func runMVCCConditionalPut(valueSize int, createFirst bool, b *testing.B) {
	rng, _ := randutil.NewPseudoRand()
	value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueSize))
	keyBuf := append(make([]byte, 0, 64), []byte("key-")...)

	stopper := stop.NewStopper()
	defer stopper.Stop()
	rocksdb := NewInMem(roachpb.Attributes{}, testCacheSize, stopper)

	b.SetBytes(int64(valueSize))
	var expected *roachpb.Value
	if createFirst {
		for i := 0; i < b.N; i++ {
			key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
			ts := makeTS(timeutil.Now().UnixNano(), 0)
			if err := MVCCPut(rocksdb, nil, key, ts, value, nil); err != nil {
				b.Fatalf("failed put: %s", err)
			}
		}
		expected = &value
	}

	b.ResetTimer()

	for i := 0; i < b.N; i++ {
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
		ts := makeTS(timeutil.Now().UnixNano(), 0)
		if err := MVCCConditionalPut(rocksdb, nil, key, ts, value, expected, nil); err != nil {
			b.Fatalf("failed put: %s", err)
		}
	}

	b.StopTimer()
}

func runMVCCConditionalPut(emk engineMaker, valueSize int, createFirst bool, b *testing.B) {
	rng, _ := randutil.NewPseudoRand()
	value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueSize))
	keyBuf := append(make([]byte, 0, 64), []byte("key-")...)

	eng, stopper := emk(b, fmt.Sprintf("cput_%d", valueSize))
	defer stopper.Stop()

	b.SetBytes(int64(valueSize))
	var expected *roachpb.Value
	if createFirst {
		for i := 0; i < b.N; i++ {
			key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
			ts := makeTS(timeutil.Now().UnixNano(), 0)
			if err := MVCCPut(context.Background(), eng, nil, key, ts, value, nil); err != nil {
				b.Fatalf("failed put: %s", err)
			}
		}
		expected = &value
	}

	b.ResetTimer()

	for i := 0; i < b.N; i++ {
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
		ts := makeTS(timeutil.Now().UnixNano(), 0)
		if err := MVCCConditionalPut(context.Background(), eng, nil, key, ts, value, expected, nil); err != nil {
			b.Fatalf("failed put: %s", err)
		}
	}

	b.StopTimer()
}

// runClientScan first creates test data (and resets the benchmarking
// timer). It then performs b.N client scans in increments of numRows
// keys over all of the data, restarting at the beginning of the
// keyspace, as many times as necessary.
func runClientScan(useSSL bool, numRows, numVersions int, b *testing.B) {
	const numKeys = 100000

	s, db := setupClientBenchData(useSSL, numVersions, numKeys, b)
	defer s.Stop()

	b.SetBytes(int64(numRows * valueSize))
	b.ResetTimer()

	b.RunParallel(func(pb *testing.PB) {
		startKeyBuf := append(make([]byte, 0, 64), []byte("key-")...)
		endKeyBuf := append(make([]byte, 0, 64), []byte("key-")...)
		for pb.Next() {
			// Choose a random key to start scan.
			keyIdx := rand.Int31n(int32(numKeys - numRows))
			startKey := roachpb.Key(encoding.EncodeUvarintAscending(
				startKeyBuf, uint64(keyIdx)))
			endKey := roachpb.Key(encoding.EncodeUvarintAscending(
				endKeyBuf, uint64(keyIdx)+uint64(numRows)))
			rows, pErr := db.Scan(startKey, endKey, int64(numRows))
			if pErr != nil {
				b.Fatalf("failed scan: %s", pErr)
			}
			if len(rows) != numRows {
				b.Fatalf("failed to scan: %d != %d", len(rows), numRows)
			}
		}
	})

	b.StopTimer()
}

func sqlKV(tableID uint32, indexID, descriptorID uint64) roachpb.KeyValue {
	k := keys.MakeTablePrefix(tableID)
	k = encoding.EncodeUvarintAscending(k, indexID)
	k = encoding.EncodeUvarintAscending(k, descriptorID)
	k = encoding.EncodeUvarintAscending(k, 12345) // Column ID, but could be anything.
	return kv(k, nil)
}

// MakeColumnKey returns the key for the column in the given row.
func MakeColumnKey(rowKey []byte, colID uint32) []byte {
	key := append([]byte(nil), rowKey...)
	size := len(key)
	key = encoding.EncodeUvarintAscending(key, uint64(colID))
	// Note that we assume that `len(key)-size` will always be encoded to a
	// single byte by EncodeUvarint. This is currently always true because the
	// varint encoding will encode 1-9 bytes.
	return encoding.EncodeUvarintAscending(key, uint64(len(key)-size))
}

