Managed service or self-hosted?

Managed costs more per month and less in total once you count patching, upgrades, backups and the night somebody has to be awake. Self-hosting pays off when you have the operational skill already and need control over placement or tuning.

What does serverless mean for a database?

Capacity scales with load and idle time costs little or nothing. Good for uneven traffic and development environments. Watch cold starts, connection limits and the bill when a background job runs constantly.

How do I avoid being locked in?

Prefer engines with an open implementation you could run elsewhere, keep schema changes in migration files, and test a restore into a plain instance once. Proprietary extensions are the part that quietly makes leaving expensive.

Where should the database sit relative to the application?

In the same region, ideally the same zone. A few milliseconds per query becomes seconds per page once an application makes fifty of them, and no amount of caching hides a badly placed primary.

How is pricing usually calculated?

Compute, storage, backups, network egress and sometimes read units, billed separately. Egress and cross-region replication are the two lines that appear larger than expected on the first full month.

Database services compared: scaling, backups, migration

Nothing matches that yet

Try a broader category, or add the product you were looking for.

Add a product

What sits on this shelf

Places to keep data and get it back quickly: relational engines, document stores, key-value caches, search indexes, time-series stores and the managed services that run any of them for you.

Most teams need fewer of these than vendors suggest. One relational database handles the majority of applications for a long time, and adding a second engine adds a second set of failures, backups and people who understand it.

Analysis and reporting over that data belongs in analytics. Where the application itself runs is hosting and cloud.

Choosing an engine

Start from the shape of your data and the questions you will ask of it.

Relational engines suit anything with relationships, transactions, constraints and reporting, which describes most business software.

The rest are specialists. Document stores suit records that vary in shape and are usually read whole. Key-value stores are for speed on simple lookups. Search indexes exist because relational text search stops being enough at a certain size, and time-series stores exist because metrics arrive faster than anything else.

The honest default is a relational engine until something proves it insufficient. Teams that begin with an exotic store to avoid a scaling problem they do not have usually meet a different, worse problem first.

Compatibility matters more than branding. An engine speaking a widely implemented protocol gives you tools, drivers, hiring and an exit. A proprietary interface gives you a migration project.

Managed, self-hosted or serverless

Managed services charge a premium and remove work you would otherwise do at inconvenient hours: patching, version upgrades, failover, backup verification.

Self-hosting is cheaper on paper and only cheaper in practice when the skill is already in the building. Count the hours honestly, including the ones spent on the evening a disk fills up.

Serverless offerings scale capacity with load and charge little when idle. That suits uneven traffic, and it suits development environments even better. Check cold start behaviour, connection limits, and how the meter treats a background job that never stops.

The questions that decide reliability

Backups: how often, how long they are kept, and whether you can restore to a chosen point in time.
Restore: how long a full one takes at your data size, tested rather than promised.
Failover: automatic or manual, how long it takes, and what the application sees while it happens.
Replication: read replicas, cross-region copies, and the lag your code has to tolerate.
Connections: the ceiling, and whether pooling is included or something you must run.

Run one restore during the trial. A backup nobody has restored is a hope, and this is the category where hope is most expensive.

Growth, migrations and the parts that hurt later

Schema changes on a large table are the operation that ruins weekends. Ask how the engine handles adding a column, changing a type or building an index while traffic continues.

Keep migrations in version control alongside the application from day one. The alternative, changes applied by hand in a console, produces environments that differ in ways nobody can reconstruct.

Watch data location as well. A database in another region than the application adds latency to every query, and an application making fifty queries a page will feel it clearly.

How the bill is built

Compute, storage, backup storage, network egress, and on some services a unit for reads and writes.

Egress is the line that surprises people, particularly when analytics or replicas pull large volumes across regions. Storage rarely shrinks by itself, so deleted rows may still be paid for until maintenance runs. Development and staging copies double everything unless they are sized down or stopped overnight.

Ask what the price becomes at three times today’s volume rather than at today’s, since that is the number you will actually pay. Related habits are collected in what pricing pages hide.

Practices that keep a database boring

The goal in this category is an absence of drama, and a small number of habits produce most of it.

Keep migrations in version control and apply them the same way in every environment. Manual changes made in a console produce differences nobody can reconstruct later.

Add indexes deliberately and review them occasionally, since unused ones cost write performance and storage while looking harmless.

Watch slow queries as a routine rather than during an incident. The query that takes four seconds today is the outage next quarter when the table has grown.

And restore a backup on a schedule. A copy that has never been restored is a file, not a safeguard, and this is the category where discovering that late is most expensive.

Questions people ask

Managed service or self-hosted?: Managed costs more per month and less in total once you count patching, upgrades, backups and the night somebody has to be awake. Self-hosting pays off when you have the operational skill already and need control over placement or tuning.
What does serverless mean for a database?: Capacity scales with load and idle time costs little or nothing. Good for uneven traffic and development environments. Watch cold starts, connection limits and the bill when a background job runs constantly.
How do I avoid being locked in?: Prefer engines with an open implementation you could run elsewhere, keep schema changes in migration files, and test a restore into a plain instance once. Proprietary extensions are the part that quietly makes leaving expensive.
Where should the database sit relative to the application?: In the same region, ideally the same zone. A few milliseconds per query becomes seconds per page once an application makes fifty of them, and no amount of caching hides a badly placed primary.
How is pricing usually calculated?: Compute, storage, backups, network egress and sometimes read units, billed separately. Egress and cross-region replication are the two lines that appear larger than expected on the first full month.

Databases