *suffixReplacementBlockCollectorWrapper) UpdateKeySuffixes(
+	oldProp []byte, from, to []byte,
+) error {
+	return w.BlockIntervalCollector.points.(SuffixReplaceableBlockCollector).UpdateKeySuffixes(oldProp, from, to)
+}
+
+// blockIntervalFilter is an implementation of BlockPropertyFilter when the
+// corresponding collector is a BlockIntervalCollector. That is, the set is of
+// the form [lower, upper).
+type blockIntervalFilter struct {
+	name           string
+	filterInterval interval
+}
+
+// NewBlockIntervalFilter constructs a BlockPropertyFilter that filters blocks
+// based on an interval property collected by BlockIntervalCollector and the
+// given [lower, upper) bounds. The given name specifies the
+// BlockIntervalCollector's properties to read.
+func NewBlockIntervalFilter(name string, lower uint64, upper uint64) BlockPropertyFilter {
+	return &blockIntervalFilter{
+		name:           name,
+		filterInterval: interval{lower: lower, upper: upper},
+	}
+}
+
+// Name implements the BlockPropertyFilter interface.
+func (b *blockIntervalFilter) Name() string {
+	return b.name
+}
+
+// Intersects implements the BlockPropertyFilter interface.
+func (b *blockIntervalFilter) Intersects(prop []byte) (bool, error) {
+	var i interval
+	if err := i.decode(prop); err != nil {
+		return false, err
+	}
+	return i.intersects(b.filterInterval), nil
+}
+
+// When encoding block properties for each block, we cannot afford to encode
+// the name. Instead, the name is mapped to a shortID, in the scope of that
+// sstable, and the shortID is encoded. Since we use a uint8, there is a limit
+// of 256 block property collectors per sstable.
+type shortID uint8
+
+type blockPropertiesEncoder struct {
+	propsBuf []byte
+	scratch  []byte
+}
+
+func (e *blockPropertiesEncoder) getScratchForProp() []byte {
+	return e.scratch[:0]
+}
+
+func (e *blockPropertiesEncoder) resetProps() {
+	e.propsBuf = e.propsBuf[:0]
+}
+
+func (e *blockPropertiesEncoder) addProp(id shortID, scratch []byte) {
+	const lenID = 1
+	lenProp := uvarintLen(uint32(len(scratch)))
+	n := lenID + lenProp + len(scratch)
+	if cap(e.propsBuf)-len(e.propsBuf) < n {
+		size := len(e.propsBuf) + 2*n
+		if size < 2*cap(e.propsBuf) {
+			size = 2 * cap(e.propsBuf)
+		}
+		buf := make([]byte, len(e.propsBuf), size)
+		copy(buf, e.propsBuf)
+		e.propsBuf = buf
+	}
+	pos := len(e.propsBuf)
+	b := e.propsBuf[pos : pos+lenID]
+	b[0] = byte(id)
+	pos += lenID
+	b = e.propsBuf[pos : pos+lenProp]
+	n = binary.PutUvarint(b, uint64(len(scratch)))
+	pos += n
+	b = e.propsBuf[pos : pos+len(scratch)]
+	pos += len(scratch)
+	copy(b, scratch)
+	e.propsBuf = e.propsBuf[0:pos]
+	e.scratch = scratch
+}
+
+func (e *blockPropertiesEncoder) unsafeProps() []byte {
+	return e.propsBuf
+}
+
+func (e *blockPropertiesEncoder) props() []byte {
+	buf := make([]byte, len(e.propsBuf))
+	copy(buf, e.propsBuf)
+	return buf
+}
+
+type blockPropertiesDecoder struct {
+	props []byte
+}
+
+func (d *blockPropertiesDecoder) done() bool {
+	return len(d.props) == 0
+}
+
+// REQUIRES: !done()
+func (d *blockPropertiesDecoder) next() (id shortID, prop []byte, err error) {
+	const lenID = 1
+	id = shortID(d.props[0])
+	propLen, m := binary.Uvarint(d.props[lenID:])
+	n := lenID + m
+	if m <= 0 || propLen == 0 || (n+int(propLen)) > len(d.props) {
+		return 0, nil, base.CorruptionErrorf("corrupt block property length")
+	}
+	prop = d.props[n : n+int(propLen)]
+	d.props = d.props[n+int(propLen):]
+	return id, prop, nil
+}
+
+// BlockPropertiesFilterer provides filtering support when reading an sstable
+// in the context of an iterator that has a slice of BlockPropertyFilters.
+// After the call to NewBlockPropertiesFilterer, the caller must call
+// IntersectsUserPropsAndFinishInit to check if the sstable intersects with
+// the filters. If it does intersect, this function also finishes initializing
+// the BlockPropertiesFilterer using the shortIDs for the relevant filters.
+// Subsequent checks for relevance of a block should use the intersects
+// method.
+type BlockPropertiesFilterer struct {
+	filters []BlockPropertyFilter
+	// Maps shortID => index in filters. This can be sparse, and shortIDs for
+	// which there is no filter are represented with an index of -1. The
+	// length of this can be shorter than the shortIDs allocated in the
+	// sstable. e.g. if the sstable used shortIDs 0, 1, 2, 3, and the iterator
+	// has two filters, corresponding to shortIDs 2, 0, this would be:
+	// len(shortIDToFiltersIndex)==3, 0=>1, 1=>-1, 2=>0.
+	shortIDToFiltersIndex []int
+}
+
+var blockPropertiesFiltererPool = sync.Pool{
+	New: func() interface{} {
+		return &BlockPropertiesFilterer{}
+	},
+}
+
+// NewBlockPropertiesFilterer returns a partially initialized filterer. To complete
+// initialization, call IntersectsUserPropsAndFinishInit.
+func NewBlockPropertiesFilterer(filters []BlockPropertyFilter) *BlockPropertiesFilterer {
+	filterer := blockPropertiesFiltererPool.Get().(*BlockPropertiesFilterer)
+	*filterer = BlockPropertiesFilterer{filters: filters}
+	return filterer
+}
+
+func releaseBlockPropertiesFilterer(filterer *BlockPropertiesFilterer) {
+	*filterer = BlockPropertiesFilterer{
+		shortIDToFiltersIndex: filterer.shortIDToFiltersIndex[:0],
+	}
+	blockPropertiesFiltererPool.Put(filterer)
+}
+
+// IntersectsUserPropsAndFinishInit is called with the user properties map for
+// the sstable and returns whether the sstable intersects the filters. It
+// additionally initializes the shortIDToFiltersIndex for the filters that are
+// relevant to this sstable.
+func (f *BlockPropertiesFilterer) IntersectsUserPropsAndFinishInit(
+	userProperties map[string]string,
+) (bool, error) {
+	for i := range f.filters {
+		props, ok := userProperties[f.filters[i].Name()]
+		if !ok {
+			// Collector was not used when writing this file, so it is
+			// considered intersecting.
+			continue
+		}
+		byteProps := []byte(props)
+		if len(byteProps) < 1 {
+			return false, base.CorruptionErrorf(
+				"block properties for %s is corrupted", f.filters[i].Name())
+		}
+		shortID := shortID(byteProps[0])
+		intersects, err := f.filters[i].Intersects(byteProps[1:])
+		if err != nil || !intersects {
+			return false, err
+		}
+		// Intersects the sstable, so need to use this filter when
+		// deciding whether to read blocks.
+		n := len(f.shortIDToFiltersIndex)
+		if n <= int(shortID) {
+			if cap(f.shortIDToFiltersIndex) <= int(shortID) {
+				index := make([]int, shortID+1, 2*(shortID+1))
+				copy(index, f.shortIDToFiltersIndex)
+				f.shortIDToFiltersIndex = index
+			} else {
+				f.shortIDToFiltersIndex = f.shortIDToFiltersIndex[:shortID+1]
+			}
+			for j := n; j < int(shortID); j++ {
+				f.shortIDToFiltersIndex[j] = -1
+			}
+		}
+		f.shortIDToFiltersIndex[shortID] = i
+	}
+	return true, nil
+}
+
+func (f *BlockPropertiesFilterer) intersects(props []byte) (bool, error) {
+	i := 0
+	decoder := blockPropertiesDecoder{props: props}
+	for i < len(f.shortIDToFiltersIndex) {
+		var id int
+		var prop []byte
+		if !decoder.done() {
+			var shortID shortID
+			var err error
+			shortID, prop, err = decoder.next()
+			if err != nil {
+				return false, err
+			}
+			id = int(shortID)
+		} else {
+			id = math.MaxUint8 + 1
+		}
+		for i < len(f.shortIDToFiltersIndex) && id > i {
+			if f.shortIDToFiltersIndex[i] >= 0 {
+				// There is a filter for this id, but the property for this id
+				// is not encoded for this block.
+				intersects, err := f.filters[f.shortIDToFiltersIndex[i]].Intersects(nil)
+				if err != nil {
+					return false, err
+				}
+				if !intersects {
+					return false, nil
+				}
+			}
+			i++
+		}
+		if i >= len(f.shortIDToFiltersIndex) {
+			return true, nil
+		}
+		// INVARIANT: id <= i. And since i is always incremented by 1, id==i.
+		if id != i {
+			panic(fmt.Sprintf("%d != %d", id, i))
+		}
+		if f.shortIDToFiltersIndex[i] >= 0 {
+			intersects, err := f.filters[f.shortIDToFiltersIndex[i]].Intersects(prop)
+			if err != nil {
+				return false, err
+			}
+			if !intersects {
+				return false, nil
+			}
+		}
+		i++
+	}
+	return true, nil
+}

vendor/github.com/cockroachdb/pebble/sstable/comparer.go

@@ -0,0 +1,34 @@
+// Copyright 2011 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 sstable
+
+import "github.com/cockroachdb/pebble/internal/base"
+
+// Compare exports the base.Compare type.
+type Compare = base.Compare
+
+// Equal exports the base.Equal type.
+type Equal = base.Equal
+
+// AbbreviatedKey exports the base.AbbreviatedKey type.
+type AbbreviatedKey = base.AbbreviatedKey
+
+// Separator exports the base.Separator type.
+type Separator = base.Separator
+
+// Successor exports the base.Successor type.
+type Successor = base.Successor
+
+// Split exports the base.Split type.
+type Split = base.Split
+
+// Comparer exports the base.Comparer type.
+type Comparer = base.Comparer
+
+// DefaultComparer exports the base.DefaultComparer variable.
+var DefaultComparer = base.DefaultComparer
+
+// Merger exports the base.Merger type.
+type Merger = base.Merger

vendor/github.com/cockroachdb/pebble/sstable/compression.go

@@ -0,0 +1,98 @@
+// Copyright 2021 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 sstable
+
+import (
+	"encoding/binary"
+
+	"github.com/cockroachdb/errors"
+	"github.com/cockroachdb/pebble/internal/base"
+	"github.com/cockroachdb/pebble/internal/cache"
+	"github.com/golang/snappy"
+)
+
+func decompressedLen(blockType blockType, b []byte) (int, int, error) {
+	switch blockType {
+	case noCompressionBlockType:
+		return 0, 0, nil
+	case snappyCompressionBlockType:
+		l, err := snappy.DecodedLen(b)
+		return l, 0, err
+	case zstdCompressionBlockType:
+		// This will also be used by zlib, bzip2 and lz4 to retrieve the decodedLen
+		// if we implement these algorithms in the future.
+		decodedLenU64, varIntLen := binary.Uvarint(b)
+		if varIntLen <= 0 {
+			return 0, 0, base.CorruptionErrorf("pebble/table: compression block has invalid length")
+		}
+		return int(decodedLenU64), varIntLen, nil
+	default:
+		return 0, 0, base.CorruptionErrorf("pebble/table: unknown block compression: %d", errors.Safe(blockType))
+	}
+}
+
+func decompressInto(blockType blockType, compressed []byte, buf []byte) ([]byte, error) {
+	var result []byte
+	var err error
+	switch blockType {
+	case snappyCompressionBlockType:
+		result, err = snappy.Decode(buf, compressed)
+	case zstdCompressionBlockType:
+		result, err = decodeZstd(buf, compressed)
+	}
+	if err != nil {
+		return nil, base.MarkCorruptionError(err)
+	}
+	if len(result) != 0 && (len(result) != len(buf) || &result[0] != &buf[0]) {
+		return nil, base.CorruptionErrorf("pebble/table: decompressed into unexpected buffer: %p != %p",
+			errors.Safe(result), errors.Safe(buf))
+	}
+	return result, nil
+}
+
+// decompressBlock decompresses an SST block, with space allocated from a cache.
+func decompressBlock(cache *cache.Cache, blockType blockType, b []byte) (*cache.Value, error) {
+	if blockType == noCompressionBlockType {
+		return nil, nil
+	}
+	// first obtain the decoded length.
+	decodedLen, prefixLen, err := decompressedLen(blockType, b)
+	if err != nil {
+		return nil, err
+	}
+	if prefixLen != 0 {
+		b = b[prefixLen:]
+	}
+	// Allocate sufficient space from the cache.
+	decoded := cache.Alloc(decodedLen)
+	decodedBuf := decoded.Buf()
+	if _, err := decompressInto(blockType, b, decodedBuf); err != nil {
+		cache.Free(decoded)
+	}
+	return decoded, nil
+}
+
+// compressBlock compresses an SST block, using compressBuf as the desired destination.
+func compressBlock(
+	compression Compression, b []byte, compressedBuf []byte,
+) (blockType blockType, compressed []byte) {
+	switch compression {
+	case SnappyCompression:
+		return snappyCompressionBlockType, snappy.Encode(compressedBuf, b)
+	case NoCompression:
+		return noCompressionBlockType, b
+	}
+
+	if len(compressedBuf) < binary.MaxVarintLen64 {
+		compressedBuf = append(compressedBuf, make([]byte, binary.MaxVarintLen64-len(compressedBuf))...)
+	}
+	varIntLen := binary.PutUvarint(compressedBuf, uint64(len(b)))
+	switch compression {
+	case ZstdCompression:
+		return zstdCompressionBlockType, encodeZstd(compressedBuf, varIntLen, b)
+	default:
+		return noCompressionBlockType, b
+	}
+}