// MakeIndexKeyPrefix returns the key prefix used for the index's data.
func MakeIndexKeyPrefix(desc *TableDescriptor, indexID IndexID) []byte {
	var key []byte
	if i, err := desc.FindIndexByID(indexID); err == nil && len(i.Interleave.Ancestors) > 0 {
		key = encoding.EncodeUvarintAscending(key, uint64(i.Interleave.Ancestors[0].TableID))
		key = encoding.EncodeUvarintAscending(key, uint64(i.Interleave.Ancestors[0].IndexID))
		return key
	}
	key = encoding.EncodeUvarintAscending(key, uint64(desc.ID))
	key = encoding.EncodeUvarintAscending(key, uint64(indexID))
	return key
}

// MakeNameMetadataKey returns the key for the name. Pass name == "" in order
// to generate the prefix key to use to scan over all of the names for the
// specified parentID.
func MakeNameMetadataKey(parentID ID, name string) roachpb.Key {
	name = NormalizeName(name)
	k := keys.MakeTablePrefix(uint32(namespaceTable.ID))
	k = encoding.EncodeUvarintAscending(k, uint64(namespaceTable.PrimaryIndex.ID))
	k = encoding.EncodeUvarintAscending(k, uint64(parentID))
	if name != "" {
		k = encoding.EncodeBytesAscending(k, []byte(name))
		k = keys.MakeColumnKey(k, uint32(namespaceTable.Columns[2].ID))
	}
	return k
}

// MakeFamilyKey returns the key for the family in the given row by appending to
// the passed key. If SentinelFamilyID is passed, a sentinel key (which is the
// first key in a sql table row) is returned.
func MakeFamilyKey(key []byte, famID uint32) []byte {
	if famID == SentinelFamilyID {
		return encoding.EncodeUvarintAscending(key, 0)
	}
	size := len(key)
	key = encoding.EncodeUvarintAscending(key, uint64(famID))
	// Note that we assume that `len(key)-size` will always be encoded to a
	// single byte by EncodeUvarint. This is currently always true because the
	// varint encoding will encode 1-9 bytes.
	return encoding.EncodeUvarintAscending(key, uint64(len(key)-size))
}

// MakeRangeIDPrefix creates a range-local key prefix from
// rangeID.
func MakeRangeIDPrefix(rangeID roachpb.RangeID) roachpb.Key {
	// Size the key buffer so that it is large enough for most callers.
	key := make(roachpb.Key, 0, 32)
	key = append(key, LocalRangeIDPrefix...)
	key = encoding.EncodeUvarintAscending(key, uint64(rangeID))
	return key
}

func BenchmarkBatchBuilderPut(b *testing.B) {
	value := make([]byte, 10)
	for i := range value {
		value[i] = byte(i)
	}
	keyBuf := append(make([]byte, 0, 64), []byte("key-")...)

	b.ResetTimer()

	const batchSize = 1000
	batch := &rocksDBBatchBuilder{}
	for i := 0; i < b.N; i += batchSize {
		end := i + batchSize
		if end > b.N {
			end = b.N
		}

		for j := i; j < end; j++ {
			key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(j)))
			ts := hlc.Timestamp{WallTime: int64(j)}
			batch.Put(MVCCKey{key, ts}, value)
		}
		batch.Finish()
	}

	b.StopTimer()
}

// runMVCCScan first creates test data (and resets the benchmarking
// timer). It then performs b.N MVCCScans in increments of numRows
// keys over all of the data in the rocksdb instance, restarting at
// the beginning of the keyspace, as many times as necessary.
func runMVCCScan(numRows, numVersions, valueSize int, b *testing.B) {
	// Use the same number of keys for all of the mvcc scan
	// benchmarks. Using a different number of keys per test gives
	// preferential treatment to tests with fewer keys. Note that the
	// datasets all fit in cache and the cache is pre-warmed.
	const numKeys = 100000

	rocksdb, stopper := setupMVCCData(numVersions, numKeys, valueSize, b)
	defer stopper.Stop()

	b.SetBytes(int64(numRows * valueSize))
	b.ResetTimer()

