Good write up! Of course.

I am an ardent follower of Pat Helland as well. Many of the potential pitfalls, I have been able to avoid, by simply following the fundamentals of distribution, elucidated by him (and some others).

Adoption of Saga pattern can be devious in operations, especially the UNDO steps, if and when necessary. We are essentially implementing the Transaction logs of traditional DBMS systems, across cooperating processes, which are only connected through a network.

The point of isolation doesn't apply, in case of related, persistent, datasets which must be consistent, post-db-transactions.

In my view, it is safer to keep DB transaction local to a service (and therefore to a DB), to the extent possible (Nader seems to suggest the same thing, anyway). It is better to replicate data and reconcile later through a different process. This helps us avoid, devising unique IDs and then work towards finding these IDs across a number of isolated services, in order to push compensation.