MOD-14826 Use try_lock in debugInfo() to avoid blocking main thread during SVS training

[meiravgri]

TieredSVSIndex::debugInfo() used a blocking std::lock_guard on updateJobMutex. Since updateSVSIndexWrapper holds that mutex for the entire duration of SVS training (which can take tens of seconds), any call to debugInfo() — including via svs_label_count() in the Python bindings — would block until training completed.

This means that reading debug info (e.g., BACKGROUND_INDEXING, INDEX_UPDATE_SCHEDULED) could stall the calling thread for the full training duration, defeating the purpose of non-blocking status queries.

Fix

Replace std::lock_guard with std::unique_lock using std::try_to_lock:

• If the lock is acquired: read indexUpdateScheduled normally.
• If the lock is not acquired (training is in progress): report indexUpdateScheduled = true, since the mutex being held by updateSVSIndexWrapper means an update is actively running.

Flow test update

The search_insert flow test was implicitly relying on the old blocking behavior — svs_label_count() (which calls debugInfo()) acted as an accidental synchronization point, blocking until background indexing finished. With the fix, debugInfo() returns immediately, so the test could observe intermediate states (e.g., flat buffer not yet drained while SVS already has vectors).

Updated the test to remove the per-iteration assert bf_curr_size < prev_bf_size check and instead assert bf_size == 0 after wait_for_index() completes.

Mark if applicable

• This PR introduces API changes
• This PR introduces serialization changes

---

Note

Low Risk
Low risk: changes only debug/observability behavior and a flow test assertion, without affecting indexing or query logic. Main risk is slightly altered BACKGROUND_INDEXING reporting while training is in progress.

Overview
Prevents TieredSVSIndex::debugInfo() from blocking for the full SVS training duration by switching the updateJobMutex acquisition to try_to_lock; when the mutex is held (training running), it reports indexUpdateScheduled=true to keep BACKGROUND_INDEXING on.

Updates test_search_insert to stop asserting the flat-buffer size monotonically decreases during background indexing, and instead asserts the buffer is fully drained (get_curr_bf_size()==0) after wait_for_index() completes.

^{Reviewed by Cursor Bugbot for commit d9edf21. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jit-ci]

🛡️ Jit Security Scan Results

✅ No security findings were detected in this PR

---

^{Security scan by Jit}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.99%. Comparing base (32cb7f2) to head (d9edf21).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #931      +/-   ##
==========================================
- Coverage   97.01%   96.99%   -0.03%     
==========================================
  Files         129      129              
  Lines        7574     7576       +2     
==========================================
  Hits         7348     7348              
- Misses        226      228       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ofiryanai]

please un-flak

[github-actions]

Successfully created backport PR for 8.2:

• [8.2] MOD-14826 Use try_lock in debugInfo() to avoid blocking main thread during SVS training #932

[github-actions]

Successfully created backport PR for 8.4:

• [8.4] MOD-14826 Use try_lock in debugInfo() to avoid blocking main thread during SVS training #933