storage: Add rocksdb-vs-pebble benchmark for ExportToSst

[itsbilal]

As part of the investigation into #49710, this change adds a
benchmark for ExportToSst that tests both RocksDB and Pebble.

Here are some example runs without contention (old = rocksdb,
new = pebble):

name                                                   old time/op  new time/op  delta
ExportToSst/rocksdb/numKeys=64/numRevisions=1-12       43.9µs ± 3%  34.5µs ± 4%  -21.33%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=64/numRevisions=10-12       281µs ± 3%   169µs ± 6%  -39.89%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=64/numRevisions=100-12     1.82ms ±22%  1.17ms ± 1%  -35.73%  (p=0.000 n=10+9)
ExportToSst/rocksdb/numKeys=512/numRevisions=1-12       212µs ± 6%   111µs ± 3%  -47.77%  (p=0.000 n=10+9)
ExportToSst/rocksdb/numKeys=512/numRevisions=10-12     1.91ms ± 1%  1.19ms ± 8%  -37.65%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=512/numRevisions=100-12    13.7ms ± 3%  10.1ms ±12%  -26.21%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=1024/numRevisions=1-12      390µs ± 1%   215µs ±12%  -44.94%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=1024/numRevisions=10-12    4.01ms ± 6%  2.40ms ±16%  -40.13%  (p=0.000 n=10+9)
ExportToSst/rocksdb/numKeys=1024/numRevisions=100-12   27.9ms ± 2%  20.8ms ± 2%  -25.48%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=8192/numRevisions=1-12     2.97ms ± 2%  1.42ms ± 5%  -52.24%  (p=0.000 n=9+10)
ExportToSst/rocksdb/numKeys=8192/numRevisions=10-12    32.8ms ± 7%  19.1ms ± 3%  -41.59%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=8192/numRevisions=100-12    224ms ± 3%   169ms ±25%  -24.64%  (p=0.000 n=9+10)
ExportToSst/rocksdb/numKeys=65536/numRevisions=1-12    23.7ms ± 4%  13.4ms ±20%  -43.65%  (p=0.000 n=9+10)
ExportToSst/rocksdb/numKeys=65536/numRevisions=10-12    264ms ± 4%   201ms ±24%  -23.92%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=65536/numRevisions=100-12   1.88s ± 6%   1.23s ± 8%  -34.70%  (p=0.000 n=10+8)

And some with contention=true:

name                                                                   old time/op  new time/op  delta
ExportToSst/rocksdb/numKeys=65536/numRevisions=10/contention=true-12    362ms ± 7%   168ms ± 3%  -53.60%  (p=0.000 n=10+10)
ExportToSst/rocksdb/numKeys=65536/numRevisions=100/contention=true-12   2.24s ± 6%   1.24s ±10%  -44.50%  (p=0.000 n=10+10)

Release note: None.

[cockroach-teamcity]

This change is 

[petermattis]

Thanks for getting to this so quickly. Too bad it didn't reveal a smoking gun. You should include RocksDB vs Pebble comparison numbers in the PR/commit messsage.

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @itsbilal, @pbardea, and @petermattis)

---

pkg/storage/bench_test.go, line 1009 at r1 (raw file):

		for j := 0; j < numRevisions; j++ {
			err := batch.Put(MVCCKey{key, hlc.Timestamp{int64(j), 1}}, []byte("foobar"))

In practice, most hlc.Timestamps have a 0 logical time.

---

pkg/storage/bench_test.go, line 1044 at r1 (raw file):

							}
						}
						engine.Flush()

This is a lot of flushing! I'm not quite sure what this contention bit is testing. Given that it isn't changing the results, perhaps we should just remove it.

---

pkg/storage/bench_test.go, line 1065 at r1 (raw file):

	b.StopTimer()
	if contention {
		closeCh <- struct{}{}

You can close(closeCh) once, and all the waiters will read from the closed channel. This allows you to change the creation to an unbuffered channel: closeCh := make(chan struct{}).

[itsbilal]

TFTR! Added some benchstat runs to the commit message and PR description.

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @pbardea and @petermattis)

---

pkg/storage/bench_test.go, line 1009 at r1 (raw file):

Previously, petermattis (Peter Mattis) wrote…

In practice, most hlc.Timestamps have a 0 logical time.

Done.

---

pkg/storage/bench_test.go, line 1044 at r1 (raw file):

Previously, petermattis (Peter Mattis) wrote…

This is a lot of flushing! I'm not quite sure what this contention bit is testing. Given that it isn't changing the results, perhaps we should just remove it.

Done.

---

pkg/storage/bench_test.go, line 1065 at r1 (raw file):

Previously, petermattis (Peter Mattis) wrote…

You can close(closeCh) once, and all the waiters will read from the closed channel. This allows you to change the creation to an unbuffered channel: closeCh := make(chan struct{}).

Done.

[petermattis]

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @itsbilal and @pbardea)

---

pkg/storage/bench_test.go, line 1044 at r1 (raw file):

Previously, itsbilal (Bilal Akhtar) wrote…

Done.

Ah, I meant that I think we could get rid of the whole contention section. It doesn't seem to be showing us anything beyond the non-contended case. I'm glad you made an attempt to see if "contention" had any impact on ExportToSst performance, but just because the code is written doesn't mean we need to keep it.

[pbardea]

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @itsbilal and @pbardea)

---

pkg/storage/bench_test.go, line 1055 at r2 (raw file):

		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, 0, IterOptions{

nit: I think comments around the 0 targetSize and maxSize args here would be helpful for readability

[itsbilal]

TFTRs!

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @itsbilal and @petermattis)

---

pkg/storage/bench_test.go, line 1044 at r1 (raw file):

Previously, petermattis (Peter Mattis) wrote…

Ah, I meant that I think we could get rid of the whole contention section. It doesn't seem to be showing us anything beyond the non-contended case. I'm glad you made an attempt to see if "contention" had any impact on ExportToSst performance, but just because the code is written doesn't mean we need to keep it.

Done. Removed.

---

pkg/storage/bench_test.go, line 1055 at r2 (raw file):

Previously, pbardea (Paul Bardea) wrote…

nit: I think comments around the 0 targetSize and maxSize args here would be helpful for readability

Done.

[petermattis]

Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @itsbilal)

[itsbilal]

TFTRs!

bors r+

[craig]

Build succeeded

• GitHub CI (Cockroach)