vendor/github.com/cockroachdb/pebble/sstable/compression_cgo.go

@@ -0,0 +1,34 @@
+// Copyright 2021 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.
+
+//go:build cgo
+// +build cgo
+
+package sstable
+
+import (
+	"bytes"
+
+	"github.com/DataDog/zstd"
+)
+
+// decodeZstd decompresses b with the Zstandard algorithm.
+// It reuses the preallocated capacity of decodedBuf if it is sufficient.
+// On success, it returns the decoded byte slice.
+func decodeZstd(decodedBuf, b []byte) ([]byte, error) {
+	return zstd.Decompress(decodedBuf, b)
+}
+
+// encodeZstd compresses b with the Zstandard algorithm at default compression
+// level (level 3). It reuses the preallocated capacity of compressedBuf if it
+// is sufficient. The subslice `compressedBuf[:varIntLen]` should already encode
+// the length of `b` before calling encodeZstd. It returns the encoded byte
+// slice, including the `compressedBuf[:varIntLen]` prefix.
+func encodeZstd(compressedBuf []byte, varIntLen int, b []byte) []byte {
+	buf := bytes.NewBuffer(compressedBuf[:varIntLen])
+	writer := zstd.NewWriterLevel(buf, 3)
+	writer.Write(b)
+	writer.Close()
+	return buf.Bytes()
+}

vendor/github.com/cockroachdb/pebble/sstable/compression_nocgo.go

@@ -0,0 +1,30 @@
+// Copyright 2021 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.
+
+//go:build !cgo
+// +build !cgo
+
+package sstable
+
+import "github.com/klauspost/compress/zstd"
+
+// decodeZstd decompresses b with the Zstandard algorithm.
+// It reuses the preallocated capacity of decodedBuf if it is sufficient.
+// On success, it returns the decoded byte slice.
+func decodeZstd(decodedBuf, b []byte) ([]byte, error) {
+	decoder, _ := zstd.NewReader(nil)
+	defer decoder.Close()
+	return decoder.DecodeAll(b, decodedBuf[:0])
+}
+
+// encodeZstd compresses b with the Zstandard algorithm at default compression
+// level (level 3). It reuses the preallocated capacity of compressedBuf if it
+// is sufficient. The subslice `compressedBuf[:varIntLen]` should already encode
+// the length of `b` before calling encodeZstd. It returns the encoded byte
+// slice, including the `compressedBuf[:varIntLen]` prefix.
+func encodeZstd(compressedBuf []byte, varIntLen int, b []byte) []byte {
+	encoder, _ := zstd.NewWriter(nil)
+	defer encoder.Close()
+	return encoder.EncodeAll(b, compressedBuf[:varIntLen])
+}

vendor/github.com/cockroachdb/pebble/sstable/filter.go

@@ -0,0 +1,102 @@
+// Copyright 2011 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 sstable
+
+import "sync/atomic"
+
+// FilterMetrics holds metrics for the filter policy.
+type FilterMetrics struct {
+	// The number of hits for the filter policy. This is the
+	// number of times the filter policy was successfully used to avoid access
+	// of a data block.
+	Hits int64
+	// The number of misses for the filter policy. This is the number of times
+	// the filter policy was checked but was unable to filter an access of a data
+	// block.
+	Misses int64
+}
+
+var dummyFilterMetrics FilterMetrics
+
+func (m *FilterMetrics) readerApply(r *Reader) {
+	if r.tableFilter != nil {
+		r.tableFilter.metrics = m
+	}
+}
+
+// BlockHandle is the file offset and length of a block.
+type BlockHandle struct {
+	Offset, Length uint64
+}
+
+// BlockHandleWithProperties is used for data blocks and first/lower level
+// index blocks, since they can be annotated using BlockPropertyCollectors.
+type BlockHandleWithProperties struct {
+	BlockHandle
+	Props []byte
+}
+
+type filterWriter interface {
+	addKey(key []byte)
+	finish() ([]byte, error)
+	metaName() string
+	policyName() string
+}
+
+type tableFilterReader struct {
+	policy  FilterPolicy
+	metrics *FilterMetrics
+}
+
+func newTableFilterReader(policy FilterPolicy) *tableFilterReader {
+	return &tableFilterReader{
+		policy:  policy,
+		metrics: &dummyFilterMetrics,
+	}
+}
+
+func (f *tableFilterReader) mayContain(data, key []byte) bool {
+	mayContain := f.policy.MayContain(TableFilter, data, key)
+	if mayContain {
+		atomic.AddInt64(&f.metrics.Misses, 1)
+	} else {
+		atomic.AddInt64(&f.metrics.Hits, 1)
+	}
+	return mayContain
+}
+
+type tableFilterWriter struct {
+	policy FilterPolicy
+	writer FilterWriter
+	// count is the count of the number of keys added to the filter.
+	count int
+}
+
+func newTableFilterWriter(policy FilterPolicy) *tableFilterWriter {
+	return &tableFilterWriter{
+		policy: policy,
+		writer: policy.NewWriter(TableFilter),
+	}
+}
+
+func (f *tableFilterWriter) addKey(key []byte) {
+	f.count++
+	f.writer.AddKey(key)
+}
+
+func (f *tableFilterWriter) finish() ([]byte, error) {
+	if f.count == 0 {
+		return nil, nil
+	}
+	return f.writer.Finish(nil), nil
+}
+
+func (f *tableFilterWriter) metaName() string {
+	return "fullfilter." + f.policy.Name()
+}
+
+func (f *tableFilterWriter) policyName() string {
+	return f.policy.Name()
+}

vendor/github.com/cockroachdb/pebble/sstable/format.go

@@ -0,0 +1,91 @@
+// 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 sstable
+
+import (
+	"github.com/cockroachdb/errors"
+	"github.com/cockroachdb/pebble/internal/base"
+)
+
+// TableFormat specifies the format version for sstables. The legacy LevelDB
+// format is format version 1.
+type TableFormat uint32
+
+// The available table formats, representing the tuple (magic number, version
+// number). Note that these values are not (and should not) be serialized to
+// disk. The ordering should follow the order the versions were introduced to
+// Pebble (i.e. the history is linear).
+const (
+	TableFormatUnspecified TableFormat = iota
+	TableFormatLevelDB
+	TableFormatRocksDBv2
+	TableFormatPebblev1 // Block properties.
+	TableFormatPebblev2 // Range keys.
+
+	TableFormatMax = TableFormatPebblev2
+)
+
+// ParseTableFormat parses the given magic bytes and version into its
+// corresponding internal TableFormat.
+func ParseTableFormat(magic []byte, version uint32) (TableFormat, error) {
+	switch string(magic) {
+	case levelDBMagic:
+		return TableFormatLevelDB, nil
+	case rocksDBMagic:
+		if version != rocksDBFormatVersion2 {
+			return TableFormatUnspecified, base.CorruptionErrorf(
+				"pebble/table: unsupported rocksdb format version %d", errors.Safe(version),
+			)
+		}
+		return TableFormatRocksDBv2, nil
+	case pebbleDBMagic:
+		switch version {
+		case 1:
+			return TableFormatPebblev1, nil
+		case 2:
+			return TableFormatPebblev2, nil
+		default:
+			return TableFormatUnspecified, base.CorruptionErrorf(
+				"pebble/table: unsupported pebble format version %d", errors.Safe(version),
+			)
+		}
+	default:
+		return TableFormatUnspecified, base.CorruptionErrorf(
+			"pebble/table: invalid table (bad magic number)",
+		)
+	}
+}
+
+// AsTuple returns the TableFormat's (Magic String, Version) tuple.
+func (f TableFormat) AsTuple() (string, uint32) {
+	switch f {
+	case TableFormatLevelDB:
+		return levelDBMagic, 0
+	case TableFormatRocksDBv2:
+		return rocksDBMagic, 2
+	case TableFormatPebblev1:
+		return pebbleDBMagic, 1
+	case TableFormatPebblev2:
+		return pebbleDBMagic, 2
+	default:
+		panic("sstable: unknown table format version tuple")
+	}
+}
+
+// String returns the TableFormat (Magic String,Version) tuple.
+func (f TableFormat) String() string {
+	switch f {
+	case TableFormatLevelDB:
+		return "(LevelDB)"
+	case TableFormatRocksDBv2:
+		return "(RocksDB,v2)"
+	case TableFormatPebblev1:
+		return "(Pebble,v1)"
+	case TableFormatPebblev2:
+		return "(Pebble,v2)"
+	default:
+		panic("sstable: unknown table format version tuple")
+	}
+}

vendor/github.com/cockroachdb/pebble/sstable/internal.go

@@ -0,0 +1,33 @@
+// 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 sstable
+
+import "github.com/cockroachdb/pebble/internal/base"
+
+// InternalKeyKind exports the base.InternalKeyKind type.
+type InternalKeyKind = base.InternalKeyKind
+
+// SeekGEFlags exports base.SeekGEFlags.
+type SeekGEFlags = base.SeekGEFlags
+
+// SeekLTFlags exports base.SeekLTFlags.
+type SeekLTFlags = base.SeekLTFlags
+
+// These constants are part of the file format, and should not be changed.
+const (
+	InternalKeyKindDelete          = base.InternalKeyKindDelete
+	InternalKeyKindSet             = base.InternalKeyKindSet
+	InternalKeyKindMerge           = base.InternalKeyKindMerge
+	InternalKeyKindLogData         = base.InternalKeyKindLogData
+	InternalKeyKindRangeDelete     = base.InternalKeyKindRangeDelete
+	InternalKeyKindMax             = base.InternalKeyKindMax
+	InternalKeyKindInvalid         = base.InternalKeyKindInvalid
+	InternalKeySeqNumBatch         = base.InternalKeySeqNumBatch
+	InternalKeySeqNumMax           = base.InternalKeySeqNumMax
+	InternalKeyRangeDeleteSentinel = base.InternalKeyRangeDeleteSentinel
+)
+
+// InternalKey exports the base.InternalKey type.
+type InternalKey = base.InternalKey