	keyBuf := append(make([]byte, 0, 64), []byte("key-")...)
	for i := 0; i < b.N; i++ {
		// Choose a random key to start scan.
		keyIdx := rand.Int31n(int32(numKeys - numRows))
		startKey := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(keyIdx)))
		walltime := int64(5 * (rand.Int31n(int32(numVersions)) + 1))
		ts := makeTS(walltime, 0)
		kvs, _, err := MVCCScan(rocksdb, startKey, keyMax, int64(numRows), ts, true, nil)
		if err != nil {
			b.Fatalf("failed scan: %s", err)
		}
		if len(kvs) != numRows {
			b.Fatalf("failed to scan: %d != %d", len(kvs), numRows)
		}
	}

	b.StopTimer()
}

// runMVCCGet first creates test data (and resets the benchmarking
// timer). It then performs b.N MVCCGets.
func runMVCCGet(numVersions, valueSize int, b *testing.B) {
	const overhead = 48          // Per key/value overhead (empirically determined)
	const targetSize = 512 << 20 // 512 MB
	// Adjust the number of keys so that each test has approximately the same
	// amount of data.
	numKeys := targetSize / ((overhead + valueSize) * (1 + (numVersions-1)/2))

	rocksdb, stopper := setupMVCCData(numVersions, numKeys, valueSize, b)
	defer stopper.Stop()

	b.SetBytes(int64(valueSize))
	b.ResetTimer()

	keyBuf := append(make([]byte, 0, 64), []byte("key-")...)
	for i := 0; i < b.N; i++ {
		// Choose a random key to retrieve.
		keyIdx := rand.Int31n(int32(numKeys))
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(keyIdx)))
		walltime := int64(5 * (rand.Int31n(int32(numVersions)) + 1))
		ts := makeTS(walltime, 0)
		if v, _, err := MVCCGet(rocksdb, key, ts, true, nil); err != nil {
			b.Fatalf("failed get: %s", err)
		} else if v == nil {
			b.Fatalf("failed get (key not found): %[email protected]%d", keyIdx, walltime)
		} else if valueBytes, err := v.GetBytes(); err != nil {
			b.Fatal(err)
		} else if len(valueBytes) != valueSize {
			b.Fatalf("unexpected value size: %d", len(valueBytes))
		}
	}

	b.StopTimer()
}

func makePrefixWithRangeID(prefix []byte, rangeID roachpb.RangeID, infix roachpb.RKey) roachpb.Key {
	// Size the key buffer so that it is large enough for most callers.
	key := make(roachpb.Key, 0, 32)
	key = append(key, prefix...)
	key = encoding.EncodeUvarintAscending(key, uint64(rangeID))
	key = append(key, infix...)
	return key
}

func (td *tableDeleter) deleteAllRowsFast(ctx context.Context) error {
	var tablePrefix []byte
	// TODO(dan): This should be moved into keys.MakeTablePrefix, but updating
	// all the uses of that will be a pain.
	if interleave := td.rd.helper.tableDesc.PrimaryIndex.Interleave; len(interleave.Ancestors) > 0 {
		tablePrefix = encoding.EncodeUvarintAscending(nil, uint64(interleave.Ancestors[0].TableID))
	}
	tablePrefix = encoding.EncodeUvarintAscending(nil, uint64(td.rd.helper.tableDesc.ID))

	// Delete rows and indexes starting with the table's prefix.
	tableStartKey := roachpb.Key(tablePrefix)
	tableEndKey := tableStartKey.PrefixEnd()
	if log.V(2) {
		log.Infof(ctx, "DelRange %s - %s", tableStartKey, tableEndKey)
	}
	td.b.DelRange(tableStartKey, tableEndKey, false)
	return td.finalize(ctx)
}

// EncodeIndexKey creates a key by concatenating keyPrefix with the encodings of
// the columns in the index.
//
// If a table or index is interleaved, `encoding.encodedNullDesc` is used in
// place of the family id (a varint) to signal the next component of the key.
// An example of one level of interleaving (a parent):
// /<parent_table_id>/<parent_index_id>/<field_1>/<field_2>/NullDesc/<table_id>/<index_id>/<field_3>/<family>
//
// Returns the key and whether any of the encoded values were NULLs.
//
// Note that ImplicitColumnIDs are not encoded, so the result isn't always a
// full index key.
func EncodeIndexKey(
	tableDesc *TableDescriptor,
	index *IndexDescriptor,
	colMap map[ColumnID]int,
	values []parser.Datum,
	keyPrefix []byte,
) (key []byte, containsNull bool, err error) {
	key = keyPrefix
	colIDs := index.ColumnIDs
	dirs := directions(index.ColumnDirections)

	if len(index.Interleave.Ancestors) > 0 {
		for i, ancestor := range index.Interleave.Ancestors {
			// The first ancestor is assumed to already be encoded in keyPrefix.
			if i != 0 {
				key = encoding.EncodeUvarintAscending(key, uint64(ancestor.TableID))
				key = encoding.EncodeUvarintAscending(key, uint64(ancestor.IndexID))
			}

			length := int(ancestor.SharedPrefixLen)
			var n bool
			key, n, err = EncodeColumns(colIDs[:length], dirs[:length], colMap, values, key)
			if err != nil {
				return key, containsNull, err
			}
			colIDs, dirs = colIDs[length:], dirs[length:]
			containsNull = containsNull || n

