I find that we use CRDT's quite often in the form of append-only sets (formally a G-Set). Rather than choose the CRDT specifically, we've seen that an append-on

I find that we use CRDT's quite often in the form of append-only sets (formally a G-Set). Rather than choose the CRDT specifically, we've seen that an append-only set was the natural choice for a particular need and then we identified the commutative and idempotent properties of the operation. Once identified, we leverage them for scale. One good example is processing items in an SQS queue, where (without enabling FIFO) there is no guarantee of ordering. With a G-Set you can read and merge batches of items (and retry naively) without affecting the outcome.