vendor/github.com/cockroachdb/pebble/sstable/options.go

@@ -0,0 +1,257 @@
+// Copyright 2019 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 sstable
+
+import (
+	"github.com/cockroachdb/pebble/internal/base"
+	"github.com/cockroachdb/pebble/internal/cache"
+)
+
+// Compression is the per-block compression algorithm to use.
+type Compression int
+
+// The available compression types.
+const (
+	DefaultCompression Compression = iota
+	NoCompression
+	SnappyCompression
+	ZstdCompression
+	NCompression
+)
+
+func (c Compression) String() string {
+	switch c {
+	case DefaultCompression:
+		return "Default"
+	case NoCompression:
+		return "NoCompression"
+	case SnappyCompression:
+		return "Snappy"
+	case ZstdCompression:
+		return "ZSTD"
+	default:
+		return "Unknown"
+	}
+}
+
+// FilterType exports the base.FilterType type.
+type FilterType = base.FilterType
+
+// Exported TableFilter constants.
+const (
+	TableFilter = base.TableFilter
+)
+
+// FilterWriter exports the base.FilterWriter type.
+type FilterWriter = base.FilterWriter
+
+// FilterPolicy exports the base.FilterPolicy type.
+type FilterPolicy = base.FilterPolicy
+
+// TablePropertyCollector provides a hook for collecting user-defined
+// properties based on the keys and values stored in an sstable. A new
+// TablePropertyCollector is created for an sstable when the sstable is being
+// written.
+type TablePropertyCollector interface {
+	// Add is called with each new entry added to the sstable. While the sstable
+	// is itself sorted by key, do not assume that the entries are added in any
+	// order. In particular, the ordering of point entries and range tombstones
+	// is unspecified.
+	Add(key InternalKey, value []byte) error
+
+	// Finish is called when all entries have been added to the sstable. The
+	// collected properties (if any) should be added to the specified map. Note
+	// that in case of an error during sstable construction, Finish may not be
+	// called.
+	Finish(userProps map[string]string) error
+
+	// The name of the property collector.
+	Name() string
+}
+
+// SuffixReplaceableTableCollector is an extension to the TablePropertyCollector
+// interface that allows a table property collector to indicate that it supports
+// being *updated* during suffix replacement, i.e. when an existing SST in which
+// all keys have the same key suffix is updated to have a new suffix.
+//
+// A collector which supports being updated in such cases must be able to derive
+// its updated value from its old value and the change being made to the suffix,
+// without needing to be passed each updated K/V.
+//
+// For example, a collector that only inspects values can simply copy its
+// previously computed property as-is, since key-suffix replacement does not
+// change values, while a collector that depends only on key suffixes, like one
+// which collected mvcc-timestamp bounds from timestamp-suffixed keys, can just
+// set its new bounds from the new suffix, as it is common to all keys, without
+// needing to recompute it from every key.
+type SuffixReplaceableTableCollector interface {
+	// UpdateKeySuffixes is called when a table is updated to change the suffix of
+	// all keys in the table, and is passed the old value for that prop, if any,
+	// for that table as well as the old and new suffix.
+	UpdateKeySuffixes(oldProps map[string]string, oldSuffix, newSuffix []byte) error
+}
+
+// ReaderOptions holds the parameters needed for reading an sstable.
+type ReaderOptions struct {
+	// Cache is used to cache uncompressed blocks from sstables.
+	//
+	// The default cache size is a zero-size cache.
+	Cache *cache.Cache
+
+	// Comparer defines a total ordering over the space of []byte keys: a 'less
+	// than' relationship. The same comparison algorithm must be used for reads
+	// and writes over the lifetime of the DB.
+	//
+	// The default value uses the same ordering as bytes.Compare.
+	Comparer *Comparer
+
+	// Filters is a map from filter policy name to filter policy. It is used for
+	// debugging tools which may be used on multiple databases configured with
+	// different filter policies. It is not necessary to populate this filters
+	// map during normal usage of a DB.
+	Filters map[string]FilterPolicy
+
+	// Merger defines the associative merge operation to use for merging values
+	// written with {Batch,DB}.Merge. The MergerName is checked for consistency
+	// with the value stored in the sstable when it was written.
+	MergerName string
+}
+
+func (o ReaderOptions) ensureDefaults() ReaderOptions {
+	if o.Comparer == nil {
+		o.Comparer = base.DefaultComparer
+	}
+	if o.MergerName == "" {
+		o.MergerName = base.DefaultMerger.Name
+	}
+	return o
+}
+
+// WriterOptions holds the parameters used to control building an sstable.
+type WriterOptions struct {
+	// BlockRestartInterval is the number of keys between restart points
+	// for delta encoding of keys.
+	//
+	// The default value is 16.
+	BlockRestartInterval int
+
+	// BlockSize is the target uncompressed size in bytes of each table block.
+	//
+	// The default value is 4096.
+	BlockSize int
+
+	// BlockSizeThreshold finishes a block if the block size is larger than the
+	// specified percentage of the target block size and adding the next entry
+	// would cause the block to be larger than the target block size.
+	//
+	// The default value is 90
+	BlockSizeThreshold int
+
+	// Cache is used to cache uncompressed blocks from sstables.
+	//
+	// The default is a nil cache.
+	Cache *cache.Cache
+
+	// Comparer defines a total ordering over the space of []byte keys: a 'less
+	// than' relationship. The same comparison algorithm must be used for reads
+	// and writes over the lifetime of the DB.
+	//
+	// The default value uses the same ordering as bytes.Compare.
+	Comparer *Comparer
+
+	// Compression defines the per-block compression to use.
+	//
+	// The default value (DefaultCompression) uses snappy compression.
+	Compression Compression
+
+	// FilterPolicy defines a filter algorithm (such as a Bloom filter) that can
+	// reduce disk reads for Get calls.
+	//
+	// One such implementation is bloom.FilterPolicy(10) from the pebble/bloom
+	// package.
+	//
+	// The default value means to use no filter.
+	FilterPolicy FilterPolicy
+
+	// FilterType defines whether an existing filter policy is applied at a
+	// block-level or table-level. Block-level filters use less memory to create,
+	// but are slower to access as a check for the key in the index must first be
+	// performed to locate the filter block. A table-level filter will require
+	// memory proportional to the number of keys in an sstable to create, but
+	// avoids the index lookup when determining if a key is present. Table-level
+	// filters should be preferred except under constrained memory situations.
+	FilterType FilterType
+
+	// IndexBlockSize is the target uncompressed size in bytes of each index
+	// block. When the index block size is larger than this target, two-level
+	// indexes are automatically enabled. Setting this option to a large value
+	// (such as math.MaxInt32) disables the automatic creation of two-level
+	// indexes.
+	//
+	// The default value is the value of BlockSize.
+	IndexBlockSize int
+
+	// Merger defines the associative merge operation to use for merging values
+	// written with {Batch,DB}.Merge. The MergerName is checked for consistency
+	// with the value stored in the sstable when it was written.
+	MergerName string
+
+	// TableFormat specifies the format version for writing sstables. The default
+	// is TableFormatRocksDBv2 which creates RocksDB compatible sstables. Use
+	// TableFormatLevelDB to create LevelDB compatible sstable which can be used
+	// by a wider range of tools and libraries.
+	TableFormat TableFormat
+
+	// TablePropertyCollectors is a list of TablePropertyCollector creation
+	// functions. A new TablePropertyCollector is created for each sstable built
+	// and lives for the lifetime of the table.
+	TablePropertyCollectors []func() TablePropertyCollector
+
+	// BlockPropertyCollectors is a list of BlockPropertyCollector creation
+	// functions. A new BlockPropertyCollector is created for each sstable
+	// built and lives for the lifetime of writing that table.
+	BlockPropertyCollectors []func() BlockPropertyCollector
+
+	// Checksum specifies which checksum to use.
+	Checksum ChecksumType
+
+	// Parallelism is used to indicate that the sstable Writer is allowed to
+	// compress data blocks and write datablocks to disk in parallel with the
+	// Writer client goroutine.
+	Parallelism bool
+}
+
+func (o WriterOptions) ensureDefaults() WriterOptions {
+	if o.BlockRestartInterval <= 0 {
+		o.BlockRestartInterval = base.DefaultBlockRestartInterval
+	}
+	if o.BlockSize <= 0 {
+		o.BlockSize = base.DefaultBlockSize
+	}
+	if o.BlockSizeThreshold <= 0 {
+		o.BlockSizeThreshold = base.DefaultBlockSizeThreshold
+	}
+	if o.Comparer == nil {
+		o.Comparer = base.DefaultComparer
+	}
+	if o.Compression <= DefaultCompression || o.Compression >= NCompression {
+		o.Compression = SnappyCompression
+	}
+	if o.IndexBlockSize <= 0 {
+		o.IndexBlockSize = o.BlockSize
+	}
+	if o.MergerName == "" {
+		o.MergerName = base.DefaultMerger.Name
+	}
+	if o.Checksum == ChecksumTypeNone {
+		o.Checksum = ChecksumTypeCRC32c
+	}
+	// By default, if the table format is not specified, fall back to using the
+	// most compatible format.
+	if o.TableFormat == TableFormatUnspecified {
+		o.TableFormat = TableFormatRocksDBv2
+	}
+	return o
+}