			// We reuse NotNullDescending (0xfe) as the interleave sentinel.
			key = encoding.EncodeNotNullDescending(key)
		}

		key = encoding.EncodeUvarintAscending(key, uint64(tableDesc.ID))
		key = encoding.EncodeUvarintAscending(key, uint64(index.ID))
	}

	var n bool
	key, n, err = EncodeColumns(colIDs, dirs, colMap, values, key)
	containsNull = containsNull || n
	return key, containsNull, err
}

func TestMakeSplitKey(t *testing.T) {
	e := func(vals ...uint64) roachpb.Key {
		var k roachpb.Key
		for _, v := range vals {
			k = encoding.EncodeUvarintAscending(k, v)
		}
		return k
	}

	goodData := []struct {
		in       roachpb.Key
		expected roachpb.Key
	}{
		{e(1, 2, 0), e(1, 2)},          // /Table/1/2/0 -> /Table/1/2
		{e(1, 2, 1), e(1)},             // /Table/1/2/1 -> /Table/1
		{e(1, 2, 2), e()},              // /Table/1/2/2 -> /Table
		{e(1, 2, 3, 0), e(1, 2, 3)},    // /Table/1/2/3/0 -> /Table/1/2/3
		{e(1, 2, 3, 1), e(1, 2)},       // /Table/1/2/3/1 -> /Table/1/2
		{e(1, 2, 200, 2), e(1, 2)},     // /Table/1/2/200/2 -> /Table/1/2
		{e(1, 2, 3, 4, 1), e(1, 2, 3)}, // /Table/1/2/3/4/1 -> /Table/1/2/3
	}
	for i, d := range goodData {
		out, err := MakeSplitKey(d.in)
		if err != nil {
			t.Fatalf("%d: %s: unexpected error: %v", i, d.in, err)
		}
		if !d.expected.Equal(out) {
			t.Fatalf("%d: %s: expected %s, but got %s", i, d.in, d.expected, out)
		}
	}

	errorData := []struct {
		in  roachpb.Key
		err string
	}{
		// Column ID suffix size is too large.
		{e(1), "malformed table key"},
		{e(1, 2), "malformed table key"},
		// The table ID is invalid.
		{e(200)[:1], "insufficient bytes to decode uvarint value"},
		// The index ID is invalid.
		{e(1, 200)[:2], "insufficient bytes to decode uvarint value"},
		// The column ID suffix is invalid.
		{e(1, 2, 200)[:3], "insufficient bytes to decode uvarint value"},
	}
	for i, d := range errorData {
		_, err := MakeSplitKey(d.in)
		if !testutils.IsError(err, d.err) {
			t.Fatalf("%d: %s: expected %s, but got %v", i, d.in, d.err, err)
		}
	}
}

func runMVCCBatchPut(valueSize, batchSize int, b *testing.B) {
	defer tracing.Disable()()
	rng, _ := randutil.NewPseudoRand()
	value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueSize))
	keyBuf := append(make([]byte, 0, 64), []byte("key-")...)

	stopper := stop.NewStopper()
	defer stopper.Stop()
	rocksdb := NewInMem(roachpb.Attributes{}, testCacheSize, stopper)

	b.SetBytes(int64(valueSize))
	b.ResetTimer()

	for i := 0; i < b.N; i += batchSize {
		end := i + batchSize
		if end > b.N {
			end = b.N
		}

		batch := rocksdb.NewBatch()

		for j := i; j < end; j++ {
			key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(j)))
			ts := makeTS(time.Now().UnixNano(), 0)
			if err := MVCCPut(batch, nil, key, ts, value, nil); err != nil {
				b.Fatalf("failed put: %s", err)
			}
		}

		if err := batch.Commit(); err != nil {
			b.Fatal(err)
		}

		batch.Close()
	}

	b.StopTimer()
}

func runMVCCBatchPut(emk engineMaker, valueSize, batchSize int, b *testing.B) {
	rng, _ := randutil.NewPseudoRand()
	value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueSize))
	keyBuf := append(make([]byte, 0, 64), []byte("key-")...)

