Dapr vs Skaffold

Dapr maintains 1.16, 1.17 and 1.18 concurrently, and the current window is entirely bug fixes backported across all three. The 1.18.3 release carries fifteen of them; the older branches receive the subset that applies. Workflow durability dominates — stalled workflows left unrecoverable after the last worker disconnected, terminate events silently dropped when batched, orphaned activity-result reminders retrying forever, and continue_as_new iterations sharing one unbounded trace. The 1.16 line has now opened a 1.16.20 candidate carrying a single placement reconnect fix.

The failure reports are notably specific about who was affected and under what configuration, and several describe components that looked healthy while silently doing nothing — input bindings that never activated because a warmup probe had a hardcoded three-second budget, an Azure credential chain that stopped at SPIFFE instead of falling back. That class of bug is what a maturing distributed runtime finds once the obvious crashes are gone. Release candidates are published openly before each patch, so the same fixes appear several times in the feed, and the newest candidate shows the oldest supported branch still receiving actor and placement corrections.

Expect 1.16.20 to ship as a final shortly and further patches across all three branches, with actor lifecycle and workflow recovery paths the likeliest sources given where this window's fixes cluster.