MS-930 Implement stateful face matcher that reduces amount of allocations

[luhmirin-s]

JIRA ticket
Will be released in: 2025.2.0

This is an experimental implementation of the suggestion in the JIRA ticket, it could be polished to be production-ready if required.

Notable changes

• Changing the face matches to pre-allocate native sample template memory and reuse it for each candidate. In theory, it should reduce memory and GC pressure.
• Still using a separate allocation per coroutine to avoid cross-coroutine memory sharing and potential concurrency issues. There is still a significant reduction in allocation calls.

Surprisingly the change is way less noticeable in my very basic benchmarking (running identification on 50k candidates using ROCv3 on a Lenovo Tab 8):

Before change	After change
114780ms	92866ms
113032ms	110177ms
107323ms	110511ms
110483ms	95675ms
111383ms	104624ms
113065ms	109125ms
112612ms	110679ms
Avg: 111811	Avg: 104808

As part of this task, I also investigated the feasibility of the same refactoring for fingerprints and decided against it for the following reasons:

• SimAfis matching supports 1:N on the library level and takes the candidate list as a parameter, therefore it is highly likely that the same allocation optimisation is happening under the hood (or at least it should be done on the library level).
• NEC wraps templates with a custom data class and there is no way to control allocations. Therefore even if we cache the template list, it would not prevent unnecessary allocations.
• Fingerprint BioSDK setup is significantly more complex and would require a lot of changes for no/negligible benefit.

Additional work checklist

• Effect on other features and security has been considered
• Design document marked as "In development" (if applicable)
• External (Gitbook) and internal (Confluence) Documentation is up to date (or ticket created)
• Test cases in Testiny are up to date (or ticket created)
• Other teams notified about the changes (if applicable)

[sonarqubecloud]

Quality Gate passed

Issues
2 New issues
0 Accepted issues

Measures
0 Security Hotspots
88.2% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud