bench: lower attach concurrency

[koivunej]

it has been observed that we get transaction serialization issues otherwise.

evidence: https://neon-github-public-dev.s3.amazonaws.com/reports/main/10402057346/index.html#suites/c62b105f3a4f00dd6be4ad88810e0e02/7c03011d9dbd5ec4/

Slack thread: https://neondb.slack.com/archives/C060CNA47S9/p1723722430992019

Alternatively we could:

• first issue attach hooks sequentially then attach concurrently
• just not do it at all (unsure why we even bother with overriding the generations)

[problame]

I remember adding the concurrent attach because it was one of the slowest parts of the setup code at the time.

The serialization failures in storcon db: why? aren't these separate rows for each tenant?
Smells like a suboptimal SQL query in storcon to me, wouldn't we hit the same problem in prod?

[problame]

Commented

[koivunej]

The serialization failures in storcon db: why? aren't these separate rows for each tenant?

Yes, but it might be that uniqueness constraint on tenant gets to witness all or some subset of other rows, maybe?

Smells like a suboptimal SQL query in storcon to me, wouldn't we hit the same problem in prod?

I don't think we do large scale generation bumps in prod, but possibly yes. Though the attach-hook mentions it's testing-only hook.

[koivunej]

The serialization failures in storcon db: why? aren't these separate rows for each tenant?

Yes, but it might be that uniqueness constraint on tenant gets to witness all or some subset of other rows, maybe?

refreshing on https://www.postgresql.org/docs/current/transaction-iso.html it would appear that we should first query instead of directly insert (and run into constraint violation). Let's fix it that way.

[github-actions]

2252 tests run: 2173 passed, 0 failed, 79 skipped (full report)

---

Code coverage* (full report)

• functions: 32.4% (7220 of 22266 functions)
• lines: 50.4% (58300 of 115727 lines)

* collected from Rust tests only

---

_{The comment gets automatically updated with the latest test results
b6a2c39 at 2024-08-15T14:21:20.589Z :recycle:}

[koivunej]

Discussed this PR, nothe better protocol of check + insert shouldn't be introduced with a test change.