We report on our ongoing effort to develop observational debuggers for SQL. This debugging paradigm—in which the evaluation of selected subexpressions may be "spied on"—fits the nature of query languages, but may lead to observations whose size can overwhelm users. Here, we tackle this challenge with the help of data provenance analysis. The analysis identifies exactly those input rows that are material in producing suspect query outputs. Running the debugger on such a minimized input will exclusively yield observations that are indeed relevant in understanding the bug.