github.com/cockroachdb/cockroach@v20.2.0-alpha.1+incompatible/pkg/storage/bench_test.go

// Copyright 2014 The Cockroach Authors.
//
// Use of this software is governed by the Business Source License
// included in the file licenses/BSL.txt.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0, included in the file
// licenses/APL.txt.

package storage

import (
	"context"
	"fmt"
	"math"
	"math/rand"
	"os"
	"path/filepath"
	"testing"
	"time"

	"github.com/cockroachdb/cockroach/pkg/base"
	"github.com/cockroachdb/cockroach/pkg/roachpb"
	"github.com/cockroachdb/cockroach/pkg/settings/cluster"
	"github.com/cockroachdb/cockroach/pkg/storage/enginepb"
	"github.com/cockroachdb/cockroach/pkg/testutils"
	"github.com/cockroachdb/cockroach/pkg/util/encoding"
	"github.com/cockroachdb/cockroach/pkg/util/fileutil"
	"github.com/cockroachdb/cockroach/pkg/util/hlc"
	"github.com/cockroachdb/cockroach/pkg/util/log"
	"github.com/cockroachdb/cockroach/pkg/util/randutil"
	"github.com/cockroachdb/cockroach/pkg/util/timeutil"
)

const overhead = 48 // Per key/value overhead (empirically determined)

type engineMaker func(testing.TB, string) Engine

type benchDataOptions struct {
	numVersions int
	numKeys     int
	valueBytes  int

	// In transactional mode, data is written by writing and later resolving
	// intents. In non-transactional mode, data is written directly, without
	// leaving intents. Transactional mode notably stresses RocksDB deletion
	// tombstones, as the metadata key is repeatedly written and deleted.
	//
	// Both modes are reflective of real workloads. Transactional mode simulates
	// data that has recently been INSERTed into a table, while non-transactional
	// mode simulates data that has been RESTOREd or is old enough to have been
	// fully compacted.
	transactional bool
}

// loadTestData writes numKeys keys in numBatches separate batches. Keys are
// written in order. Every key in a given batch has the same MVCC timestamp;
// batch timestamps start at batchTimeSpan and increase in intervals of
// batchTimeSpan.
//
// Importantly, writing keys in order convinces RocksDB to output one SST per
// batch, where each SST contains keys of only one timestamp. E.g., writing A,B
// at t0 and C at t1 will create two SSTs: one for A,B that only contains keys
// at t0, and one for C that only contains keys at t1. Conversely, writing A, C
// at t0 and B at t1 would create just one SST that contained A,B,C (due to an
// immediate compaction).
//
// The creation of the database is time consuming, so the caller can choose
// whether to use a temporary or permanent location.
func loadTestData(dir string, numKeys, numBatches, batchTimeSpan, valueBytes int) (Engine, error) {
	ctx := context.Background()

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

	eng, err := NewRocksDB(
		RocksDBConfig{
			StorageConfig: base.StorageConfig{
				Settings: cluster.MakeTestingClusterSettings(),
				Dir:      dir,
			},
		},
		RocksDBCache{},
	)
	if err != nil {
		return nil, err
	}

	if exists {
		testutils.ReadAllFiles(filepath.Join(dir, "*"))
		return eng, nil
	}

	log.Infof(context.Background(), "creating test data: %s", dir)

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

	keys := make([]roachpb.Key, numKeys)
	for i := 0; i < numKeys; i++ {
		keys[i] = roachpb.Key(encoding.EncodeUvarintAscending([]byte("key-"), uint64(i)))
	}

	sstTimestamps := make([]int64, numBatches)
	for i := 0; i < len(sstTimestamps); i++ {
		sstTimestamps[i] = int64((i + 1) * batchTimeSpan)
	}

	var batch Batch
	var minWallTime int64
	for i, key := range keys {
		if scaled := len(keys) / numBatches; (i % scaled) == 0 {
			if i > 0 {
				log.Infof(ctx, "committing (%d/~%d)", i/scaled, numBatches)
				if err := batch.Commit(false /* sync */); err != nil {
					return nil, err
				}
				batch.Close()
				if err := eng.Flush(); err != nil {
					return nil, err
				}
			}
			batch = eng.NewBatch()
			minWallTime = sstTimestamps[i/scaled]
		}
		timestamp := hlc.Timestamp{WallTime: minWallTime + rand.Int63n(int64(batchTimeSpan))}
		value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueBytes))
		value.InitChecksum(key)
		if err := MVCCPut(ctx, batch, nil, key, timestamp, value, nil); err != nil {
			return nil, err
		}
	}
	if err := batch.Commit(false /* sync */); err != nil {
		return nil, err
	}
	batch.Close()
	if err := eng.Flush(); err != nil {
		return nil, err
	}

	return eng, nil
}

// 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(
	ctx context.Context, b *testing.B, emk engineMaker, opts benchDataOptions,
) (Engine, string) {
	loc := fmt.Sprintf("mvcc_data_%d_%d_%d", opts.numVersions, opts.numKeys, opts.valueBytes)
	if opts.transactional {
		loc += "_txn"
	}

	exists := true
	if _, err := os.Stat(loc); os.IsNotExist(err) {
		exists = false
	} else if err != nil {
		b.Fatal(err)
	}

	eng := emk(b, loc)

	if exists {
		testutils.ReadAllFiles(filepath.Join(loc, "*"))
		return eng, loc
	}

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

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

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

	var txn *roachpb.Transaction
	if opts.transactional {
		txnCopy := *txn1Commit
		txn = &txnCopy
	}

	writeKey := func(batch Batch, idx int) {
		key := keys[idx]
		value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, opts.valueBytes))
		value.InitChecksum(key)
		counts[idx]++
		ts := hlc.Timestamp{WallTime: int64(counts[idx] * 5)}
		if txn != nil {
			txn.ReadTimestamp = ts
			txn.WriteTimestamp = ts
		}
		if err := MVCCPut(ctx, batch, nil /* ms */, key, ts, value, txn); err != nil {
			b.Fatal(err)
		}
	}

	resolveLastIntent := func(batch Batch, idx int) {
		key := keys[idx]
		txnMeta := txn.TxnMeta
		txnMeta.WriteTimestamp = hlc.Timestamp{WallTime: int64(counts[idx]) * 5}
		if _, err := MVCCResolveWriteIntent(ctx, batch, nil /* ms */, roachpb.LockUpdate{
			Span:   roachpb.Span{Key: key},
			Status: roachpb.COMMITTED,
			Txn:    txnMeta,
		}); err != nil {
			b.Fatal(err)
		}
	}

	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(ctx, "committing (%d/~%d)", i/scaled, 20)
			if err := batch.Commit(false /* sync */); err != nil {
				b.Fatal(err)
			}
			batch.Close()
			batch = eng.NewBatch()
			if err := eng.Flush(); err != nil {
				b.Fatal(err)
			}
		}

		if opts.transactional {
			// If we've previously written this key transactionally, we need to
			// resolve the intent we left. We don't do this immediately after writing
			// the key to introduce the possibility that the intent's resolution ends
			// up in a different batch than writing the intent itself. Note that the
			// first time through this loop for any given key we'll attempt to resolve
			// a non-existent intent, but that's OK.
			resolveLastIntent(batch, idx)
		}
		writeKey(batch, idx)
	}
	if opts.transactional {
		// If we were writing transactionally, we need to do one last round of
		// intent resolution. Just stuff it all into the last batch.
		for idx := range keys {
			resolveLastIntent(batch, idx)
		}
	}
	if err := batch.Commit(false /* sync */); err != nil {
		b.Fatal(err)
	}
	batch.Close()
	if err := eng.Flush(); err != nil {
		b.Fatal(err)
	}

	return eng, loc
}

type benchScanOptions struct {
	benchDataOptions
	numRows int
	reverse bool
}

// 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 Engine instance, restarting at
// the beginning of the keyspace, as many times as necessary.
func runMVCCScan(ctx context.Context, b *testing.B, emk engineMaker, opts benchScanOptions) {
	// 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.
	if opts.numKeys != 0 {
		b.Fatal("test error: cannot call runMVCCScan with non-zero numKeys")
	}
	opts.numKeys = 100000

	eng, _ := setupMVCCData(ctx, b, emk, opts.benchDataOptions)
	defer eng.Close()

