A system and method for selection of a provenance dependency function in a stream-based data processing infrastructure to optimize backtracing performance in response to a provenance query includes determining performance of a set of dependency functions associated with an analysis component for determining relevancy of each input event received by the analysis component. The relevancy of each input event is determined according to each dependency function and storing a record of relevant events according to a recording method. Relevancy results of the dependency functions are aggregated, and the dependency functions are ordered according to a criterion. Data provenance is provided for a given output event using the input event recording method associated with a best dependency function according to the criterion.