[Mongo] Clear the streamID from the request/response maps after stitc…

…hing (#1878)

Summary: This PR modifies mongo's stitcher logic to clear streamID’s
from request/response maps once all frames of a streamID have been
consumed. We observed high CPU use allocated to `FramesSize()` &
`EraseExpiredFrames()`, this was due to the size of the maps increasing
with new streamIDs and having to continuously loop over those growing
maps to cleanup. Clearing the streamID's from the maps after stitching
significantly reduces the CPU allocated to the cleanup logic for mongo,
the exact details for mongo's streamID reuse needs to be determined to
further adapt the stitcher logic.

Type of change: /kind bug

Test Plan: Existing tests still pass, ran the px-mongo demo and observed
lower CPU allocated to mongo in the PEM through flamegraph.

---------

Signed-off-by: Kartik Pattaswamy <[email protected]>

src/stirling/source_connectors/socket_tracer/protocols/mongodb/stitcher.cc

@@ -183,7 +183,13 @@ RecordsWithErrorCount<mongodb::Record> StitchFrames(
       ++erase_until_iter;
     }
 
-    req_deque.erase(req_deque.begin(), erase_until_iter);
+    // Determine whether to clear the StreamID from the map or only the frames that have been
+    // consumed.
+    if (erase_until_iter == req_deque.end()) {
+      reqs->erase(stream_id);
+    } else {
+      req_deque.erase(req_deque.begin(), erase_until_iter);
+    }
     stream_id_pair.second = true;
   }
 
@@ -195,9 +201,11 @@ RecordsWithErrorCount<mongodb::Record> StitchFrames(
         error_count++;
       }
     }
-    resp_deque.clear();
   }
 
+  // Clear the response map.
+  resps->clear();
+
   // Clear the state.
   auto it = state->stream_order.begin();
   while (it != state->stream_order.end()) {