vendor/github.com/cockroachdb/pebble/sstable/properties.go

@@ -0,0 +1,363 @@
+// 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 sstable
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"math"
+	"reflect"
+	"sort"
+	"unsafe"
+
+	"github.com/cockroachdb/pebble/internal/intern"
+)
+
+const propertiesBlockRestartInterval = math.MaxInt32
+const propGlobalSeqnumName = "rocksdb.external_sst_file.global_seqno"
+
+var propTagMap = make(map[string]reflect.StructField)
+var propBoolTrue = []byte{'1'}
+var propBoolFalse = []byte{'0'}
+
+var columnFamilyIDField = func() reflect.StructField {
+	f, ok := reflect.TypeOf(Properties{}).FieldByName("ColumnFamilyID")
+	if !ok {
+		panic("Properties.ColumnFamilyID field not found")
+	}
+	return f
+}()
+
+var propOffsetTagMap = make(map[uintptr]string)
+
+func init() {
+	t := reflect.TypeOf(Properties{})
+	for i := 0; i < t.NumField(); i++ {
+		f := t.Field(i)
+		if tag := f.Tag.Get("prop"); tag != "" {
+			switch f.Type.Kind() {
+			case reflect.Bool:
+			case reflect.Uint32:
+			case reflect.Uint64:
+			case reflect.String:
+			default:
+				panic(fmt.Sprintf("unsupported property field type: %s %s", f.Name, f.Type))
+			}
+			propTagMap[tag] = f
+			propOffsetTagMap[f.Offset] = tag
+		}
+	}
+}
+
+// Properties holds the sstable property values. The properties are
+// automatically populated during sstable creation and load from the properties
+// meta block when an sstable is opened.
+type Properties struct {
+	// ID of column family for this SST file, corresponding to the CF identified
+	// by column_family_name.
+	ColumnFamilyID uint64 `prop:"rocksdb.column.family.id"`
+	// Name of the column family with which this SST file is associated. Empty if
+	// the column family is unknown.
+	ColumnFamilyName string `prop:"rocksdb.column.family.name"`
+	// The name of the comparer used in this table.
+	ComparerName string `prop:"rocksdb.comparator"`
+	// The compression algorithm used to compress blocks.
+	CompressionName string `prop:"rocksdb.compression"`
+	// The compression options used to compress blocks.
+	CompressionOptions string `prop:"rocksdb.compression_options"`
+	// The time when the SST file was created. Since SST files are immutable,
+	// this is equivalent to last modified time.
+	CreationTime uint64 `prop:"rocksdb.creation.time"`
+	// The total size of all data blocks.
+	DataSize uint64 `prop:"rocksdb.data.size"`
+	// The external sstable version format. Version 2 is the one RocksDB has been
+	// using since 5.13. RocksDB only uses the global sequence number for an
+	// sstable if this property has been set.
+	ExternalFormatVersion uint32 `prop:"rocksdb.external_sst_file.version"`
+	// Actual SST file creation time. 0 means unknown.
+	FileCreationTime uint64 `prop:"rocksdb.file.creation.time"`
+	// The name of the filter policy used in this table. Empty if no filter
+	// policy is used.
+	FilterPolicyName string `prop:"rocksdb.filter.policy"`
+	// The size of filter block.
+	FilterSize uint64 `prop:"rocksdb.filter.size"`
+	// If 0, key is variable length. Otherwise number of bytes for each key.
+	FixedKeyLen uint64 `prop:"rocksdb.fixed.key.length"`
+	// Format version, reserved for backward compatibility.
+	FormatVersion uint64 `prop:"rocksdb.format.version"`
+	// The global sequence number to use for all entries in the table. Present if
+	// the table was created externally and ingested whole.
+	GlobalSeqNum uint64 `prop:"rocksdb.external_sst_file.global_seqno"`
+	// Whether the index key is user key or an internal key.
+	IndexKeyIsUserKey uint64 `prop:"rocksdb.index.key.is.user.key"`
+	// Total number of index partitions if kTwoLevelIndexSearch is used.
+	IndexPartitions uint64 `prop:"rocksdb.index.partitions"`
+	// The size of index block.
+	IndexSize uint64 `prop:"rocksdb.index.size"`
+	// The index type. TODO(peter): add a more detailed description.
+	IndexType uint32 `prop:"rocksdb.block.based.table.index.type"`
+	// Whether delta encoding is used to encode the index values.
+	IndexValueIsDeltaEncoded uint64 `prop:"rocksdb.index.value.is.delta.encoded"`
+	// The name of the merger used in this table. Empty if no merger is used.
+	MergerName string `prop:"rocksdb.merge.operator"`
+	// The number of blocks in this table.
+	NumDataBlocks uint64 `prop:"rocksdb.num.data.blocks"`
+	// The number of deletion entries in this table, including both point and
+	// range deletions.
+	NumDeletions uint64 `prop:"rocksdb.deleted.keys"`
+	// The number of entries in this table.
+	NumEntries uint64 `prop:"rocksdb.num.entries"`
+	// The number of merge operands in the table.
+	NumMergeOperands uint64 `prop:"rocksdb.merge.operands"`
+	// The number of range deletions in this table.
+	NumRangeDeletions uint64 `prop:"rocksdb.num.range-deletions"`
+	// The number of RANGEKEYDELs in this table.
+	NumRangeKeyDels uint64 `prop:"pebble.num.range-key-dels"`
+	// The number of RANGEKEYSETs in this table.
+	NumRangeKeySets uint64 `prop:"pebble.num.range-key-sets"`
+	// The number of RANGEKEYUNSETs in this table.
+	NumRangeKeyUnsets uint64 `prop:"pebble.num.range-key-unsets"`
+	// Timestamp of the earliest key. 0 if unknown.
+	OldestKeyTime uint64 `prop:"rocksdb.oldest.key.time"`
+	// The name of the prefix extractor used in this table. Empty if no prefix
+	// extractor is used.
+	PrefixExtractorName string `prop:"rocksdb.prefix.extractor.name"`
+	// If filtering is enabled, was the filter created on the key prefix.
+	PrefixFiltering bool `prop:"rocksdb.block.based.table.prefix.filtering"`
+	// A comma separated list of names of the property collectors used in this
+	// table.
+	PropertyCollectorNames string `prop:"rocksdb.property.collectors"`
+	// Total raw key size.
+	RawKeySize uint64 `prop:"rocksdb.raw.key.size"`
+	// Total raw rangekey key size.
+	RawRangeKeyKeySize uint64 `prop:"pebble.raw.range-key.key.size"`
+	// Total raw rangekey value size.
+	RawRangeKeyValueSize uint64 `prop:"pebble.raw.range-key.value.size"`
+	// Total raw value size.
+	RawValueSize uint64 `prop:"rocksdb.raw.value.size"`
+	// Size of the top-level index if kTwoLevelIndexSearch is used.
+	TopLevelIndexSize uint64 `prop:"rocksdb.top-level.index.size"`
+	// User collected properties.
+	UserProperties map[string]string
+	// If filtering is enabled, was the filter created on the whole key.
+	WholeKeyFiltering bool `prop:"rocksdb.block.based.table.whole.key.filtering"`
+
+	// Loaded set indicating which fields have been loaded from disk. Indexed by
+	// the field's byte offset within the struct
+	// (reflect.StructField.Offset). Only set if the properties have been loaded
+	// from a file. Only exported for testing purposes.
+	Loaded map[uintptr]struct{}
+}
+
+// NumPointDeletions returns the number of point deletions in this table.
+func (p *Properties) NumPointDeletions() uint64 {
+	return p.NumDeletions - p.NumRangeDeletions
+}
+
+// NumRangeKeys returns a count of the number of range keys in this table.
+func (p *Properties) NumRangeKeys() uint64 {
+	return p.NumRangeKeyDels + p.NumRangeKeySets + p.NumRangeKeyUnsets
+}
+
+func (p *Properties) String() string {
+	var buf bytes.Buffer
+	v := reflect.ValueOf(*p)
+	vt := v.Type()
+	for i := 0; i < v.NumField(); i++ {
+		ft := vt.Field(i)
+		tag := ft.Tag.Get("prop")
+		if tag == "" {
+			continue
+		}
+
+		f := v.Field(i)
+		// TODO(peter): Use f.IsZero() when we can rely on go1.13.
+		if zero := reflect.Zero(f.Type()); zero.Interface() == f.Interface() {
+			// Skip printing of zero values which were not loaded from disk.
+			if _, ok := p.Loaded[ft.Offset]; !ok {
+				continue
+			}
+		}
+
+		fmt.Fprintf(&buf, "%s: ", tag)
+		switch ft.Type.Kind() {
+		case reflect.Bool:
+			fmt.Fprintf(&buf, "%t\n", f.Bool())
+		case reflect.Uint32:
+			fmt.Fprintf(&buf, "%d\n", f.Uint())
+		case reflect.Uint64:
+			u := f.Uint()
+			if ft.Offset == columnFamilyIDField.Offset && u == math.MaxInt32 {
+				fmt.Fprintf(&buf, "-\n")
+			} else {
+				fmt.Fprintf(&buf, "%d\n", f.Uint())
+			}
+		case reflect.String:
+			fmt.Fprintf(&buf, "%s\n", f.String())
+		default:
+			panic("not reached")
+		}
+	}
+	keys := make([]string, 0, len(p.UserProperties))
+	for key := range p.UserProperties {
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	for _, key := range keys {
+		fmt.Fprintf(&buf, "%s: %s\n", key, p.UserProperties[key])
+	}
+	return buf.String()
+}
+
+func (p *Properties) load(b block, blockOffset uint64) error {
+	i, err := newRawBlockIter(bytes.Compare, b)
+	if err != nil {
+		return err
+	}
+	p.Loaded = make(map[uintptr]struct{})
+	v := reflect.ValueOf(p).Elem()
+	for valid := i.First(); valid; valid = i.Next() {
+		tag := intern.Bytes(i.Key().UserKey)
+		if f, ok := propTagMap[tag]; ok {
+			p.Loaded[f.Offset] = struct{}{}
+			field := v.FieldByIndex(f.Index)
+			switch f.Type.Kind() {
+			case reflect.Bool:
+				field.SetBool(bytes.Equal(i.Value(), propBoolTrue))
+			case reflect.Uint32:
+				field.SetUint(uint64(binary.LittleEndian.Uint32(i.Value())))
+			case reflect.Uint64:
+				var n uint64
+				if tag == propGlobalSeqnumName {
+					n = binary.LittleEndian.Uint64(i.Value())
+				} else {
+					n, _ = binary.Uvarint(i.Value())
+				}
+				field.SetUint(n)
+			case reflect.String:
+				field.SetString(intern.Bytes(i.Value()))
+			default:
+				panic("not reached")
+			}
+			continue
+		}
+		if p.UserProperties == nil {
+			p.UserProperties = make(map[string]string)
+		}
+		p.UserProperties[tag] = string(i.Value())
+	}
+	return nil
+}
+
+func (p *Properties) saveBool(m map[string][]byte, offset uintptr, value bool) {
+	tag := propOffsetTagMap[offset]
+	if value {
+		m[tag] = propBoolTrue
+	} else {
+		m[tag] = propBoolFalse
+	}
+}
+
+func (p *Properties) saveUint32(m map[string][]byte, offset uintptr, value uint32) {
+	var buf [4]byte
+	binary.LittleEndian.PutUint32(buf[:], value)
+	m[propOffsetTagMap[offset]] = buf[:]
+}
+
+func (p *Properties) saveUint64(m map[string][]byte, offset uintptr, value uint64) {
+	var buf [8]byte
+	binary.LittleEndian.PutUint64(buf[:], value)
+	m[propOffsetTagMap[offset]] = buf[:]
+}
+
+func (p *Properties) saveUvarint(m map[string][]byte, offset uintptr, value uint64) {
+	var buf [10]byte
+	n := binary.PutUvarint(buf[:], value)
+	m[propOffsetTagMap[offset]] = buf[:n]
+}
+
+func (p *Properties) saveString(m map[string][]byte, offset uintptr, value string) {
+	m[propOffsetTagMap[offset]] = []byte(value)
+}
+
+func (p *Properties) save(w *rawBlockWriter) {
+	m := make(map[string][]byte)
+	for k, v := range p.UserProperties {
+		m[k] = []byte(v)
+	}
+
+	p.saveUvarint(m, unsafe.Offsetof(p.ColumnFamilyID), p.ColumnFamilyID)
+	if p.ColumnFamilyName != "" {
+		p.saveString(m, unsafe.Offsetof(p.ColumnFamilyName), p.ColumnFamilyName)
+	}
+	if p.ComparerName != "" {
+		p.saveString(m, unsafe.Offsetof(p.ComparerName), p.ComparerName)
+	}
+	if p.CompressionName != "" {
+		p.saveString(m, unsafe.Offsetof(p.CompressionName), p.CompressionName)
+	}
+	if p.CompressionOptions != "" {
+		p.saveString(m, unsafe.Offsetof(p.CompressionOptions), p.CompressionOptions)
+	}
+	p.saveUvarint(m, unsafe.Offsetof(p.CreationTime), p.CreationTime)
+	p.saveUvarint(m, unsafe.Offsetof(p.DataSize), p.DataSize)
+	if p.ExternalFormatVersion != 0 {
+		p.saveUint32(m, unsafe.Offsetof(p.ExternalFormatVersion), p.ExternalFormatVersion)
+		p.saveUint64(m, unsafe.Offsetof(p.GlobalSeqNum), p.GlobalSeqNum)
+	}
+	if p.FileCreationTime > 0 {
+		p.saveUvarint(m, unsafe.Offsetof(p.FileCreationTime), p.FileCreationTime)
+	}
+	if p.FilterPolicyName != "" {
+		p.saveString(m, unsafe.Offsetof(p.FilterPolicyName), p.FilterPolicyName)
+	}
+	p.saveUvarint(m, unsafe.Offsetof(p.FilterSize), p.FilterSize)
+	p.saveUvarint(m, unsafe.Offsetof(p.FixedKeyLen), p.FixedKeyLen)
+	p.saveUvarint(m, unsafe.Offsetof(p.FormatVersion), p.FormatVersion)
+	p.saveUvarint(m, unsafe.Offsetof(p.IndexKeyIsUserKey), p.IndexKeyIsUserKey)
+	if p.IndexPartitions != 0 {
+		p.saveUvarint(m, unsafe.Offsetof(p.IndexPartitions), p.IndexPartitions)
+		p.saveUvarint(m, unsafe.Offsetof(p.TopLevelIndexSize), p.TopLevelIndexSize)
+	}
+	p.saveUvarint(m, unsafe.Offsetof(p.IndexSize), p.IndexSize)
+	p.saveUint32(m, unsafe.Offsetof(p.IndexType), p.IndexType)
+	p.saveUvarint(m, unsafe.Offsetof(p.IndexValueIsDeltaEncoded), p.IndexValueIsDeltaEncoded)
+	if p.MergerName != "" {
+		p.saveString(m, unsafe.Offsetof(p.MergerName), p.MergerName)
+	}
+	p.saveUvarint(m, unsafe.Offsetof(p.NumDataBlocks), p.NumDataBlocks)
+	p.saveUvarint(m, unsafe.Offsetof(p.NumEntries), p.NumEntries)
+	p.saveUvarint(m, unsafe.Offsetof(p.NumDeletions), p.NumDeletions)
+	p.saveUvarint(m, unsafe.Offsetof(p.NumMergeOperands), p.NumMergeOperands)
+	p.saveUvarint(m, unsafe.Offsetof(p.NumRangeDeletions), p.NumRangeDeletions)
+	if p.NumRangeKeys() > 0 {
+		p.saveUvarint(m, unsafe.Offsetof(p.NumRangeKeyDels), p.NumRangeKeyDels)
+		p.saveUvarint(m, unsafe.Offsetof(p.NumRangeKeySets), p.NumRangeKeySets)
+		p.saveUvarint(m, unsafe.Offsetof(p.NumRangeKeyUnsets), p.NumRangeKeyUnsets)
+		p.saveUvarint(m, unsafe.Offsetof(p.RawRangeKeyKeySize), p.RawRangeKeyKeySize)
+		p.saveUvarint(m, unsafe.Offsetof(p.RawRangeKeyValueSize), p.RawRangeKeyValueSize)
+	}
+	p.saveUvarint(m, unsafe.Offsetof(p.OldestKeyTime), p.OldestKeyTime)
+	if p.PrefixExtractorName != "" {
+		p.saveString(m, unsafe.Offsetof(p.PrefixExtractorName), p.PrefixExtractorName)
+	}
+	p.saveBool(m, unsafe.Offsetof(p.PrefixFiltering), p.PrefixFiltering)
+	if p.PropertyCollectorNames != "" {
+		p.saveString(m, unsafe.Offsetof(p.PropertyCollectorNames), p.PropertyCollectorNames)
+	}
+	p.saveUvarint(m, unsafe.Offsetof(p.RawKeySize), p.RawKeySize)
+	p.saveUvarint(m, unsafe.Offsetof(p.RawValueSize), p.RawValueSize)
+	p.saveBool(m, unsafe.Offsetof(p.WholeKeyFiltering), p.WholeKeyFiltering)
+
+	keys := make([]string, 0, len(m))
+	for key := range m {
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	for _, key := range keys {
+		w.add(InternalKey{UserKey: []byte(key)}, m[key])
+	}
+}

