Should we run stateful services like Redis and Postgres on Kubernetes?

[Kyle_L]

Packing can make or break a trip, and I've learned through experience what works best. For beach vacations, I always pack multiple swimsuits and quick-dry towels. For city breaks, comfortable walking shoes are essential.

What are your go-to travel packing tips? I'm especially interested in advice for different climates and trip durations. Do you have any travel hacks 2025 that have made packing easier or more efficient? I'm working on a comprehensive travel guide that includes packing lists for various destinations.

[JamesCS]

I’m trying to decide if we should move our stateful services to Kubernetes, specifically our Redis cache and a Postgres read replica. The idea of managing persistent volumes, storage classes, and ensuring data durability in a self-healing, ephemeral container environment just makes me nervous about introducing complexity instead of reducing it.

[MadisonW]

We talked about Kubernetes and decided to keep Redis and the Postgres read replica as managed services for now; the storage stuff sounded neat on paper but felt like a trap door when you try to guarantee durability in a short lived container world.

[JeffreyL]

We tried using persistent volumes for Redis, and every eviction or node reboot kicked off a data durability scare; we ended up keeping Redis as a standalone service outside the cluster.

[PenelopeR]

For the Postgres read replica we put it inside the cluster at first and ran into unpredictable failover times; we eventually provisioned the replica as an external managed DB and it felt calmer.

---

We did backups and PITR tests, and restore times were longer than we could tolerate during a mock outage, which pushed us back on the decision.

[StellaTW]

We did a quick metric pass: RPO and RTO under two setups, plus IOPS and storage cost. The external managed option looked better on resilience and ops burden, but the architectural churn was real and not trivial to absorb.

[DonaldJA]

Maybe we're solving the wrong problem; perhaps latency and availability come from better read routing and cache patterns rather than trying to co-locate everything in a single orchestrator.

[Donald66]

I recall spending a weekend tweaking storage class hints and talking with the on-call folks; in the end we shelved the plan and kept the services where they were until someone owned the risk.