	{
		// Pull all of the sstables into the RocksDB cache in order to make the
		// timings more stable. Otherwise, the first run will be penalized pulling
		// data into the cache while later runs will not.
		iter := eng.NewIterator(IterOptions{UpperBound: roachpb.KeyMax})
		_, _ = iter.ComputeStats(roachpb.KeyMin, roachpb.KeyMax, 0)
		iter.Close()
	}

	b.SetBytes(int64(opts.numRows * opts.valueBytes))
	b.ResetTimer()

	startKeyBuf := append(make([]byte, 0, 64), []byte("key-")...)
	endKeyBuf := 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(opts.numKeys - opts.numRows))
		startKey := roachpb.Key(encoding.EncodeUvarintAscending(startKeyBuf[:4], uint64(keyIdx)))
		endKey := roachpb.Key(encoding.EncodeUvarintAscending(endKeyBuf[:4], uint64(keyIdx+int32(opts.numRows)-1)))
		endKey = endKey.Next()
		walltime := int64(5 * (rand.Int31n(int32(opts.numVersions)) + 1))
		ts := hlc.Timestamp{WallTime: walltime}
		res, err := MVCCScan(ctx, eng, startKey, endKey, ts, MVCCScanOptions{
			MaxKeys: int64(opts.numRows),
			Reverse: opts.reverse,
		})
		if err != nil {
			b.Fatalf("failed scan: %+v", err)
		}
		if len(res.KVs) != opts.numRows {
			b.Fatalf("failed to scan: %d != %d", len(res.KVs), opts.numRows)
		}
	}

	b.StopTimer()
}

// runMVCCGet first creates test data (and resets the benchmarking
// timer). It then performs b.N MVCCGets.
func runMVCCGet(ctx context.Context, b *testing.B, emk engineMaker, opts benchDataOptions) {
	// 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.
	if opts.numKeys != 0 {
		b.Fatal("test error: cannot call runMVCCGet with non-zero numKeys")
	}
	opts.numKeys = 100000

	eng, _ := setupMVCCData(ctx, b, emk, opts)
	defer eng.Close()

	b.SetBytes(int64(opts.valueBytes))
	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(opts.numKeys))
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(keyIdx)))
		walltime := int64(5 * (rand.Int31n(int32(opts.numVersions)) + 1))
		ts := hlc.Timestamp{WallTime: walltime}
		if v, _, err := MVCCGet(ctx, eng, key, ts, MVCCGetOptions{}); err != nil {
			b.Fatalf("failed get: %+v", err)
		} else if v == nil {
			b.Fatalf("failed get (key not found): %d@%d", keyIdx, walltime)
		} else if valueBytes, err := v.GetBytes(); err != nil {
			b.Fatal(err)
		} else if len(valueBytes) != opts.valueBytes {
			b.Fatalf("unexpected value size: %d", len(valueBytes))
		}
	}

	b.StopTimer()
}

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

	eng := emk(b, fmt.Sprintf("put_%d", valueSize))
	defer eng.Close()

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

	for i := 0; i < b.N; i++ {
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
		ts := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
		if err := MVCCPut(ctx, eng, nil, key, ts, value, nil); err != nil {
			b.Fatalf("failed put: %+v", err)
		}
	}

	b.StopTimer()
}

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

	eng := emk(b, fmt.Sprintf("put_%d", valueSize))
	defer eng.Close()

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

	for i := 0; i < b.N; i++ {
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
		ts := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
		if err := MVCCBlindPut(ctx, eng, nil, key, ts, value, nil); err != nil {
			b.Fatalf("failed put: %+v", err)
		}
	}

	b.StopTimer()
}

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

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

	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 := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
			if err := MVCCPut(ctx, eng, nil, key, ts, value, nil); err != nil {
				b.Fatalf("failed put: %+v", err)
			}
		}
		expected = &value
	}

	b.ResetTimer()

