Sawzall as a language is quite a bit uglier than the vast majority of general purpose languages. Couple this with the read/emit nature of the language, and it's either useful as a stream processing language, or as a step in a mapreduce chain.
Given how easy other languages are at processing streams, tagging output, etc., and that Sawzall doesn't really have an idea of shared state between "records" (aside from data emitted), it's hard to find things that Sawzall is good at other than mapreduce.
Or I could use any one of a dozen other languages that are more convenient to use, already available on my system, already works with S4, and with a syntax that doesn't make me want to cry :P
Given how easy other languages are at processing streams, tagging output, etc., and that Sawzall doesn't really have an idea of shared state between "records" (aside from data emitted), it's hard to find things that Sawzall is good at other than mapreduce.