The questions that matter: is history point-in-time, are delisted securities included, how are corporate actions handled, what is the timestamp convention, how often is history silently revised, and does the live feed match the historical one in format and content.
That last point causes more live failures than any other. Research on one vendor's cleaned daily bars, executed on a broker's raw real-time feed, means the numbers driving live decisions are not the numbers the model was fitted on.
Cheap and free sources are usable for learning and for ideas, but they typically lack delisted names, restate history, and vary in adjustment methodology. Know which limitation applies before you use a result to size real risk.
Related: point-in-time-data, survivorship-bias, data-quality-check, live-vs-backtest-divergence