roachtest: add read committed variants of ycsb

[nvanbenschoten]

Closes #107112.

This PR adds the following six roachtest variants:

ycsb/A/nodes=3/cpu=32/isolation-level=read-committed
ycsb/B/nodes=3/cpu=32/isolation-level=read-committed
ycsb/C/nodes=3/cpu=32/isolation-level=read-committed
ycsb/D/nodes=3/cpu=32/isolation-level=read-committed
ycsb/E/nodes=3/cpu=32/isolation-level=read-committed
ycsb/F/nodes=3/cpu=32/isolation-level=read-committed

It does so after adding an --isolation-level flag to ycsb, which controls the isolation level to run the workload transactions under. If unset, the workload will run with the default isolation level of the database.

Release note: None

[cockroach-teamcity]

This change is 

[nvanbenschoten]

I'll kick off some iterations of each workload to compare vs. serializable.

[nvanbenschoten]

Here is the result of the comparison:

                      │ ycsb_ssi.txt │               ycsb_rc.txt               │
                      │    ops/s     │     ops/s       vs base                 │
ycsb/A/nodes=3/cpu=32   30.97k ± ∞ ¹    31.92k ± 6%          ~ (p=0.662 n=5+6)
ycsb/B/nodes=3/cpu=32   82.73k ± ∞ ¹    91.95k ±  ∞ ¹  +11.16% (p=0.016 n=5)
ycsb/C/nodes=3/cpu=32   99.24k ± ∞ ¹   103.40k ± 6%          ~ (p=0.931 n=5+6)
ycsb/D/nodes=3/cpu=32   85.23k ± ∞ ¹    96.24k ± 2%          ~ (p=0.126 n=5+6)
ycsb/E/nodes=3/cpu=32   6.386k ± ∞ ¹    7.080k ± 1%          ~ (p=0.126 n=5+6)
ycsb/F/nodes=3/cpu=32   17.17k ± ∞ ¹    30.24k ± 3%    +76.11% (p=0.004 n=5+6)

                      │ ycsb_ssi.txt │              ycsb_rc.txt              │
                      │   avg(ms)    │   avg(ms)     vs base                 │
ycsb/A/nodes=3/cpu=32    4.600 ± ∞ ¹   4.500 ± 4%          ~ (p=0.771 n=5+6)
ycsb/B/nodes=3/cpu=32    2.300 ± ∞ ¹   2.100 ±  ∞ ¹   -8.70% (p=0.048 n=5)
ycsb/C/nodes=3/cpu=32    1.900 ± ∞ ¹   1.900 ± 5%          ~ (p=0.732 n=5+6)
ycsb/D/nodes=3/cpu=32    1.700 ± ∞ ¹   1.500 ± 0%          ~ (p=0.065 n=5+6)
ycsb/E/nodes=3/cpu=32    22.60 ± ∞ ¹   20.35 ± 1%          ~ (p=0.113 n=5+6)
ycsb/F/nodes=3/cpu=32    8.400 ± ∞ ¹   4.800 ± 4%    -42.86% (n=5+6)

                      │ ycsb_ssi.txt │              ycsb_rc.txt               │
                      │   p99(ms)    │    p99(ms)     vs base                 │
ycsb/A/nodes=3/cpu=32    54.50 ± ∞ ¹   67.10 ±  9%    +23.12% (p=0.017 n=5+6)
ycsb/B/nodes=3/cpu=32    12.60 ± ∞ ¹   10.50 ±   ∞ ¹  -16.67% (p=0.008 n=5)
ycsb/C/nodes=3/cpu=32    8.100 ± ∞ ¹   7.900 ± 13%          ~ (p=0.905 n=5+6)
ycsb/D/nodes=3/cpu=32    6.800 ± ∞ ¹   6.150 ±  2%          ~ (p=0.082 n=5+6)
ycsb/E/nodes=3/cpu=32    67.10 ± ∞ ¹   62.90 ±  3%          ~ (p=0.182 n=5+6)
ycsb/F/nodes=3/cpu=32   130.00 ± ∞ ¹   54.50 ± 12%    -58.08% (p=0.004 n=5+6)

The largest difference is from ycsb F (50% read, 50% read-modify-write). That's likely due to a combination of two factors:

1. this PR passes --read-modify-write-in-txn=false for read committed. I think I'm going to switch that to be the default even for serializable. We've seen external users of our YCSB benchmark get tripped up on this difference vs. the official benchmark.
2. reads under RC don't block on exclusive locks. This eliminates blocking between SFU locks in the read-modify-write transaction and non-locking readers.

We also see a meaningful improvement in ycsb A (50% read, 50% update) and ycsb B (95% read, 5% update). I think this is also due to RC reads not blocking on exclusive locks. This eliminates blocking between the implicit SFU locks acquired by the updates and the non-locking readers.

[nvanbenschoten]

I think I'm going to switch that to be the default even for serializable. We've seen external users of our YCSB benchmark get tripped up on this difference vs. the official benchmark.

This change is in #107537.

[nvanbenschoten]

I updated this PR to reflect the change to the default value of --read-modify-write-in-txn in #107537. @michae2 PTAL when you get a chance.

[michae2]

Nice!

Reviewed 2 of 2 files at r1, 1 of 1 files at r3, all commit messages.
Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @herkolategan, @nvanbenschoten, and @smg260)

---

pkg/workload/connection.go line 123 at r1 (raw file):

	// NOTE: validation of the isolation level value is done by the server during
	// connection establishment.
	return setUrlParam(urls, "default_transaction_isolation", isoLevel)

If isoLevel is snapshot, should this also set sql.txn.snapshot_isolation.enabled?

[nvanbenschoten]

TFTR!

bors r=michae2

Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @herkolategan, @michae2, and @smg260)

---

pkg/workload/connection.go line 123 at r1 (raw file):

Previously, michae2 (Michael Erickson) wrote…

If isoLevel is snapshot, should this also set sql.txn.snapshot_isolation.enabled?

If there was a session variable to enable snapshot isolation, I think it would be a good idea to set it here. As is, we only have the cluster setting, and we generally don't like workloads silently manipulating cluster settings, so I'll leave for now. Please let me know if you'd like me to revisit.

[craig]

Build succeeded:

• Bazel Essential CI (Cockroach)