Backup modes: the interruption cost is written down, the consistency cost is not

When taking a backup you choose between three modes, and interfaces usually explain those three in terms of downtime: one with none, one with a few seconds, one that shuts the machine down. That is true but incomplete. The part not written down is consistency, and that is where the real difference sits.

What the no downtime mode actually does

The disk is captured as it stands at that instant while the machine keeps running. The phrase "as it stands at that instant" is the critical one: whatever was half written is inside, half written. This resembles pulling the power on the machine. A journalled filesystem sorts itself out at boot and nothing goes wrong. But a database that was in the middle of a write at that exact moment may have to run its own recovery when restored.

This is called crash consistency, and for most machines it is enough. Where it is not enough is a machine holding something that writes constantly and cares about what it writes.

Closing the gap without stopping the machine

When the guest agent is installed, the filesystem can be told at the moment of the backup to "wait, flush what you are holding to disk first". A brief freeze happens, the image is taken, it thaws. That turns the "power pulled" state into a "cleanly paused" state, and there is still no downtime.

This is the most practical argument for installing the guest agent. The agent is not there only to report an address, it is there for exactly this moment.

The application's own copy is a separate thing

For applications that keep their own state, such as databases and mail stores, a machine backup and an application dump do different jobs. The machine backup brings the machine back; the application's own dump brings the data back in a form the application trusts. On machines holding data that matters you do both, and neither replaces the other.

Suspend and stop

Suspend pauses the machine briefly. The memory state is consistent, but it adds nothing to the application's own consistency beyond what a frozen filesystem gives, and in return there is downtime.

Stop is the most honest option: a backup taken while the machine is off leaves no questions. If a machine can afford a minute of downtime overnight, this option removes the whole question.

The practical rule

Default to the no downtime mode, install the guest agent, and on machines holding real data take the application's own dump as well. Putting everything on stop out of fear is wrong: an arrangement that shuts down every machine every night becomes an arrangement nobody wants to touch, and one day it gets turned off. Keep stop for the few machines where the data matters more than the uptime.

What Atlas does

When taking a backup Atlas shows the three modes side by side and states in plain words the interruption each one causes, so you do not have to guess which pauses what. The mode that causes no interruption is selected by default. The consistency side depends on what is running inside the machine, so it is your decision rather than the product's: this entry was written so you can make that decision.