Limitations and considerations

--only-restart-stale-services requires --rolling-restart for the filter to actually take effect. In normal restart mode Cloudera Manager (CM) restarts every service regardless.
Rolling restart is not always available. If CM reports that the rolling-restart command is not currently available (a known CM condition), the operation falls back to a normal, non-rolling restart. This means a --rolling-restart request can still cause downtime. Treat the fallback as a rare event.
Single-instance roles cause per-role downtime even during a rolling restart. Design and size your cluster accordingly if uninterrupted availability is required.
Client connections are interrupted to the specific role instances recycled during a rolling restart. Client applications should be written to reconnect on failure.
Concurrent restart flows are rejected. CM will refuse to start a second restart while another one is in progress. Instead, the operation attaches to the existing one.