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

	b.StopTimer()
}

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

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

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

	for i := 0; i < b.N; i++ {
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
		ts := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
		if err := MVCCBlindConditionalPut(ctx, eng, nil, key, ts, value, nil, CPutFailIfMissing, nil); err != nil {
			b.Fatalf("failed put: %+v", err)
		}
	}

	b.StopTimer()
}

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

	eng := emk(b, fmt.Sprintf("iput_%d", valueSize))
	defer eng.Close()

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

	for i := 0; i < b.N; i++ {
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
		ts := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
		if err := MVCCInitPut(ctx, eng, nil, key, ts, value, false, nil); err != nil {
			b.Fatalf("failed put: %+v", err)
		}
	}

	b.StopTimer()
}

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

	eng := emk(b, fmt.Sprintf("iput_%d", valueSize))
	defer eng.Close()

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

	for i := 0; i < b.N; i++ {
		key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
		ts := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
		if err := MVCCBlindInitPut(ctx, eng, nil, key, ts, value, false, nil); err != nil {
			b.Fatalf("failed put: %+v", err)
		}
	}

	b.StopTimer()
}

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

	eng := emk(b, fmt.Sprintf("batch_put_%d_%d", valueSize, batchSize))
	defer eng.Close()

	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 := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
			if err := MVCCPut(ctx, batch, nil, key, ts, value, nil); err != nil {
				b.Fatalf("failed put: %+v", err)
			}
		}

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

		batch.Close()
	}

	b.StopTimer()
}

// Benchmark batch time series merge operations. This benchmark does not
// perform any reads and is only used to measure the cost of the periodic time
// series updates.
func runMVCCBatchTimeSeries(ctx context.Context, b *testing.B, emk engineMaker, batchSize int) {
	// Precompute keys so we don't waste time formatting them at each iteration.
	numKeys := batchSize
	keys := make([]roachpb.Key, numKeys)
	for i := 0; i < numKeys; i++ {
		keys[i] = roachpb.Key(fmt.Sprintf("key-%d", i))
	}

	// We always write the same time series data (containing a single unchanging
	// sample). This isn't realistic but is fine because we're never reading the
	// data.
	var value roachpb.Value
	if err := value.SetProto(&roachpb.InternalTimeSeriesData{
		StartTimestampNanos: 0,
		SampleDurationNanos: 1000,
		Samples: []roachpb.InternalTimeSeriesSample{
			{Offset: 0, Count: 1, Sum: 5.0},
		},
	}); err != nil {
		b.Fatal(err)
	}

	eng := emk(b, fmt.Sprintf("batch_merge_%d", batchSize))
	defer eng.Close()

	b.ResetTimer()

	var ts hlc.Timestamp
	for i := 0; i < b.N; i++ {
		batch := eng.NewBatch()

		for j := 0; j < batchSize; j++ {
			ts.Logical++
			if err := MVCCMerge(ctx, batch, nil, keys[j], ts, value); err != nil {
				b.Fatalf("failed put: %+v", err)
			}
		}

		if err := batch.Commit(false /* sync */); err != nil {
			b.Fatal(err)
		}
		batch.Close()
	}

	b.StopTimer()
}

// runMVCCMerge merges value into numKeys separate keys.
func runMVCCMerge(
	ctx context.Context, b *testing.B, emk engineMaker, value *roachpb.Value, numKeys int,
) {
	eng := emk(b, fmt.Sprintf("merge_%d", numKeys))
	defer eng.Close()

	// Precompute keys so we don't waste time formatting them at each iteration.
	keys := make([]roachpb.Key, numKeys)
	for i := 0; i < numKeys; i++ {
		keys[i] = roachpb.Key(fmt.Sprintf("key-%d", i))
	}

	b.ResetTimer()

	ts := hlc.Timestamp{}
	// Use parallelism if specified when test is run.
	b.RunParallel(func(pb *testing.PB) {
		for pb.Next() {
			ms := enginepb.MVCCStats{}
			ts.Logical++
			err := MVCCMerge(ctx, eng, &ms, keys[rand.Intn(numKeys)], ts, *value)
			if err != nil {
				b.Fatal(err)
			}
		}
	})
	b.StopTimer()

	// Read values out to force merge.
	for _, key := range keys {
		val, _, err := MVCCGet(ctx, eng, key, hlc.Timestamp{}, MVCCGetOptions{})
		if err != nil {
			b.Fatal(err)
		} else if val == nil {
			continue
		}
	}
}

