Managed vs self-hosted

For every component you face the same choice. You can use the provider's managed service, or run the open-source software yourself on VMs or Kubernetes. Judge it on three things: the operational burden, the total cost and the lock-in.

Operational burden

Running Postgres yourself means backups and restore tests, point-in-time recovery, failover, minor and major upgrades, OS patching, disk growth and replication lag. Someone is on call for all of it. A rough estimate for a production-grade self-run database cluster is a quarter to half of an engineer's time, ongoing.

Total cost, not hourly rate

Managed services often charge 1.5 to 2 times the equivalent VM. Compare the totals, including people.

Self-hosted Postgres, primary + standby + backups
  2 x r7g.xlarge VMs + EBS               ~ $500/month
  0.3 engineer x $200k/yr loaded cost    ~ $5,000/month
  Total                                  ~ $5,500/month

RDS Multi-AZ, same instance size         ~ $1,000-1,200/month
  0.05 engineer for upgrades and tuning  ~ $830/month
  Total                                  ~ $2,000/month

At small and medium scale, managed wins by a wide margin. The crossover comes at large scale, where a dedicated team spreads its cost across hundreds of nodes. It also comes when you need an extension the provider does not support, or kernel-level tuning.

Lock-in

Managed Postgres has low lock-in, because you can pg_dump your data out. DynamoDB, Spanner, BigQuery and SQS are higher. Their APIs and data models are proprietary, so leaving means rewriting code.

Lock-in is a cost to price in, not a reason to refuse. Most companies never switch clouds. The engineering time spent staying portable is often more than a migration that never happens.

Common mistakes

• Comparing hourly rates. The cheap VM bill leaves out the engineers who run it.
• Refusing managed services to avoid lock-in. Usually the cost of avoiding it is larger than the risk.
• Self-hosting with no owner. A database nobody is on call for will fail at night.

Check yourself

1. Why does the self-hosted total include engineer time? Answer: Because the people who run the database are often the largest cost, larger than the servers.
2. Is lock-in a reason to avoid a managed service? Answer: Not by itself. Price it in, and weigh it against the cost of staying portable.

Key takeaways

• Compare total cost, including the engineer time to run the service.
• Default to managed databases, caches, queues and Kubernetes or serverless compute.
• Self-hosting makes sense at large scale, or with a need the provider cannot meet, and only with a clear owner.
• Treat lock-in as a cost to price in.

Further learning

Next, see how the networks that these services sit in are built: cloud networking and VPCs.