vendor/github.com/cockroachdb/pebble/sstable/raw_block.go

@@ -0,0 +1,273 @@
+// 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 sstable
+
+import (
+	"encoding/binary"
+	"sort"
+	"unsafe"
+
+	"github.com/cockroachdb/pebble/internal/base"
+)
+
+type rawBlockWriter struct {
+	blockWriter
+}
+
+func (w *rawBlockWriter) add(key InternalKey, value []byte) {
+	w.curKey, w.prevKey = w.prevKey, w.curKey
+
+	size := len(key.UserKey)
+	if cap(w.curKey) < size {
+		w.curKey = make([]byte, 0, size*2)
+	}
+	w.curKey = w.curKey[:size]
+	copy(w.curKey, key.UserKey)
+
+	w.store(size, value)
+}
+
+// rawBlockIter is an iterator over a single block of data. Unlike blockIter,
+// keys are stored in "raw" format (i.e. not as internal keys). Note that there
+// is significant similarity between this code and the code in blockIter. Yet
+// reducing duplication is difficult due to the blockIter being performance
+// critical.
+type rawBlockIter struct {
+	cmp         Compare
+	offset      int32
+	nextOffset  int32
+	restarts    int32
+	numRestarts int32
+	ptr         unsafe.Pointer
+	data        []byte
+	key, val    []byte
+	ikey        InternalKey
+	cached      []blockEntry
+	cachedBuf   []byte
+}
+
+func newRawBlockIter(cmp Compare, block block) (*rawBlockIter, error) {
+	i := &rawBlockIter{}
+	return i, i.init(cmp, block)
+}
+
+func (i *rawBlockIter) init(cmp Compare, block block) error {
+	numRestarts := int32(binary.LittleEndian.Uint32(block[len(block)-4:]))
+	if numRestarts == 0 {
+		return base.CorruptionErrorf("pebble/table: invalid table (block has no restart points)")
+	}
+	i.cmp = cmp
+	i.restarts = int32(len(block)) - 4*(1+numRestarts)
+	i.numRestarts = numRestarts
+	i.ptr = unsafe.Pointer(&block[0])
+	i.data = block
+	if i.key == nil {
+		i.key = make([]byte, 0, 256)
+	} else {
+		i.key = i.key[:0]
+	}
+	i.val = nil
+	i.clearCache()
+	return nil
+}
+
+func (i *rawBlockIter) readEntry() {
+	ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(i.offset))
+	shared, ptr := decodeVarint(ptr)
+	unshared, ptr := decodeVarint(ptr)
+	value, ptr := decodeVarint(ptr)
+	i.key = append(i.key[:shared], getBytes(ptr, int(unshared))...)
+	i.key = i.key[:len(i.key):len(i.key)]
+	ptr = unsafe.Pointer(uintptr(ptr) + uintptr(unshared))
+	i.val = getBytes(ptr, int(value))
+	i.nextOffset = int32(uintptr(ptr)-uintptr(i.ptr)) + int32(value)
+}
+
+func (i *rawBlockIter) loadEntry() {
+	i.readEntry()
+	i.ikey.UserKey = i.key
+}
+
+func (i *rawBlockIter) clearCache() {
+	i.cached = i.cached[:0]
+	i.cachedBuf = i.cachedBuf[:0]
+}
+
+func (i *rawBlockIter) cacheEntry() {
+	var valStart int32
+	valSize := int32(len(i.val))
+	if valSize > 0 {
+		valStart = int32(uintptr(unsafe.Pointer(&i.val[0])) - uintptr(i.ptr))
+	}
+
+	i.cached = append(i.cached, blockEntry{
+		offset:   i.offset,
+		keyStart: int32(len(i.cachedBuf)),
+		keyEnd:   int32(len(i.cachedBuf) + len(i.key)),
+		valStart: valStart,
+		valSize:  valSize,
+	})
+	i.cachedBuf = append(i.cachedBuf, i.key...)
+}
+
+// SeekGE implements internalIterator.SeekGE, as documented in the pebble
+// package.
+func (i *rawBlockIter) SeekGE(key []byte) bool {
+	// Find the index of the smallest restart point whose key is > the key
+	// sought; index will be numRestarts if there is no such restart point.
+	i.offset = 0
+	index := sort.Search(int(i.numRestarts), func(j int) bool {
+		offset := int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*j:]))
+		// For a restart point, there are 0 bytes shared with the previous key.
+		// The varint encoding of 0 occupies 1 byte.
+		ptr := unsafe.Pointer(uintptr(i.ptr) + uintptr(offset+1))
+		// Decode the key at that restart point, and compare it to the key sought.
+		v1, ptr := decodeVarint(ptr)
+		_, ptr = decodeVarint(ptr)
+		s := getBytes(ptr, int(v1))
+		return i.cmp(key, s) < 0
+	})
+
+	// Since keys are strictly increasing, if index > 0 then the restart point at
+	// index-1 will be the largest whose key is <= the key sought.  If index ==
+	// 0, then all keys in this block are larger than the key sought, and offset
+	// remains at zero.
+	if index > 0 {
+		i.offset = int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*(index-1):]))
+	}
+	i.loadEntry()
+
+	// Iterate from that restart point to somewhere >= the key sought.
+	for valid := i.Valid(); valid; valid = i.Next() {
+		if i.cmp(key, i.key) <= 0 {
+			break
+		}
+	}
+	return i.Valid()
+}
+
+// SeekPrefixGE implements internalIterator.SeekPrefixGE, as documented in the
+// pebble package.
+func (i *rawBlockIter) SeekPrefixGE(key []byte) bool {
+	// This should never be called as prefix iteration is never used with raw blocks.
+	panic("pebble: SeekPrefixGE unimplemented")
+}
+
+// SeekLT implements internalIterator.SeekLT, as documented in the pebble
+// package.
+func (i *rawBlockIter) SeekLT(key []byte) bool {
+	panic("pebble/table: SeekLT unimplemented")
+}
+
+// First implements internalIterator.First, as documented in the pebble
+// package.
+func (i *rawBlockIter) First() bool {
+	i.offset = 0
+	i.loadEntry()
+	return i.Valid()
+}
+
+// Last implements internalIterator.Last, as documented in the pebble package.
+func (i *rawBlockIter) Last() bool {
+	// Seek forward from the last restart point.
+	i.offset = int32(binary.LittleEndian.Uint32(i.data[i.restarts+4*(i.numRestarts-1):]))
+
+	i.readEntry()
+	i.clearCache()
+	i.cacheEntry()
+
+	for i.nextOffset < i.restarts {
+		i.offset = i.nextOffset
+		i.readEntry()
+		i.cacheEntry()
+	}
+
+	i.ikey.UserKey = i.key
+	return i.Valid()
+}
+
+// Next implements internalIterator.Next, as documented in the pebble
+// package.
+func (i *rawBlockIter) Next() bool {
+	i.offset = i.nextOffset
+	if !i.Valid() {
+		return false
+	}
+	i.loadEntry()
+	return true
+}
+
+// Prev implements internalIterator.Prev, as documented in the pebble
+// package.
+func (i *rawBlockIter) Prev() bool {
+	if n := len(i.cached) - 1; n > 0 && i.cached[n].offset == i.offset {
+		i.nextOffset = i.offset
+		e := &i.cached[n-1]
+		i.offset = e.offset
+		i.val = getBytes(unsafe.Pointer(uintptr(i.ptr)+uintptr(e.valStart)), int(e.valSize))
+		i.ikey.UserKey = i.cachedBuf[e.keyStart:e.keyEnd]
+		i.cached = i.cached[:n]
+		return true
+	}
+
+	if i.offset == 0 {
+		i.offset = -1
+		i.nextOffset = 0
+		return false
+	}
+
+	targetOffset := i.offset
+	index := sort.Search(int(i.numRestarts), func(j int) bool {
+		offset := int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*j:]))
+		return offset >= targetOffset
+	})
+	i.offset = 0
+	if index > 0 {
+		i.offset = int32(binary.LittleEndian.Uint32(i.data[int(i.restarts)+4*(index-1):]))
+	}
+
+	i.readEntry()
+	i.clearCache()
+	i.cacheEntry()
+
+	for i.nextOffset < targetOffset {
+		i.offset = i.nextOffset
+		i.readEntry()
+		i.cacheEntry()
+	}
+
+	i.ikey.UserKey = i.key
+	return true
+}
+
+// Key implements internalIterator.Key, as documented in the pebble package.
+func (i *rawBlockIter) Key() InternalKey {
+	return i.ikey
+}
+
+// Value implements internalIterator.Value, as documented in the pebble
+// package.
+func (i *rawBlockIter) Value() []byte {
+	return i.val
+}
+
+// Valid implements internalIterator.Valid, as documented in the pebble
+// package.
+func (i *rawBlockIter) Valid() bool {
+	return i.offset >= 0 && i.offset < i.restarts
+}
+
+// Error implements internalIterator.Error, as documented in the pebble
+// package.
+func (i *rawBlockIter) Error() error {
+	return nil
+}
+
+// Close implements internalIterator.Close, as documented in the pebble
+// package.
+func (i *rawBlockIter) Close() error {
+	i.val = nil
+	return nil
+}