// runMVCCGetMergedValue reads merged values for numKeys separate keys and mergesPerKey
// operands per key.
func runMVCCGetMergedValue(
	ctx context.Context, b *testing.B, emk engineMaker, numKeys, mergesPerKey int,
) {
	eng := emk(b, fmt.Sprintf("get_merged_%d_%d", numKeys, mergesPerKey))
	defer eng.Close()

	// Precompute keys so we don't waste time formatting them at each iteration.
	keys := make([]roachpb.Key, numKeys)
	for i := 0; i < numKeys; i++ {
		keys[i] = roachpb.Key(fmt.Sprintf("key-%d", i))
	}

	timestamp := hlc.Timestamp{}
	for i := 0; i < numKeys; i++ {
		for j := 0; j < mergesPerKey; j++ {
			timeseries := &roachpb.InternalTimeSeriesData{
				StartTimestampNanos: 0,
				SampleDurationNanos: 1000,
				Samples: []roachpb.InternalTimeSeriesSample{
					{Offset: int32(j), Count: 1, Sum: 5.0},
				},
			}
			var value roachpb.Value
			if err := value.SetProto(timeseries); err != nil {
				b.Fatal(err)
			}
			ms := enginepb.MVCCStats{}
			timestamp.Logical++
			err := MVCCMerge(ctx, eng, &ms, keys[i], timestamp, value)
			if err != nil {
				b.Fatal(err)
			}
		}
	}

	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		_, _, err := MVCCGet(ctx, eng, keys[rand.Intn(numKeys)], timestamp, MVCCGetOptions{})
		if err != nil {
			b.Fatal(err)
		}
	}
	b.StopTimer()
}

func runMVCCDeleteRange(ctx context.Context, b *testing.B, emk engineMaker, valueBytes int) {
	// 512 KB ranges so the benchmark doesn't take forever
	const rangeBytes = 512 * 1024
	numKeys := rangeBytes / (overhead + valueBytes)
	eng, dir := setupMVCCData(ctx, b, emk, benchDataOptions{
		numVersions: 1,
		numKeys:     numKeys,
		valueBytes:  valueBytes,
	})
	eng.Close()

	b.SetBytes(rangeBytes)
	b.StopTimer()
	b.ResetTimer()

	locDirty := dir + "_dirty"

	for i := 0; i < b.N; i++ {
		if err := os.RemoveAll(locDirty); err != nil {
			b.Fatal(err)
		}
		if err := fileutil.CopyDir(dir, locDirty); err != nil {
			b.Fatal(err)
		}
		func() {
			eng := emk(b, locDirty)
			defer eng.Close()

			b.StartTimer()
			if _, _, _, err := MVCCDeleteRange(
				ctx,
				eng,
				&enginepb.MVCCStats{},
				roachpb.KeyMin,
				roachpb.KeyMax,
				math.MaxInt64,
				hlc.MaxTimestamp,
				nil,
				false,
			); err != nil {
				b.Fatal(err)
			}
			b.StopTimer()
		}()
	}
}

func runClearRange(
	ctx context.Context,
	b *testing.B,
	emk engineMaker,
	clearRange func(e Engine, b Batch, start, end MVCCKey) error,
) {
	const rangeBytes = 64 << 20
	const valueBytes = 92
	numKeys := rangeBytes / (overhead + valueBytes)
	eng, _ := setupMVCCData(ctx, b, emk, benchDataOptions{
		numVersions: 1,
		numKeys:     numKeys,
		valueBytes:  valueBytes,
	})
	defer eng.Close()

