We want to define a new builtin metric to periodically count the number of RPCs with error per connection, for better diagnostics.