Proposal

Try to send killCursors on client-side cursor timeout

Background & Motivation

The C driver deliberately does not re-establish connection to send killCursors to the server in _mongoc_client_kill_cursor.

CSOT suggests it is OK to extend the timeout to send killCursors:

If a cursor times out client-side (e.g. a non-tailable cursor created with timeoutMode=CURSOR_LIFETIME), it’s imperative that drivers make a good-faith effort to close the server-side cursor even though the timeout has expired because failing to do so would leave resources open on the server for a potentially long time. It was decided that timeoutMS will be refreshed for close operations to allow the cursor to be killed server-side.

CSOT is a large project and not-yet implemented in the C driver. This ticket proposes a smaller change to improve a timeout caused by socketTimeoutMS.

Here is a failing test that shows killCursors is not sent after a client-side timeout (even though the cursor ID was returned by the server):

mongoc_cursor_t *cursor =
   mongoc_collection_aggregate(coll, MONGOC_QUERY_NONE, pipeline, tmp_bson("{'batchSize': 1 }"), NULL);

// First document returned fast:
ASSERT(mongoc_cursor_next(cursor, &got));
ASSERT(mongoc_cursor_get_id(cursor)); // Cursor established.

// Second document triggers timeout:
ASSERT(!mongoc_cursor_next(cursor, &got));
ASSERT(mongoc_cursor_error(cursor, &error));
ASSERT_ERROR_CONTAINS(error, MONGOC_ERROR_STREAM, MONGOC_ERROR_STREAM_SOCKET, "socket error or timeout");

mongoc_cursor_destroy(cursor); // Does not send killCursors!