They help relieve pressure on the garbage collector by eliminating intermediate sequence allocations (eg. (comp (map a) (map b) (filter c) (map d)) would create 4 sequences, transducers create none). Consequently you can work with results incrementally unless there is an intermediate flush required such as for a windowed aggregation.

I imagine most applications aren't very sensitive to the performance gain, but it's good to know for when you need it. In addition it's a nice and testable, compos-able pattern. I've only done backend work, no cljs yet, but I'd call it reasonable practice to think in terms of compositions/transducers for most data transformation pipelines in the future though, throughout the stack, whether performance matters or not.