Amey-Thakur Managed vs self-hosted
The sticker price comparison is always wrong in both directions: self-hosting hides engineer-hours, managed hides egress, per-request fees, and the price of leaving. Compare total costs including the exit.
Method
- Default managed for anything stateful and boring. Databases, queues, object storage, Kubernetes control planes: the undifferentiated heavy lifting where your custom requirements are near zero and the failure cost is high. Reserve self-hosting appetite for components where you have real differentiating requirements or crushing scale economics.
- Compute TCO with engineer-hours priced in. Self-hosted = instances + storage + the fraction of engineers who patch, upgrade, back up, tune, and get paged for it (0.25-1.0 FTE per serious stateful system is realistic) + the incident cost of doing it worse than a provider's dedicated team. Managed = service fees + the traffic-shaped charges (egress, per-request, per-connection) modeled at 10x your current volume, because that is when renegotiating is hardest.
- Grade the lock-in by interface, not vendor. Open-protocol managed services (Postgres-compatible, S3-API, Kafka-compatible, OpenTelemetry) are rentals you can walk from; proprietary APIs with gravity (bespoke query languages, integrated event buses, ML platforms) are marriages. Prefer the open-interface managed option even at a modest premium; it converts exit from a rewrite to a data move.
- Write the exit path before signing. For each critical managed dependency: how does the data leave (export tooling, egress cost, downtime), what replaces it, and roughly how many engineer-weeks? If the answer is "we could not, realistically", that is a strategic dependency: fine, but priced and acknowledged in the decision record (see architecture-decision-records).
- Re-decide at the trigger points. Crossovers happen: the managed bill passing the loaded cost of the FTEs it saves, a compliance need the provider cannot meet, scale where per-request pricing dwarfs hardware. Put the decision on a yearly review with current numbers (see vendor-evaluation for the assessment frame); grandfathered defaults are how six-figure line items hide.
- If self-hosting, budget the whole job. HA topology, backups with tested restores (see backup-restore), upgrade cadence, security patching SLA, monitoring and runbooks, and a named owning team. Self-hosted without that budget is not cheaper; it is deferred-incident financing.
Boundaries
- Control-plane outages of a provider are correlated risk you cannot engineer away from inside; if that is unacceptable, the answer is multi-provider architecture, which costs more than either option alone.
- Compliance sometimes decides unilaterally (data residency, audit requirements, air-gapped environments); check before running the economics.
- This is a per-component decision, not a philosophy; healthy stacks mix both and re-evaluate component by component.