[BugFix] Fix pindex read-write concurrency between commit and apply (#…

…33220)

During the execution of apply on the primary key table, the apply thread maybe clear or modify _l1_vec of persistent index.
And during the execution commit, the commit thread will update primary index memory usage and need to access _l1_vec.
So there are read/write conflicts between apply thread and commit thread which may cause BE crash. And we add _lock to prevent the read-write concurrency in previous pr but there are some scenarios where locks were inadvertently missed.

Signed-off-by: zhangqiang <qiangzh95@gmail.com>
(cherry picked from commit 6bf3e88)

be/src/storage/persistent_index.cpp

@@ -4167,14 +4167,15 @@ Status PersistentIndex::_merge_compaction_advance() {
     RETURN_IF_ERROR(writer->finish());
     std::vector<std::unique_ptr<ImmutableIndex>> new_l1_vec;
     std::vector<int> new_l1_merged_num;
-    size_t merge_num = _l1_merged_num[merge_l1_start_idx];
-    for (int i = 0; i < merge_l1_start_idx; i++) {
-        new_l1_vec.emplace_back(std::move(_l1_vec[i]));
-        new_l1_merged_num.emplace_back(_l1_merged_num[i]);
-    }
-
+    size_t merge_num = 0;
     {
         std::unique_lock wrlock(_lock);
+        merge_num = _l1_merged_num[merge_l1_start_idx];
+        for (int i = 0; i < merge_l1_start_idx; i++) {
+            new_l1_vec.emplace_back(std::move(_l1_vec[i]));
+            new_l1_merged_num.emplace_back(_l1_merged_num[i]);
+        }
+
         for (int i = merge_l1_start_idx; i < _l1_vec.size(); i++) {
             _l1_vec[i]->destroy();
         }