Likeliness to change is certainly a factor, but I think the level to which you are agnostic to the particular technology is more important. If I am choosing the

Likeliness to change is certainly a factor, but I think the level to which you are agnostic to the particular technology is more important. If I am choosing the technology to provide functionality that can be provided with pretty much the same semantics by other technologies, then I think it often makes sense to create the abstraction so that they can be swapped out. On the other hand, if I am choosing technology for semantics that are particular to that technology, I don't think creating the extra abstraction makes sense. For instance, caching is functionality that could be backed by redis, memcached, etc. and I think it makes sense to create a cache API in most cases. On the other hand, if I am processing data with Kafka, I want to interface with Kafka directly for all the special functionality and nuance that makes it great for particular use-cases.