	stopper := stop.NewStopper()
	eng, stopper := emk(b, fmt.Sprintf("batch_put_%d_%d", valueSize, batchSize))
	defer stopper.Stop()

	b.SetBytes(int64(valueSize))
	b.ResetTimer()

	for i := 0; i < b.N; i += batchSize {
		end := i + batchSize
		if end > b.N {
			end = b.N
		}

		batch := eng.NewBatch()

		for j := i; j < end; j++ {
			key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(j)))
			ts := makeTS(timeutil.Now().UnixNano(), 0)
			if err := MVCCPut(context.Background(), batch, nil, key, ts, value, nil); err != nil {
				b.Fatalf("failed put: %s", err)
			}
		}

		if err := batch.Commit(); err != nil {
			b.Fatal(err)
		}

		batch.Close()
	}

	b.StopTimer()
}

// setupMVCCData writes up to numVersions values at each of numKeys
// keys. The number of versions written for each key is chosen
// randomly according to a uniform distribution. Each successive
// version is written starting at 5ns and then in 5ns increments. This
// allows scans at various times, starting at t=5ns, and continuing to
// t=5ns*(numVersions+1). A version for each key will be read on every
// such scan, but the dynamics of the scan will change depending on
// the historical timestamp. Earlier timestamps mean scans which must
// skip more historical versions; later timestamps mean scans which
// skip fewer.
//
// The creation of the database is time consuming, especially for larger
// numbers of versions. The database is persisted between runs and stored in
// the current directory as "mvcc_scan_<versions>_<keys>_<valueBytes>" (which
// is also returned).
func setupMVCCData(emk engineMaker, numVersions, numKeys, valueBytes int, b *testing.B) (Engine, string, *stop.Stopper) {
	loc := fmt.Sprintf("mvcc_data_%d_%d_%d", numVersions, numKeys, valueBytes)

	exists := true
	if _, err := os.Stat(loc); os.IsNotExist(err) {
		exists = false
	}

	eng, stopper := emk(b, loc)

	if exists {
		readAllFiles(filepath.Join(loc, "*"))
		return eng, loc, stopper
	}

	log.Infof(context.Background(), "creating mvcc data: %s", loc)

	// Generate the same data every time.
	rng := rand.New(rand.NewSource(1449168817))

	keys := make([]roachpb.Key, numKeys)
	var order []int
	for i := 0; i < numKeys; i++ {
		keys[i] = roachpb.Key(encoding.EncodeUvarintAscending([]byte("key-"), uint64(i)))
		keyVersions := rng.Intn(numVersions) + 1
		for j := 0; j < keyVersions; j++ {
			order = append(order, i)
		}
	}

	// Randomize the order in which the keys are written.
	for i, n := 0, len(order); i < n-1; i++ {
		j := i + rng.Intn(n-i)
		order[i], order[j] = order[j], order[i]
	}

	counts := make([]int, numKeys)
	batch := eng.NewBatch()
	for i, idx := range order {
		// Output the keys in ~20 batches. If we used a single batch to output all
		// of the keys rocksdb would create a single sstable. We want multiple
		// sstables in order to exercise filtering of which sstables are examined
		// during iterator seeking. We fix the number of batches we output so that
		// optimizations which change the data size result in the same number of
		// sstables.
		if scaled := len(order) / 20; i > 0 && (i%scaled) == 0 {
			log.Infof(context.Background(), "committing (%d/~%d)", i/scaled, 20)
			if err := batch.Commit(); err != nil {
				b.Fatal(err)
			}
			batch.Close()
			batch = eng.NewBatch()
			if err := eng.Flush(); err != nil {
				b.Fatal(err)
			}
		}

		key := keys[idx]
		ts := makeTS(int64(counts[idx]+1)*5, 0)
		counts[idx]++
		value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueBytes))
		value.InitChecksum(key)
		if err := MVCCPut(context.Background(), batch, nil, key, ts, value, nil); err != nil {
			b.Fatal(err)
		}
	}
	if err := batch.Commit(); err != nil {
		b.Fatal(err)
	}
	batch.Close()
	if err := eng.Flush(); err != nil {
		b.Fatal(err)
	}

	return eng, loc, stopper
}

// MakeTablePrefix returns the key prefix used for the table's data.
func MakeTablePrefix(tableID uint32) []byte {
	return encoding.EncodeUvarintAscending(nil, uint64(tableID))
}