Read the Performance section of Index Selection article:
The actual performance depends on the shape of the data. Specifically, the average number of entities considered for each result returned is O(S/R). This indicates that poor performance is likely when many entities match each scan, but few entities match the query as a whole (R is small and S is large).
As article notes this only affects normal indexes. If you want O(log n) speed you should define a composite index.