Question

After each rolling deploy, freshly started pods show a 20-30 second window of elevated p99 and a burst of `connection reset` / handshake-timeout errors to your database and a downstream gRPC service, then settle. It's not GC or JIT — profiling shows the time going into establishing new TCP+TLS connections and the DB driver doing per-connection auth. The connection pools are configured lazy (`min-idle = 0`), so a fresh pod opens connections only as the first requests arrive — right when the LB hits it at full traffic. How do you triage and fix this cold-start?