Deciding whether a boundary should be a process boundary, and designing it when it must be: what distribution actually costs (latency, serialisation, partial failure, lost atomicity, independent deployment), caller-driven remote granularity, and choosing between synchronous call, messaging and replication.