vendor/github.com/cockroachdb/pebble/sstable/suffix_rewriter.go

@@ -0,0 +1,480 @@
+package sstable
+
+import (
+	"bytes"
+	"math"
+	"os"
+	"sync"
+	"time"
+
+	"github.com/cespare/xxhash/v2"
+	"github.com/cockroachdb/errors"
+	"github.com/cockroachdb/pebble/internal/base"
+	"github.com/cockroachdb/pebble/internal/rangekey"
+)
+
+// RewriteKeySuffixes copies the content of the passed SSTable bytes to a new
+// sstable, written to `out`, in which the suffix `from` has is replaced with
+// `to` in every key. The input sstable must consist of only Sets or RangeKeySets
+// and every key must have `from` as its suffix as determined by the Split
+// function of the Comparer in the passed WriterOptions. Range deletes must not
+// exist in this sstable, as they will be ignored.
+//
+// Data blocks are rewritten in parallel by `concurrency` workers and then
+// assembled into a final SST. Filters are copied from the original SST without
+// modification as they are not affected by the suffix, while block and table
+// properties are only minimally recomputed.
+//
+// Any block and table property collectors configured in the WriterOptions must
+// implement SuffixReplaceableTableCollector/SuffixReplaceableBlockCollector.
+func RewriteKeySuffixes(
+	sst []byte,
+	rOpts ReaderOptions,
+	out writeCloseSyncer,
+	o WriterOptions,
+	from, to []byte,
+	concurrency int,
+) (*WriterMetadata, error) {
+	r, err := NewMemReader(sst, rOpts)
+	if err != nil {
+		return nil, err
+	}
+	defer r.Close()
+	return rewriteKeySuffixesInBlocks(r, out, o, from, to, concurrency)
+}
+
+func rewriteKeySuffixesInBlocks(
+	r *Reader, out writeCloseSyncer, o WriterOptions, from, to []byte, concurrency int,
+) (*WriterMetadata, error) {
+	if o.Comparer == nil || o.Comparer.Split == nil {
+		return nil, errors.New("a valid splitter is required to define suffix to replace replace suffix")
+	}
+	if concurrency < 1 {
+		return nil, errors.New("concurrency must be >= 1")
+	}
+
+	w := NewWriter(out, o)
+	defer w.Close()
+
+	for _, c := range w.propCollectors {
+		if _, ok := c.(SuffixReplaceableTableCollector); !ok {
+			return nil, errors.Errorf("property collector %s does not support suffix replacement", c.Name())
+		}
+	}
+	for _, c := range w.blockPropCollectors {
+		if _, ok := c.(SuffixReplaceableBlockCollector); !ok {
+			return nil, errors.Errorf("block property collector %s does not support suffix replacement", c.Name())
+		}
+	}
+
+	l, err := r.Layout()
+	if err != nil {
+		return nil, errors.Wrap(err, "reading layout")
+	}
+
+	if err := rewriteDataBlocksToWriter(r, w, l.Data, from, to, w.split, concurrency); err != nil {
+		return nil, errors.Wrap(err, "rewriting data blocks")
+	}
+
+	// Copy over the range key block and replace suffixes in it if it exists.
+	if err := rewriteRangeKeyBlockToWriter(r, w, from, to); err != nil {
+		return nil, errors.Wrap(err, "rewriting range key blocks")
+	}
+
+	// Copy over the filter block if it exists (rewriteDataBlocksToWriter will
+	// already have ensured this is valid if it exists).
+	if w.filter != nil && l.Filter.Length > 0 {
+		filterBlock, _, err := readBlockBuf(r, l.Filter, nil)
+		if err != nil {
+			return nil, errors.Wrap(err, "reading filter")
+		}
+		w.filter = copyFilterWriter{
+			origPolicyName: w.filter.policyName(), origMetaName: w.filter.metaName(), data: filterBlock,
+		}
+	}
+
+	if err := w.Close(); err != nil {
+		return nil, err
+	}
+
+	return w.Metadata()
+}
+
+var errBadKind = errors.New("key does not have expected kind (set)")
+
+type blockWithSpan struct {
+	start, end InternalKey
+	data       []byte
+}
+
+func rewriteBlocks(
+	r *Reader,
+	restartInterval int,
+	checksumType ChecksumType,
+	compression Compression,
+	input []BlockHandleWithProperties,
+	output []blockWithSpan,
+	totalWorkers, worker int,
+	from, to []byte,
+	split Split,
+) error {
+	bw := blockWriter{
+		restartInterval: restartInterval,
+	}
+	buf := blockBuf{checksummer: checksummer{checksumType: checksumType}}
+	if checksumType == ChecksumTypeXXHash {
+		buf.checksummer.xxHasher = xxhash.New()
+	}
+
+	var blockAlloc []byte
+	var keyAlloc []byte
+	var scratch InternalKey
+
+	var inputBlock, inputBlockBuf []byte
+
+	iter := &blockIter{}
+
+	// We'll assume all blocks are _roughly_ equal so round-robin static partition
+	// of each worker doing every ith block is probably enough.
+	for i := worker; i < len(input); i += totalWorkers {
+		bh := input[i]
+
+		var err error
+		inputBlock, inputBlockBuf, err = readBlockBuf(r, bh.BlockHandle, inputBlockBuf)
+		if err != nil {
+			return err
+		}
+		if err := iter.init(r.Compare, inputBlock, r.Properties.GlobalSeqNum); err != nil {
+			return err
+		}
+
+		if cap(bw.restarts) < int(iter.restarts) {
+			bw.restarts = make([]uint32, 0, iter.restarts)
+		}
+		if cap(bw.buf) == 0 {
+			bw.buf = make([]byte, 0, len(inputBlock))
+		}
+		if cap(bw.restarts) < int(iter.numRestarts) {
+			bw.restarts = make([]uint32, 0, iter.numRestarts)
+		}
+
+		for key, val := iter.First(); key != nil; key, val = iter.Next() {
+			if key.Kind() != InternalKeyKindSet {
+				return errBadKind
+			}
+			si := split(key.UserKey)
+			oldSuffix := key.UserKey[si:]
+			if !bytes.Equal(oldSuffix, from) {
+				err := errors.Errorf("key has suffix %q, expected %q", oldSuffix, from)
+				return err
+			}
+			newLen := si + len(to)
+			if cap(scratch.UserKey) < newLen {
+				scratch.UserKey = make([]byte, 0, len(key.UserKey)*2+len(to)-len(from))
+			}
+
+			scratch.Trailer = key.Trailer
+			scratch.UserKey = scratch.UserKey[:newLen]
+			copy(scratch.UserKey, key.UserKey[:si])
+			copy(scratch.UserKey[si:], to)
+
+			bw.add(scratch, val)
+			if output[i].start.UserKey == nil {
+				keyAlloc, output[i].start = cloneKeyWithBuf(scratch, keyAlloc)
+			}
+		}
+		*iter = iter.resetForReuse()
+
+		keyAlloc, output[i].end = cloneKeyWithBuf(scratch, keyAlloc)
+
+		finished := compressAndChecksum(bw.finish(), compression, &buf)
+
+		// copy our finished block into the output buffer.
+		sz := len(finished) + blockTrailerLen
+		if cap(blockAlloc) < sz {
+			blockAlloc = make([]byte, sz*128)
+		}
+		output[i].data = blockAlloc[:sz:sz]
+		blockAlloc = blockAlloc[sz:]
+		copy(output[i].data, finished)
+		copy(output[i].data[len(finished):], buf.tmp[:blockTrailerLen])
+	}
+	return nil
+}
+
+func rewriteDataBlocksToWriter(
+	r *Reader,
+	w *Writer,
+	data []BlockHandleWithProperties,
+	from, to []byte,
+	split Split,
+	concurrency int,
+) error {
+	if r.Properties.NumEntries == 0 {
+		// No point keys.
+		return nil
+	}
+	blocks := make([]blockWithSpan, len(data))
+
+	if w.filter != nil {
+		if r.Properties.FilterPolicyName != w.filter.policyName() {
+			return errors.New("mismatched filters")
+		}
+		if was, is := r.Properties.ComparerName, w.props.ComparerName; was != is {
+			return errors.Errorf("mismatched Comparer %s vs %s, replacement requires same splitter to copy filters", was, is)
+		}
+	}
+
+	g := &sync.WaitGroup{}
+	g.Add(concurrency)
+	errCh := make(chan error, concurrency)
+	for i := 0; i < concurrency; i++ {
+		worker := i
+		go func() {
+			defer g.Done()
+			err := rewriteBlocks(
+				r,
+				w.dataBlockBuf.dataBlock.restartInterval,
+				w.blockBuf.checksummer.checksumType,
+				w.compression,
+				data,
+				blocks,
+				concurrency,
+				worker,
+				from, to,
+				split,
+			)
+			if err != nil {
+				errCh <- err
+			}
+		}()
+	}
+	g.Wait()
+	close(errCh)
+	if err, ok := <-errCh; ok {
+		return err
+	}
+
+	for _, p := range w.propCollectors {
+		if err := p.(SuffixReplaceableTableCollector).UpdateKeySuffixes(r.Properties.UserProperties, from, to); err != nil {
+			return err
+		}
+	}
+
+	var decoder blockPropertiesDecoder
+	var oldShortIDs []shortID
+	var oldProps [][]byte
+	if len(w.blockPropCollectors) > 0 {
+		oldProps = make([][]byte, len(w.blockPropCollectors))
+		oldShortIDs = make([]shortID, math.MaxUint8)
+		for i, p := range w.blockPropCollectors {
+			if prop, ok := r.Properties.UserProperties[p.Name()]; ok {
+				was, is := shortID(byte(prop[0])), shortID(i)
+				oldShortIDs[was] = is
+			}
+		}
+	}
+
+	for i := range blocks {
+		// Write the rewritten block to the file.
+		n, err := w.writer.Write(blocks[i].data)
+		if err != nil {
+			return err
+		}
+
+		bh := BlockHandle{Offset: w.meta.Size, Length: uint64(n) - blockTrailerLen}
+		// Update the overall size.
+		w.meta.Size += uint64(n)
+
+		// Load any previous values for our prop collectors into oldProps.
+		for i := range oldProps {
+			oldProps[i] = nil
+		}
+		decoder.props = data[i].Props
+		for !decoder.done() {
+			id, val, err := decoder.next()
+			if err != nil {
+				return err
+			}
+			oldProps[oldShortIDs[id]] = val
+		}
+
+		for i, p := range w.blockPropCollectors {
+			if err := p.(SuffixReplaceableBlockCollector).UpdateKeySuffixes(oldProps[i], from, to); err != nil {
+				return err
+			}
+		}
+
+		var bhp BlockHandleWithProperties
+		if bhp, err = w.maybeAddBlockPropertiesToBlockHandle(bh); err != nil {
+			return err
+		}
+		var nextKey InternalKey
+		if i+1 < len(blocks) {
+			nextKey = blocks[i+1].start
+		}
+		if err = w.addIndexEntrySync(blocks[i].end, nextKey, bhp, w.dataBlockBuf.tmp[:]); err != nil {
+			return err
+		}
+	}
+
+	w.meta.updateSeqNum(blocks[0].start.SeqNum())
+	w.props.NumEntries = r.Properties.NumEntries
+	w.props.RawKeySize = r.Properties.RawKeySize
+	w.props.RawValueSize = r.Properties.RawValueSize
+	w.meta.SetSmallestPointKey(blocks[0].start)
+	w.meta.SetLargestPointKey(blocks[len(blocks)-1].end)
+	return nil
+}
+
+func rewriteRangeKeyBlockToWriter(r *Reader, w *Writer, from, to []byte) error {
+	iter, err := r.NewRawRangeKeyIter()
+	if err != nil {
+		return err
+	}
+	if iter == nil {
+		// No range keys.
+		return nil
+	}
+	defer iter.Close()
+
+	for s := iter.First(); s != nil; s = iter.Next() {
+		if !s.Valid() {
+			break
+		}
+		for i := range s.Keys {
+			if s.Keys[i].Kind() != base.InternalKeyKindRangeKeySet {
+				return errBadKind
+			}
+			if !bytes.Equal(s.Keys[i].Suffix, from) {
+				return errors.Errorf("key has suffix %q, expected %q", s.Keys[i].Suffix, from)
+			}
+			s.Keys[i].Suffix = to
+		}
+
+		err := rangekey.Encode(s, func(k base.InternalKey, v []byte) error {
+			// Calling AddRangeKey instead of addRangeKeySpan bypasses the fragmenter.
+			// This is okay because the raw fragments off of `iter` are already
+			// fragmented, and suffix replacement should not affect fragmentation.
+			return w.AddRangeKey(k, v)
+		})
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+type copyFilterWriter struct {
+	origMetaName   string
+	origPolicyName string
+	data           []byte
+}
+
+func (copyFilterWriter) addKey(key []byte)         { panic("unimplemented") }
+func (c copyFilterWriter) finish() ([]byte, error) { return c.data, nil }
+func (c copyFilterWriter) metaName() string        { return c.origMetaName }
+func (c copyFilterWriter) policyName() string      { return c.origPolicyName }
+
+// RewriteKeySuffixesViaWriter is similar to RewriteKeySuffixes but uses just a
+// single loop over the Reader that writes each key to the Writer with the new
+// suffix. The is significantly slower than the parallelized rewriter, and does
+// more work to rederive filters, props, etc, however re-doing that work makes
+// it less restrictive -- props no longer need to
+func RewriteKeySuffixesViaWriter(
+	r *Reader, out writeCloseSyncer, o WriterOptions, from, to []byte,
+) (*WriterMetadata, error) {
+	if o.Comparer == nil || o.Comparer.Split == nil {
+		return nil, errors.New("a valid splitter is required to define suffix to replace replace suffix")
+	}
+
+	w := NewWriter(out, o)
+	i, err := r.NewIter(nil, nil)
+	if err != nil {
+		return nil, err
+	}
+	defer i.Close()
+
+	k, v := i.First()
+	var scratch InternalKey
+	for k != nil {
+		if k.Kind() != InternalKeyKindSet {
+			return nil, errors.New("invalid key type")
+		}
+		oldSuffix := k.UserKey[r.Split(k.UserKey):]
+		if !bytes.Equal(oldSuffix, from) {
+			return nil, errors.Errorf("key has suffix %q, expected %q", oldSuffix, from)
+		}
+		scratch.UserKey = append(scratch.UserKey[:0], k.UserKey[:len(k.UserKey)-len(from)]...)
+		scratch.UserKey = append(scratch.UserKey, to...)
+		scratch.Trailer = k.Trailer
+
+		if w.addPoint(scratch, v); err != nil {
+			return nil, err
+		}
+		k, v = i.Next()
+	}
+	if err := rewriteRangeKeyBlockToWriter(r, w, from, to); err != nil {
+		return nil, err
+	}
+	if err := w.Close(); err != nil {
+		return nil, err
+	}
+	return &w.meta, nil
+}
+
+// NewMemReader opens a reader over the SST stored in the passed []byte.
+func NewMemReader(sst []byte, o ReaderOptions) (*Reader, error) {
+	return NewReader(memReader{sst, bytes.NewReader(sst), sizeOnlyStat(int64(len(sst)))}, o)
+}
+
+func readBlockBuf(r *Reader, bh BlockHandle, buf []byte) ([]byte, []byte, error) {
+	raw := r.file.(memReader).b[bh.Offset : bh.Offset+bh.Length+blockTrailerLen]
+	if err := checkChecksum(r.checksumType, raw, bh, 0); err != nil {
+		return nil, buf, err
+	}
+	typ := blockType(raw[bh.Length])
+	raw = raw[:bh.Length]
+	if typ == noCompressionBlockType {
+		return raw, buf, nil
+	}
+	decompressedLen, prefix, err := decompressedLen(typ, raw)
+	if err != nil {
+		return nil, buf, err
+	}
+	if cap(buf) < decompressedLen {
+		buf = make([]byte, decompressedLen)
+	}
+	res, err := decompressInto(typ, raw[prefix:], buf[:decompressedLen])
+	return res, buf, err
+}
+
+// memReader is a thin wrapper around a []byte such that it can be passed to an
+// sstable.Reader. It supports concurrent use, and does so without locking in
+// contrast to the heavier read/write vfs.MemFile.
+type memReader struct {
+	b []byte
+	r *bytes.Reader
+	s sizeOnlyStat
+}
+
+var _ ReadableFile = memReader{}
+
+// ReadAt implements io.ReaderAt.
+func (m memReader) ReadAt(p []byte, off int64) (n int, err error) { return m.r.ReadAt(p, off) }
+
+// Close implements io.Closer.
+func (memReader) Close() error { return nil }
+
+// Stat implements ReadableFile.
+func (m memReader) Stat() (os.FileInfo, error) { return m.s, nil }
+
+type sizeOnlyStat int64
+
+func (s sizeOnlyStat) Size() int64      { return int64(s) }
+func (sizeOnlyStat) IsDir() bool        { panic(errors.AssertionFailedf("unimplemented")) }
+func (sizeOnlyStat) ModTime() time.Time { panic(errors.AssertionFailedf("unimplemented")) }
+func (sizeOnlyStat) Mode() os.FileMode  { panic(errors.AssertionFailedf("unimplemented")) }
+func (sizeOnlyStat) Name() string       { panic(errors.AssertionFailedf("unimplemented")) }
+func (sizeOnlyStat) Sys() interface{}   { panic(errors.AssertionFailedf("unimplemented")) }

