Databases
Postgres on your own box or as a managed service
The work a managed database actually removes, the control it takes away, the failure modes that surprise people in both directions, and a checklist for deciding per project rather than per company.
The argument about managed databases usually collapses into "managed is easier" versus "self-hosted is cheaper and more flexible". Both are true and neither decides anything. The useful version of the question is more specific: which operational tasks are you actually handing over, and which capabilities are you actually giving up?
What a managed service really removes
Be concrete about this, because "no ops" is marketing.
Provisioning and version upgrades. A managed engine gives you a running instance in a minute and handles minor version bumps. Major version upgrades are usually a guided process rather than an apt adventure, though on many platforms the major version is still your call and your migration.
Backups and point-in-time recovery. This is the single most valuable thing a managed service provides, and the thing most self-hosted setups get wrong. A cron with pg_dump is not a backup strategy. It has no point-in-time recovery, it misses the WAL, and nobody tests restoring it. A managed service that offers PITR is genuinely replacing a discipline most small teams do not maintain.
Failover, if it exists. Careful here. Some managed offerings are single-instance with a managed backup, not a cluster with automatic failover. That is still useful but it is not high availability, and vendor pages blur the difference. Ask whether failover is automatic, how long it takes, and whether the connection endpoint stays constant.
Monitoring and alerting on the boring things. Disk filling up, connections exhausting, long-running queries. A managed platform that surfaces these is doing work you probably were not doing.
Patching. Security patches for the engine arrive without you scheduling a maintenance window. This is real value if you are the only person who would otherwise be doing it.
What you give up
Superuser access. Managed engines frequently restrict SUPERUSER. That means no arbitrary extensions, no pg_ctl flags, no tuning parameters outside an allowlist, no filesystem-level tricks. If your application depends on an extension the provider does not offer, this is a hard blocker.
Version agility. You get the versions the platform offers. If you need a specific patch level for a compatibility reason, or you want the newest major version on release day, a managed service is the wrong shape.
Access to the machine. No perf, no strace, no reading the OS-level metrics yourself when a query is inexplicably slow. This matters less than people expect and more than vendors admit.
Cost predictability at scale. Managed pricing at small sizes is attractive and at large sizes often becomes poor value compared to a dedicated machine you tune yourself. The crossover point differs per provider and is worth checking rather than assuming.
Some failure-mode control. When the platform decides to fail over or restart, you do not get a vote. Usually that is fine. Occasionally it is not.
The failure modes that surprise people
Connections. Managed services cap connections and connection storms are the most common cause of an otherwise inexplicable outage. Application servers that open a pool per process and multiply can exhaust the limit without a single heavy query. Use a pooler.
Storage growth you cannot see. Some platforms grow storage automatically and bill for it, others stop accepting writes at a limit. Know which one you are on before a busy week fills the disk.
Backups that exist but are unproven. A backup you have never restored is a hypothesis. Managed or not, restore one into a scratch instance and look at the data. This should be a scheduled exercise, not a thought you have during an incident.
Memory sizing against cache. Postgres performance depends heavily on shared_buffers and the OS page cache. A managed instance with 1 GB of RAM and a 20 GB dataset will do a lot of disk reads, and no amount of indexing fixes a working set that does not fit. Size the instance against the dataset and the working set, not against the CPU count.
Locks and migrations. A long ALTER TABLE on a large table holds locks and can stall an application. Managed or self-hosted, this is your migration's fault, and it is more dangerous on a shared instance because you cannot observe the OS.
A decision checklist
Answer these per project. The answer is often different for the production database and the staging one, and that is fine.
- Does anything need a
SUPERUSER-only extension or setting? If yes, managed is out unless the provider offers it. - Is there someone who will genuinely maintain backups and test restores? If not, managed wins on evidence rather than convenience.
- How big will the dataset be in a year? Managed is comfortable at small sizes and gets expensive as it grows.
- Do you need automatic failover and a stable endpoint? If yes, check the offering actually provides it rather than assuming.
- How many connections will the app open at peak? Multiply the pool size by the worker count before you size anything.
- What is the cost of a couple of hours of downtime? If it is high, a service with automatic failover is not a luxury.
- Will you need point-in-time recovery? If you have ever wanted "restore to ten minutes ago",
pg_dumpcannot do it.
If the answers are "no exotic extensions, no DBA, modest size, wants PITR", a managed service is the right call and the flexibility you gave up was theoretical. If the answers are "custom extension, tight tuning, large dataset, we have a DBA", self-hosted is right and the convenience you gave up was not worth the price.
The cost shape nobody models
Managed pricing is usually per instance with fixed CPU and RAM, while self-hosted is per machine that you can also use for other things. That means a managed database looks expensive per unit compared to a box you already own, unless you honestly account for the time spent keeping the self-hosted one correct.
Those hours are real. Backups, restores, upgrades, monitoring, incident response, and the year where nobody noticed the backups had been failing since March. If you would not bill those hours to this project, you are not comparing like for like.
A note on slot-based pricing
One pattern worth understanding is buying database slots rather than individual instances: one price for N databases, each with a stated CPU and RAM cap on a shared cluster. It is cheap per database and it is not the same as dedicated capacity. It suits many small databases and applications where traffic is modest. It is a poor fit for one large database that wants the whole machine, and a provider that does not explain that distinction is overselling.
Ask which one you are buying before you assume the price reflects the same thing.
- postgresql
- managed databases
- operations