	// It is not currently possible to ClearRange(NilKey, MVCCKeyMax) thanks to a
	// variety of hacks inside of ClearRange that explode if provided the NilKey.
	// So instead we start our ClearRange at the first key that actually exists.
	//
	// TODO(benesch): when those hacks are removed, don't bother computing the
	// first key and simply ClearRange(NilKey, MVCCKeyMax).
	iter := eng.NewIterator(IterOptions{UpperBound: roachpb.KeyMax})
	defer iter.Close()
	iter.SeekGE(NilKey)
	if ok, err := iter.Valid(); !ok {
		b.Fatalf("unable to find first key (err: %v)", err)
	}
	firstKey := iter.Key()

	b.SetBytes(rangeBytes)
	b.ResetTimer()

	for i := 0; i < b.N; i++ {
		batch := eng.NewWriteOnlyBatch()
		if err := clearRange(eng, batch, firstKey, MVCCKeyMax); err != nil {
			b.Fatal(err)
		}
		// NB: We don't actually commit the batch here as we don't want to delete
		// the data. Doing so would require repopulating on every iteration of the
		// loop which was ok when ClearRange was slow but now causes the benchmark
		// to take an exceptionally long time since ClearRange is very fast.
		batch.Close()
	}

	b.StopTimer()
}

// runMVCCComputeStats benchmarks computing MVCC stats on a 64MB range of data.
func runMVCCComputeStats(ctx context.Context, b *testing.B, emk engineMaker, valueBytes int) {
	const rangeBytes = 64 * 1024 * 1024
	numKeys := rangeBytes / (overhead + valueBytes)
	eng, _ := setupMVCCData(ctx, b, emk, benchDataOptions{
		numVersions: 1,
		numKeys:     numKeys,
		valueBytes:  valueBytes,
	})
	defer eng.Close()

	b.SetBytes(rangeBytes)
	b.ResetTimer()

	var stats enginepb.MVCCStats
	var err error
	for i := 0; i < b.N; i++ {
		iter := eng.NewIterator(IterOptions{UpperBound: roachpb.KeyMax})
		stats, err = iter.ComputeStats(roachpb.KeyMin, roachpb.KeyMax, 0)
		iter.Close()
		if err != nil {
			b.Fatal(err)
		}
	}

	b.StopTimer()
	log.Infof(ctx, "live_bytes: %d", stats.LiveBytes)
}

// runMVCCCFindSplitKey benchmarks MVCCFindSplitKey on a 64MB range of data.
func runMVCCFindSplitKey(ctx context.Context, b *testing.B, emk engineMaker, valueBytes int) {
	const rangeBytes = 64 * 1024 * 1024
	numKeys := rangeBytes / (overhead + valueBytes)
	eng, _ := setupMVCCData(ctx, b, emk, benchDataOptions{
		numVersions: 1,
		numKeys:     numKeys,
		valueBytes:  valueBytes,
	})
	defer eng.Close()

	b.SetBytes(rangeBytes)
	b.ResetTimer()

	var err error
	for i := 0; i < b.N; i++ {
		_, err = MVCCFindSplitKey(ctx, eng, roachpb.RKeyMin,
			roachpb.RKeyMax, rangeBytes/2)
		if err != nil {
			b.Fatal(err)
		}
	}

	b.StopTimer()
}

type benchGarbageCollectOptions struct {
	benchDataOptions
	keyBytes       int
	deleteVersions int
}

func runMVCCGarbageCollect(
	ctx context.Context, b *testing.B, emk engineMaker, opts benchGarbageCollectOptions,
) {
	rng, _ := randutil.NewPseudoRand()
	eng := emk(b, "mvcc_gc")
	defer eng.Close()

	ts := hlc.Timestamp{}.Add(time.Date(2000, 1, 1, 0, 0, 0, 0, time.UTC).UnixNano(), 0)
	val := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, opts.valueBytes))

	// We write values at ts+(0,i), set now=ts+(1,0) so that we're ahead of all
	// the writes. This value doesn't matter in practice, as it's used only for
	// stats updates.
	now := ts.Add(1, 0)

