Mogged by microservices

Gist from the article:

centralizing migration to one team for maximum coordination
avoid decentralizing migrations to owners - it stalls deployments indefinitely. centralizing it through a queue and jobs (with team notification enabled) streamlines it a lot better, takes less time and scales better (while also having retries available).
slow rollout, and quick rollback
not everything should be automated (80/20), rest case by case basis.
wrap up old sdk interfaces with an in-house wrapper sdk, use tools to prevent people using old sdk directly - anything should be invoked through in-house wrapper. Also, in-house wrapper can switch bw old sdk and new sdk based on config (easier rollback without redeploy).
targetting least critical first (labels) for rollout, then working through the critical systems later.
using rollback rules in argocd (through prometheus metrics - RPC errors, crashes / panics, database errors)
use centralized config store (that serivices talk to directly, not configmap or secrets) for enabling / turning off features.
stripping out legacy code.