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

* MOD-14826 Use try_lock in debugInfo() to avoid blocking main thread during SVS training (#931)

* use try lock

* fix test

(cherry picked from commit d7a5afb)

* remove assert

---------

Co-authored-by: meiravgri <109056284+meiravgri@users.noreply.github.com>
Co-authored-by: meiravgri <meirav.grimberg@redis.com>

Author: github-actions[bot]

src/VecSim/algorithms/svs/svs_tiered.h

@@ -832,9 +832,19 @@ class TieredSVSIndex : public VecSimTieredIndex<DataType, float> {
             .updateJobWaitTime = this->updateJobWaitTime,
         };
         {
-            std::lock_guard<std::mutex> lock(this->updateJobMutex);
-            svsTieredInfo.indexUpdateScheduled =
-                this->indexUpdateScheduled.test() == VecSimBool_TRUE;
+            // Use try_lock to avoid blocking the main thread during long-running
+            // training operations. updateSVSIndexWrapper holds updateJobMutex for
+            // the entire training duration (which can take 40-85s on slow machines).
+            // If the mutex is held, training is actively running, so we report
+            // indexUpdateScheduled = true (BACKGROUND_INDEXING = 1).
+            std::unique_lock<std::mutex> lock(this->updateJobMutex, std::try_to_lock);
+            if (lock.owns_lock()) {
+                svsTieredInfo.indexUpdateScheduled =
+                    this->indexUpdateScheduled.test() == VecSimBool_TRUE;
+            } else {
+                // Mutex is held by updateSVSIndexWrapper — training is in progress.
+                svsTieredInfo.indexUpdateScheduled = true;
+            }
         }
         info.tieredInfo.specificTieredBackendInfo.svsTieredInfo = svsTieredInfo;
         info.tieredInfo.backgroundIndexing =

tests/flow/test_svs_tiered.py

@@ -235,32 +235,24 @@ def search_insert(test_logger, is_multi: bool, num_per_label=1, data_type=VecSim
     searches_number = 0
     # run knn query every 1 s.
     total_tiered_search_time = 0
-    prev_bf_size = num_labels
     cur_svs_label_count = index.svs_label_count()
 
     test_logger.info(f"SVS labels number = {cur_svs_label_count}")
     while searches_number == 0 or cur_svs_label_count < num_labels - updateThreshold:
         # For each run get the current svs size and the query time.
-        bf_curr_size = index.get_curr_bf_size()
         query_start = time.time()
         tiered_labels, _ = index.knn_query(query_data, k)
         query_dur = time.time() - query_start
         total_tiered_search_time += query_dur
 
         test_logger.info(f"query time = {round_ms(query_dur)} ms")
 
-        # BF size should decrease.
-        test_logger.info(f"bf size = {bf_curr_size}")
-        assert bf_curr_size < prev_bf_size
-
         # Run the query also in the bf index to get the ground truth results.
         bf_labels, _ = bf_index.knn_query(query_data, k)
         correct += len(np.intersect1d(tiered_labels[0], bf_labels[0]))
         time.sleep(1)
         searches_number += 1
-        prev_bf_size = bf_curr_size
         cur_svs_label_count = index.svs_label_count()
-
     # SVS labels count updates before the job is done, so we need to wait for the queue to be empty.
     index.wait_for_index(1)
     index_dur = time.time() - index_start