The Home Assistant that runs my house is on 2024.11.1, released in November 2024. The current version is 2026.8.2. That is roughly 21 monthly releases of distance, and it has been running the whole time without complaining.
Before deciding anything I checked what was actually in there, against the running instance's own REST API rather than from memory: 751 entities, 150 components, no themes at all.
I did not try to upgrade
This matters more than the rest of the post, so it goes first. There was no failed upgrade. Nothing broke, because nothing was attempted. The decision was made against upgrading, not after one went wrong.
The reasons were unglamorous and they accumulated. The old instance runs from a plain docker run with no compose file, so there is no single document describing how it starts. Its configuration has grown for two years and nobody has read all of it recently. And a rollback would have meant restoring a VM snapshot, which takes the whole machine back, including anything else on it. That is the failure I would fix first on any instance meant to last: splitting the automations into separate YAML files is what keeps a configuration readable long enough to upgrade it.
None of that is impossible. Together it was enough that a clean build on a new machine looked like less work than an upgrade whose failure mode I could not describe in advance.
The screen that shows this was a choice
When you start a fresh Home Assistant, the very first screen offers two things side by side: create a new home, or restore from a backup. The restore option is not hidden behind an expert menu, it sits right there with equal weight.
The first screen. Restoring from a backup sits next to creating a new home, not behind a menu.
I mention it because it is the honest counterweight to everything above. The way back was directly available, and I did not take it. This was a preference for a clean start, not a rescue operation.
What a fresh install starts with
Three things surprised me on the way through, and all three take one sentence each in a guide, which is presumably why nobody writes them down.
The pre-selected country is the United States. That screen has exactly one field and it starts wrong for most of the people reading this. Click through without looking and you get imperial units and dollars.
One field, pre-set to the United States.
The language selector is on the welcome screen, before you create the user. Switching it there saves you from having field names and area names created in English and then living with them.
Home Assistant creates three areas by itself — living room, kitchen, bedroom — and puts them on the first dashboard next to an empty summaries column. That is the actual zero point, and it is further along than an empty page.
The first dashboard: three invented areas and an empty summaries column.
What rebuilding costs
The part I underestimated is that a rebuild is not one decision, it is 751 of them. Every entity in the old instance exists because an integration was added and configured at some point, and none of that comes along. The new machine starts with the three areas Home Assistant invents and nothing else.
The history goes too. Two years of recorder data stays with the old instance, so every long-run statistic starts from zero on the new one — energy totals, temperature trends, anything with a graph behind it. A backup restore would have kept all of it, which is exactly the trade being made here.
And the tuning is the quiet one. Automations that were adjusted over two years because they fired at the wrong moment carry no record of why they look the way they do. Copying them across is easy; knowing which of the odd-looking conditions were deliberate is not.
So the honest summary is that I chose a clean machine over two years of accumulated state, on a house that is not urgently broken. If the old instance were failing, I would have restored the backup and dealt with the mess afterwards.
The old one is still running
The new build binds no device in its first phase. The old instance continues to run the house, untouched, which is the only reason this was a comfortable decision at all. There is no cutover date yet and I am not setting one until the new machine has done something useful for a few weeks.
If you are in the same position, the thing I would check first is whether your current instance can be described. If you can point at one file that says how it starts, upgrading is a normal afternoon. If you cannot, that is the actual problem, and rebuilding is one way of fixing it.
What made this safe was having a restore path I trusted. I wrote up what a backup has to cover separately, and it is worth reading before you touch a two-year-old instance.