vendor/github.com/cockroachdb/pebble/sstable/table.go

@@ -0,0 +1,428 @@
+// Copyright 2011 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 sstable implements readers and writers of pebble tables.
+//
+// Tables are either opened for reading or created for writing but not both.
+//
+// A reader can create iterators, which allow seeking and next/prev
+// iteration. There may be multiple key/value pairs that have the same key and
+// different sequence numbers.
+//
+// A reader can be used concurrently. Multiple goroutines can call NewIter
+// concurrently, and each iterator can run concurrently with other iterators.
+// However, any particular iterator should not be used concurrently, and iterators
+// should not be used once a reader is closed.
+//
+// A writer writes key/value pairs in increasing key order, and cannot be used
+// concurrently. A table cannot be read until the writer has finished.
+//
+// Readers and writers can be created with various options. Passing a nil
+// Options pointer is valid and means to use the default values.
+//
+// One such option is to define the 'less than' ordering for keys. The default
+// Comparer uses the natural ordering consistent with bytes.Compare. The same
+// ordering should be used for reading and writing a table.
+//
+// To return the value for a key:
+//
+// 	r := table.NewReader(file, options)
+// 	defer r.Close()
+// 	i := r.NewIter(nil, nil)
+// 	defer i.Close()
+// 	ikey, value := r.SeekGE(key)
+// 	if options.Comparer.Compare(ikey.UserKey, key) != 0 {
+// 	  // not found
+// 	} else {
+// 	  // value is the first record containing key
+// 	}
+//
+// To count the number of entries in a table:
+//
+// 	i, n := r.NewIter(nil, nil), 0
+// 	for key, value := i.First(); key != nil; key, value = i.Next() {
+// 		n++
+// 	}
+// 	if err := i.Close(); err != nil {
+// 		return 0, err
+// 	}
+// 	return n, nil
+//
+// To write a table with three entries:
+//
+// 	w := table.NewWriter(file, options)
+// 	if err := w.Set([]byte("apple"), []byte("red")); err != nil {
+// 		w.Close()
+// 		return err
+// 	}
+// 	if err := w.Set([]byte("banana"), []byte("yellow")); err != nil {
+// 		w.Close()
+// 		return err
+// 	}
+// 	if err := w.Set([]byte("cherry"), []byte("red")); err != nil {
+// 		w.Close()
+// 		return err
+// 	}
+// 	return w.Close()
+package sstable // import "github.com/cockroachdb/pebble/sstable"
+
+import (
+	"encoding/binary"
+	"io"
+
+	"github.com/cockroachdb/errors"
+	"github.com/cockroachdb/pebble/internal/base"
+)
+
+/*
+The table file format looks like:
+
+<start_of_file>
+[data block 0]
+[data block 1]
+...
+[data block N-1]
+[meta filter block] (optional)
+[index block] (for single level index)
+[meta rangedel block] (optional)
+[meta range key block] (optional)
+[meta properties block]
+[metaindex block]
+[footer]
+<end_of_file>
+
+A Reader eagerly loads the footer, metaindex block and meta properties block,
+because the data contained in those blocks is needed on every read, and even
+before reading. For example, the meta properties block is used to verify the
+comparer and merger are compatible, and the metaindex block contains the
+location of the meta properties (and other meta blocks). In situations where
+file system locality matters, or one wants to minimize number of read
+requests when eagerly loading these blocks, having these three as a suffix
+of the file is convenient.
+
+The interleaving of the index block(s) between the meta blocks is done to
+match RocksDB/LevelDB behavior.
+
+Each block consists of some data and a 5 byte trailer: a 1 byte block type and a
+4 byte checksum. The checksum is computed over the compressed data and the first
+byte of the trailer (i.e. the block type), and is serialized as little-endian.
+The block type gives the per-block compression used; each block is compressed
+independently. The checksum algorithm is described in the pebble/crc package.
+
+Most blocks, other than the meta filter block, contain key/value pairs. The
+remainder of this comment refers to the decompressed block, which has its 5 byte
+trailer stripped. The decompressed block data consists of a sequence of such
+key/value entries followed by a block suffix. Each key is encoded as a shared
+prefix length and a remainder string. For example, if two adjacent keys are
+"tweedledee" and "tweedledum", then the second key would be encoded as {8,
+"um"}. The shared prefix length is varint encoded. The remainder string and the
+value are encoded as a varint-encoded length followed by the literal contents.
+To continue the example, suppose that the key "tweedledum" mapped to the value
+"socks". The encoded key/value entry would be: "\x08\x02\x05umsocks".
+
+Every block has a restart interval I. Every I'th key/value entry in that block
+is called a restart point, and shares no key prefix with the previous entry.
+Continuing the example above, if the key after "tweedledum" was "two", but was
+part of a restart point, then that key would be encoded as {0, "two"} instead
+of {2, "o"}. If a block has P restart points, then the block suffix consists
+of (P+1)*4 bytes: (P+1) little-endian uint32 values. The first P of these
+uint32 values are the block offsets of each restart point. The final uint32
+value is P itself. Thus, when seeking for a particular key, one can use binary
+search to find the largest restart point whose key is <= the key sought.
+
+An index block is a block with N key/value entries. The i'th value is the
+encoded block handle of the i'th data block. The i'th key is a separator for
+i < N-1, and a successor for i == N-1. The separator between blocks i and i+1
+is a key that is >= every key in block i and is < every key i block i+1. The
+successor for the final block is a key that is >= every key in block N-1. The
+index block restart interval is 1: every entry is a restart point.
+
+A block handle is an offset, a length, and optional block properties (for data
+blocks and first/lower level index blocks); the length does not include the 5
+byte trailer. All numbers are varint-encoded, with no padding between the two
+values. The maximum size of an encoded block handle without properties is 20
+bytes. It is not advised to have properties that accumulate to be longer than
+100 bytes.
+
+Instead of a single index block, the sstable can have a two-level index (this
+is used to prevent a single huge index block). A two-level index consists of a
+sequence of lower-level index blocks with block handles for data blocks
+followed by a single top-level index block with block handles for the
+lower-level index blocks.
+
+The metaindex block also contains block handles as values, with keys being
+the names of the meta blocks.
+
+*/
+
+const (
+	blockTrailerLen                    = 5
+	blockHandleMaxLenWithoutProperties = 10 + 10
+	// blockHandleLikelyMaxLen can be used for pre-allocating buffers to
+	// reduce memory copies. It is not guaranteed that a block handle will not
+	// exceed this length.
+	blockHandleLikelyMaxLen = blockHandleMaxLenWithoutProperties + 100
+
+	levelDBFooterLen   = 48
+	levelDBMagic       = "\x57\xfb\x80\x8b\x24\x75\x47\xdb"
+	levelDBMagicOffset = levelDBFooterLen - len(levelDBMagic)
+
+	rocksDBFooterLen             = 1 + 2*blockHandleMaxLenWithoutProperties + 4 + 8
+	rocksDBMagic                 = "\xf7\xcf\xf4\x85\xb7\x41\xe2\x88"
+	rocksDBMagicOffset           = rocksDBFooterLen - len(rocksDBMagic)
+	rocksDBVersionOffset         = rocksDBMagicOffset - 4
+	rocksDBExternalFormatVersion = 2
+
+	pebbleDBMagic = "\xf0\x9f\xaa\xb3\xf0\x9f\xaa\xb3" // 🪳🪳
+
+	minFooterLen = levelDBFooterLen
+	maxFooterLen = rocksDBFooterLen
+
+	levelDBFormatVersion  = 0
+	rocksDBFormatVersion2 = 2
+
+	metaRangeKeyName   = "pebble.range_key"
+	metaPropertiesName = "rocksdb.properties"
+	metaRangeDelName   = "rocksdb.range_del"
+	metaRangeDelV2Name = "rocksdb.range_del2"
+
+	// Index Types.
+	// A space efficient index block that is optimized for binary-search-based
+	// index.
+	binarySearchIndex = 0
+	// hashSearchIndex               = 1
+	// A two-level index implementation. Both levels are binary search indexes.
+	twoLevelIndex = 2
+	// binarySearchWithFirstKeyIndex = 3
+
+	// RocksDB always includes this in the properties block. Since Pebble
+	// doesn't use zstd compression, the string will always be the same.
+	// This should be removed if we ever decide to diverge from the RocksDB
+	// properties block.
+	rocksDBCompressionOptions = "window_bits=-14; level=32767; strategy=0; max_dict_bytes=0; zstd_max_train_bytes=0; enabled=0; "
+)
+
+// ChecksumType specifies the checksum used for blocks.
+type ChecksumType byte
+
+// The available checksum types.
+const (
+	ChecksumTypeNone     ChecksumType = 0
+	ChecksumTypeCRC32c   ChecksumType = 1
+	ChecksumTypeXXHash   ChecksumType = 2
+	ChecksumTypeXXHash64 ChecksumType = 3
+)
+
+// String implements fmt.Stringer.
+func (t ChecksumType) String() string {
+	switch t {
+	case ChecksumTypeCRC32c:
+		return "crc32c"
+	case ChecksumTypeNone:
+		return "none"
+	case ChecksumTypeXXHash:
+		return "xxhash"
+	case ChecksumTypeXXHash64:
+		return "xxhash64"
+	default:
+		panic(errors.Newf("sstable: unknown checksum type: %d", t))
+	}
+}
+
+type blockType byte
+
+const (
+	// The block type gives the per-block compression format.
+	// These constants are part of the file format and should not be changed.
+	// They are different from the Compression constants because the latter
+	// are designed so that the zero value of the Compression type means to
+	// use the default compression (which is snappy).
+	// Not all compression types listed here are supported.
+	noCompressionBlockType     blockType = 0
+	snappyCompressionBlockType blockType = 1
+	zlibCompressionBlockType   blockType = 2
+	bzip2CompressionBlockType  blockType = 3
+	lz4CompressionBlockType    blockType = 4
+	lz4hcCompressionBlockType  blockType = 5
+	xpressCompressionBlockType blockType = 6
+	zstdCompressionBlockType   blockType = 7
+)
+
+// String implements fmt.Stringer.
+func (t blockType) String() string {
+	switch t {
+	case 0:
+		return "none"
+	case 1:
+		return "snappy"
+	case 2:
+		return "zlib"
+	case 3:
+		return "bzip2"
+	case 4:
+		return "lz4"
+	case 5:
+		return "lz4hc"
+	case 6:
+		return "xpress"
+	case 7:
+		return "zstd"
+	default:
+		panic(errors.Newf("sstable: unknown block type: %d", t))
+	}
+}
+
+// legacy (LevelDB) footer format:
+//    metaindex handle (varint64 offset, varint64 size)
+//    index handle     (varint64 offset, varint64 size)
+//    <padding> to make the total size 2 * BlockHandle::kMaxEncodedLength
+//    table_magic_number (8 bytes)
+// new (RocksDB) footer format:
+//    checksum type (char, 1 byte)
+//    metaindex handle (varint64 offset, varint64 size)
+//    index handle     (varint64 offset, varint64 size)
+//    <padding> to make the total size 2 * BlockHandle::kMaxEncodedLength + 1
+//    footer version (4 bytes)
+//    table_magic_number (8 bytes)
+type footer struct {
+	format      TableFormat
+	checksum    ChecksumType
+	metaindexBH BlockHandle
+	indexBH     BlockHandle
+	footerBH    BlockHandle
+}
+
+func readFooter(f ReadableFile) (footer, error) {
+	var footer footer
+	stat, err := f.Stat()
+	if err != nil {
+		return footer, errors.Wrap(err, "pebble/table: invalid table (could not stat file)")
+	}
+	if stat.Size() < minFooterLen {
+		return footer, base.CorruptionErrorf("pebble/table: invalid table (file size is too small)")
+	}
+
+	buf := make([]byte, maxFooterLen)
+	off := stat.Size() - maxFooterLen
+	if off < 0 {
+		off = 0
+	}
+	n, err := f.ReadAt(buf, off)
+	if err != nil && err != io.EOF {
+		return footer, errors.Wrap(err, "pebble/table: invalid table (could not read footer)")
+	}
+	buf = buf[:n]
+
+	switch magic := buf[len(buf)-len(rocksDBMagic):]; string(magic) {
+	case levelDBMagic:
+		if len(buf) < levelDBFooterLen {
+			return footer, base.CorruptionErrorf(
+				"pebble/table: invalid table (footer too short): %d", errors.Safe(len(buf)))
+		}
+		footer.footerBH.Offset = uint64(off+int64(len(buf))) - levelDBFooterLen
+		buf = buf[len(buf)-levelDBFooterLen:]
+		footer.footerBH.Length = uint64(len(buf))
+		footer.format = TableFormatLevelDB
+		footer.checksum = ChecksumTypeCRC32c
+
+	case rocksDBMagic, pebbleDBMagic:
+		// NOTE: The Pebble magic string implies the same footer format as that used
+		// by the RocksDBv2 table format.
+		if len(buf) < rocksDBFooterLen {
+			return footer, base.CorruptionErrorf("pebble/table: invalid table (footer too short): %d", errors.Safe(len(buf)))
+		}
+		footer.footerBH.Offset = uint64(off+int64(len(buf))) - rocksDBFooterLen
+		buf = buf[len(buf)-rocksDBFooterLen:]
+		footer.footerBH.Length = uint64(len(buf))
+		version := binary.LittleEndian.Uint32(buf[rocksDBVersionOffset:rocksDBMagicOffset])
+
+		format, err := ParseTableFormat(magic, version)
+		if err != nil {
+			return footer, err
+		}
+		footer.format = format
+
+		switch ChecksumType(buf[0]) {
+		case ChecksumTypeCRC32c:
+			footer.checksum = ChecksumTypeCRC32c
+		case ChecksumTypeXXHash64:
+			footer.checksum = ChecksumTypeXXHash64
+		default:
+			return footer, base.CorruptionErrorf("pebble/table: unsupported checksum type %d", errors.Safe(footer.checksum))
+		}
+		buf = buf[1:]
+
+	default:
+		return footer, base.CorruptionErrorf("pebble/table: invalid table (bad magic number)")
+	}
+
+	{
+		end := uint64(stat.Size())
+		var n int
+		footer.metaindexBH, n = decodeBlockHandle(buf)
+		if n == 0 || footer.metaindexBH.Offset+footer.metaindexBH.Length > end {
+			return footer, base.CorruptionErrorf("pebble/table: invalid table (bad metaindex block handle)")
+		}
+		buf = buf[n:]
+
+		footer.indexBH, n = decodeBlockHandle(buf)
+		if n == 0 || footer.indexBH.Offset+footer.indexBH.Length > end {
+			return footer, base.CorruptionErrorf("pebble/table: invalid table (bad index block handle)")
+		}
+	}
+
+	return footer, nil
+}
+
+func (f footer) encode(buf []byte) []byte {
+	switch magic, version := f.format.AsTuple(); magic {
+	case levelDBMagic:
+		buf = buf[:levelDBFooterLen]
+		for i := range buf {
+			buf[i] = 0
+		}
+		n := encodeBlockHandle(buf[0:], f.metaindexBH)
+		encodeBlockHandle(buf[n:], f.indexBH)
+		copy(buf[len(buf)-len(levelDBMagic):], levelDBMagic)
+
+	case rocksDBMagic, pebbleDBMagic:
+		buf = buf[:rocksDBFooterLen]
+		for i := range buf {
+			buf[i] = 0
+		}
+		switch f.checksum {
+		case ChecksumTypeNone:
+			buf[0] = byte(ChecksumTypeNone)
+		case ChecksumTypeCRC32c:
+			buf[0] = byte(ChecksumTypeCRC32c)
+		case ChecksumTypeXXHash:
+			buf[0] = byte(ChecksumTypeXXHash)
+		case ChecksumTypeXXHash64:
+			buf[0] = byte(ChecksumTypeXXHash64)
+		default:
+			panic("unknown checksum type")
+		}
+		n := 1
+		n += encodeBlockHandle(buf[n:], f.metaindexBH)
+		encodeBlockHandle(buf[n:], f.indexBH)
+		binary.LittleEndian.PutUint32(buf[rocksDBVersionOffset:], version)
+		copy(buf[len(buf)-len(rocksDBMagic):], magic)
+
+	default:
+		panic("sstable: unspecified table format version")
+	}
+
+	return buf
+}
+
+func supportsTwoLevelIndex(format TableFormat) bool {
+	switch format {
+	case TableFormatLevelDB:
+		return false
+	case TableFormatRocksDBv2, TableFormatPebblev1, TableFormatPebblev2:
+		return true
+	default:
+		panic("sstable: unspecified table format version")
+	}
+}

