Fixes #37152.

🔢

💭

The Manual benchmark is now very close to what it used to be in 9. The 2% difference in speed (some runs showed 5%, the Error/StdDev here is pretty high) is the "price" for now much more nicer parameter names.

The Contains benchmark is not comparing 🍎 and 🍎. In 9, the query was translated to OPENJSON, while in 10 this is translated to multiple parameters. Before the fix the translation was very slow (~188x) and created a lot of allocations (~448x). Now the numbers are ~5x and ~10x. Again, this is comparing 1 parameter vs 1000 parameters - 🍎 vs 🍊.

Fair to say, I know about a decent room for improvement in allocations (hence theoretically in speed as well, at least in benchmarks). But it is too risky for servicing. I might do it for 11.

It is also important to know that this benchmark measures only the translation. Once the query hits the database, the translation to multiple parameters should show its real benefits.

Description

EF 10 changed the default translation for parameterized collections from single parameter and JSON function to multiple parameters. We missed O(n^2) performance degradation in code path where we generate unique parameter names resulting in subpar performance.

Customer impact

Query translation takes significant time compared to other types of translation. Excessive allocations.

How found

Customer reported on 10.0.

Regression

Yes.

Testing

The change in the PR is covered by existing tests. We don't have specific benchmarks for translations.

Risk

Low. Added fast path optimization. Quirk added.