	// Write 'numKeys' of the given 'keySize' and 'valSize' to the given engine.
	// For each key, write 'numVersions' versions, and add a GCRequest_GCKey to
	// the returned slice that affects the oldest 'deleteVersions' versions. The
	// first write for each key will be at `ts`, the second one at `ts+(0,1)`,
	// etc.
	//
	// NB: a real invocation of MVCCGarbageCollect typically has most of the keys
	// in sorted order. Here they will be ordered randomly.
	setup := func() (gcKeys []roachpb.GCRequest_GCKey) {
		batch := eng.NewBatch()
		for i := 0; i < opts.numKeys; i++ {
			key := randutil.RandBytes(rng, opts.keyBytes)
			if opts.deleteVersions > 0 {
				gcKeys = append(gcKeys, roachpb.GCRequest_GCKey{
					Timestamp: ts.Add(0, int32(opts.deleteVersions-1)),
					Key:       key,
				})
			}
			for j := 0; j < opts.numVersions; j++ {
				if err := MVCCPut(ctx, batch, nil /* ms */, key, ts.Add(0, int32(j)), val, nil); err != nil {
					b.Fatal(err)
				}
			}
		}
		if err := batch.Commit(false); err != nil {
			b.Fatal(err)
		}
		batch.Close()
		return gcKeys
	}

	gcKeys := setup()

	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		batch := eng.NewWriteOnlyBatch()
		distinct := batch.Distinct()
		if err := MVCCGarbageCollect(ctx, distinct, nil /* ms */, gcKeys, now); err != nil {
			b.Fatal(err)
		}
		distinct.Close()
		batch.Close()
	}
}

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

	eng := emk(b, fmt.Sprintf("batch_apply_batch_repr_%d_%d", valueSize, batchSize))
	defer eng.Close()

	var repr []byte
	{
		order := make([]int, batchSize)
		for i := range order {
			order[i] = i
		}
		if !sequential {
			rng.Shuffle(len(order), func(i, j int) {
				order[i], order[j] = order[j], order[i]
			})
		}

		batch := eng.NewWriteOnlyBatch()
		defer batch.Close() // NB: hold open so batch.Repr() doesn't get reused

		for i := 0; i < batchSize; i++ {
			key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(order[i])))
			ts := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
			if err := MVCCBlindPut(ctx, batch, nil, key, ts, value, nil); err != nil {
				b.Fatal(err)
			}
		}
		repr = batch.Repr()
	}

	b.SetBytes(int64(len(repr)))
	b.ResetTimer()

	for i := 0; i < b.N; i++ {
		var batch Batch
		if !indexed {
			batch = eng.NewWriteOnlyBatch()
		} else {
			batch = eng.NewBatch()
		}
		if err := batch.ApplyBatchRepr(repr, false /* sync */); err != nil {
			b.Fatal(err)
		}
		if r, ok := batch.(*rocksDBBatch); ok {
			// Ensure mutations are flushed for RocksDB indexed batches.
			r.flushMutations()
		}
		batch.Close()
	}

	b.StopTimer()
}

func runExportToSst(
	ctx context.Context,
	b *testing.B,
	emk engineMaker,
	numKeys int,
	numRevisions int,
	exportAllRevisions bool,
	contention bool,
) {
	dir, cleanup := testutils.TempDir(b)
	defer cleanup()
	engine := emk(b, dir)
	defer engine.Close()

	batch := engine.NewWriteOnlyBatch()
	for i := 0; i < numKeys; i++ {
		key := make([]byte, 16)
		key = append(key, 'a', 'a', 'a')
		key = encoding.EncodeUint32Ascending(key, uint32(i))

		for j := 0; j < numRevisions; j++ {
			err := batch.Put(MVCCKey{Key: key, Timestamp: hlc.Timestamp{WallTime: int64(j + 1), Logical: 0}}, []byte("foobar"))
			if err != nil {
				b.Fatal(err)
			}
		}
	}
	if err := batch.Commit(true); err != nil {
		b.Fatal(err)
	}
	batch.Close()
	if err := engine.Flush(); err != nil {
		b.Fatal(err)
	}

	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		startTS := hlc.Timestamp{WallTime: int64(numRevisions / 2)}
		endTS := hlc.Timestamp{WallTime: int64(numRevisions + 2)}
		_, _, _, err := engine.ExportToSst(roachpb.KeyMin, roachpb.KeyMax, startTS, endTS, exportAllRevisions, 0 /* targetSize */, 0 /* maxSize */, IterOptions{
			LowerBound: roachpb.KeyMin,
			UpperBound: roachpb.KeyMax,
		})
		if err != nil {
			b.Fatal(err)
		}
	}
	b.StopTimer()
}