GenStage is perfect for this kind of workload, as it allows the program to be throttled using back-pressure based on what your target system can take (as well as how fast your data source(s) can provide). If you can partition the source data and your transform steps are a bottleneck (as opposed to the extract or load being the bottlenecks due to capacity at either source or destination), then you can also build your application to be clustered, allowing for horizontal scaling of the ETL workload.
Moreover, Elixir makes it trivial to run multiple work orders at the same time, spreading the work out across however many cores you have in the machine running your ETL application.
That said: I would expect that writing this application in Elixir would not provide for the absolute fastest execution times for a single-core, single-threaded approach when compared to other languages available out there (C++, Java, ..), but for concurrency (local multi-core and/or multi-system distribution), durability (fault tollerance), and developer productivity it is really hard to beat for these sorts of applications IME.
