Why three nodes and not two
## The problem with two
Two machines, one fails. The remaining one does not know whether the other is dead or whether only the connection has gone. If it keeps writing, and the other does the same, two diverging data sets arise. That is worse than an outage: an outage is visible, a split is not.
That is why it is not the number of running machines that decides, but the **majority**. A group that has the majority may write. One without it stops.
## Why odd numbers
With four machines, a partition can create two halves of two each. Neither side has the majority, both stop — and a deadlock is exactly the same as an outage, only more expensive.
| Nodes | Majority at | Failures tolerated |
|---|---|---|
| 3 | 2 | 1 |
| 4 | 3 | 1 |
| 5 | 3 | 2 |
| 7 | 4 | 3 |
Four nodes tolerate exactly as much as three and cost one more machine.
That is why the configurator only offers odd numbers.
## What this means for your configuration
The step from three to five is not a small one: it costs **two additional machines per critical role**, not two in total. With three critical roles, that is six machines.
In return, you can then absorb two simultaneous failures instead of one — which in practice means: one machine may be under maintenance while another fails. That is precisely why IT-Grundschutz requires five.
## What triple capacity is not
A cluster with three database nodes does not have three times the performance. The three hold **the same** data — that is the point. Anyone who calculates a cluster's headroom against the total capacity arrives at figures such as “2,554 % headroom” and has measured nothing.
The configurator therefore only calculates with the load-bearing capacity.