vendor/github.com/cockroachdb/pebble/sstable/unsafe.go

@@ -0,0 +1,35 @@
+// 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 sstable
+
+import (
+	"unsafe"
+
+	"github.com/cockroachdb/pebble/internal/manual"
+)
+
+func getBytes(ptr unsafe.Pointer, length int) []byte {
+	return (*[manual.MaxArrayLen]byte)(ptr)[:length:length]
+}
+
+func decodeVarint(ptr unsafe.Pointer) (uint32, unsafe.Pointer) {
+	if a := *((*uint8)(ptr)); a < 128 {
+		return uint32(a),
+			unsafe.Pointer(uintptr(ptr) + 1)
+	} else if a, b := a&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 1))); b < 128 {
+		return uint32(b)<<7 | uint32(a),
+			unsafe.Pointer(uintptr(ptr) + 2)
+	} else if b, c := b&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 2))); c < 128 {
+		return uint32(c)<<14 | uint32(b)<<7 | uint32(a),
+			unsafe.Pointer(uintptr(ptr) + 3)
+	} else if c, d := c&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 3))); d < 128 {
+		return uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a),
+			unsafe.Pointer(uintptr(ptr) + 4)
+	} else {
+		d, e := d&0x7f, *((*uint8)(unsafe.Pointer(uintptr(ptr) + 4)))
+		return uint32(e)<<28 | uint32(d)<<21 | uint32(c)<<14 | uint32(b)<<7 | uint32(a),
+			unsafe.Pointer(uintptr(ptr) + 5)
+	}
+}

vendor/github.com/cockroachdb/pebble/sstable/write_queue.go

@@ -0,0 +1,147 @@
+package sstable
+
+import (
+	"sync"
+
+	"github.com/cockroachdb/pebble/internal/base"
+)
+
+type writeTask struct {
+	// Since writeTasks are pooled, the compressionDone channel will be re-used.
+	// It is necessary that any writes to the channel have already been read,
+	// before adding the writeTask back to the pool.
+	compressionDone chan bool
+	buf             *dataBlockBuf
+	// If this is not nil, then this index block will be flushed.
+	flushableIndexBlock *indexBlockBuf
+	// currIndexBlock is the index block on which indexBlock.add must be called.
+	currIndexBlock *indexBlockBuf
+	indexEntrySep  InternalKey
+	// inflightSize is used to decrement Writer.coordination.sizeEstimate.inflightSize.
+	inflightSize int
+	// inflightIndexEntrySize is used to decrement Writer.indexBlock.sizeEstimate.inflightSize.
+	indexInflightSize int
+	// If the index block is finished, then we set the finishedIndexProps here.
+	finishedIndexProps []byte
+}
+
+// It is not the responsibility of the writeTask to clear the
+// task.flushableIndexBlock, and task.buf.
+func (task *writeTask) clear() {
+	*task = writeTask{
+		indexEntrySep:   base.InvalidInternalKey,
+		compressionDone: task.compressionDone,
+	}
+}
+
+// Note that only the Writer client goroutine will be adding tasks to the writeQueue.
+// Both the Writer client and the compression goroutines will be able to write to
+// writeTask.compressionDone to indicate that the compression job associated with
+// a writeTask has finished.
+type writeQueue struct {
+	tasks  chan *writeTask
+	wg     sync.WaitGroup
+	writer *Writer
+
+	// err represents an error which is encountered when the write queue attempts
+	// to write a block to disk. The error is stored here to skip unnecessary block
+	// writes once the first error is encountered.
+	err    error
+	closed bool
+}
+
+func newWriteQueue(size int, writer *Writer) *writeQueue {
+	w := &writeQueue{}
+	w.tasks = make(chan *writeTask, size)
+	w.writer = writer
+
+	w.wg.Add(1)
+	go w.runWorker()
+	return w
+}
+
+func (w *writeQueue) performWrite(task *writeTask) error {
+	var bh BlockHandle
+	var bhp BlockHandleWithProperties
+
+	var err error
+	if bh, err = w.writer.writeCompressedBlock(task.buf.compressed, task.buf.tmp[:]); err != nil {
+		return err
+	}
+
+	// Update the size estimates after writing the data block to disk.
+	w.writer.coordination.sizeEstimate.dataBlockWritten(
+		w.writer.meta.Size, task.inflightSize, int(bh.Length),
+	)
+
+	bhp = BlockHandleWithProperties{BlockHandle: bh, Props: task.buf.dataBlockProps}
+	if err = w.writer.addIndexEntry(
+		task.indexEntrySep, bhp, task.buf.tmp[:], task.flushableIndexBlock, task.currIndexBlock,
+		task.indexInflightSize, task.finishedIndexProps); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+// It is necessary to ensure that none of the buffers in the writeTask,
+// dataBlockBuf, indexBlockBuf, are pointed to by another struct.
+func (w *writeQueue) releaseBuffers(task *writeTask) {
+	task.buf.clear()
+	dataBlockBufPool.Put(task.buf)
+
+	// This index block is no longer used by the Writer, so we can add it back
+	// to the pool.
+	if task.flushableIndexBlock != nil {
+		task.flushableIndexBlock.clear()
+		indexBlockBufPool.Put(task.flushableIndexBlock)
+	}
+
+	task.clear()
+	writeTaskPool.Put(task)
+}
+
+func (w *writeQueue) runWorker() {
+	for task := range w.tasks {
+		<-task.compressionDone
+
+		if w.err == nil {
+			w.err = w.performWrite(task)
+		}
+
+		w.releaseBuffers(task)
+	}
+	w.wg.Done()
+}
+
+func (w *writeQueue) add(task *writeTask) {
+	w.tasks <- task
+}
+
+// addSync will perform the writeTask synchronously with the caller goroutine. Calls to addSync
+// are no longer valid once writeQueue.add has been called at least once.
+func (w *writeQueue) addSync(task *writeTask) error {
+	// This should instantly return without blocking.
+	<-task.compressionDone
+
+	if w.err == nil {
+		w.err = w.performWrite(task)
+	}
+
+	w.releaseBuffers(task)
+
+	return w.err
+}
+
+// finish should only be called once no more tasks will be added to the writeQueue.
+// finish will return any error which was encountered while tasks were processed.
+func (w *writeQueue) finish() error {
+	if w.closed {
+		return w.err
+	}
+
+	close(w.tasks)
+	w.wg.Wait()
+	w.closed = true
+	return w.err
+}