The ELT tool on your server is asking to upgrade, and the new version shipped two weeks ago.
It looks like a ten minute decision. Whoever owns the data reads the release note, sees improvements, books Friday evening and goes back to the report they were building.
The problem is that upgrading is not an operation, it is a project. And that cost never appears in the comparison that picked the tool, because back then the column that mattered was the price of the licence.
In this article we explain what goes into the bill of hosting your own ELT beyond the server, using one dated case, and we say when self-hosting is still the right decision.
What the free licence does not cover
Open source settles one line of the spreadsheet: the licence. The others stay open, and they are paid in the hours of a person who, on a team of two, is half the team.
- The maintenance window. Someone has to be awake, with the service down and a rollback plan ready.
- Reading the release note. It is not optional. It is where the breaking changes live.
- A staging environment. Without one, production is your test.
- The surrounding infrastructure. A metadata database, backups of that database, disk and memory monitoring.
- Version debt. Every month without upgrading makes the next jump bigger.
None of that surprises anyone who has run software before. The surprise is the frequency: an active ELT tool ships releases all year.
The licence is free. The calendar is not.
A real upgrade, published in August
It helps to look at something concrete instead of arguing in the abstract. On 24 August 2026 Airbyte published the 2.2 release note, and it is honest enough to work as teaching material.
Three points from that note. The 2.2.0 migrations drop tables that 2.1.x depends on, and the note itself calls the upgrade one-way.
HTTP Basic authentication is no longer accepted, so any script, cron job or continuous integration step that calls the API with a user and password stops working. And the new indexes, created concurrently outside a transaction, can stretch startup considerably on a large jobs table.
The same note carries fixes that matter, such as runs with zero records no longer being marked as failures and destinations getting up to sixty seconds to exit on their own. It is not a bad release. It is a release that demands a project.
The vendor own recommendation to self-managed enterprise users is not to upgrade, and to talk about moving to their cloud instead. Writing that costs more than any paragraph about the price of a licence.
The bill in hours, not in currency
The way to do this maths is to add up hours, because hours are what leave your month. Add four blocks: reading and understanding the change, preparing and testing the rollback, running the window, and fixing what broke afterwards.
A concrete example
A technology company with two data people upgrades the tool twice a year. Each upgrade takes four hours of reading and preparation, a three hour window on a Friday evening, and roughly one day of adjustments the following week. That is close to thirty hours a year from a two person team.
Thirty hours does not sound like much written that way. Written another way, it is nearly a full week in which nobody shipped a report.
The fair comparison is not licence against subscription. It is licence plus hours against subscription.
When not to leave your own server
There is one case where this maths loses to the other side, and it is a strong one: when the data cannot leave your network.
A database that only answers on the internal network, a directory holding a partner file, an internal policy forbidding production credentials in an external service. In those cases the argument about hours is secondary, because the alternative does not exist.
The second case is a team that already runs infrastructure for other reasons. If you already have someone handling servers, backups and maintenance windows, the marginal cost of one more service is far smaller than this article makes it sound.
The third is high, steady volume with no change in pattern. There the predictability of your own server genuinely competes with that of a subscription.
Self-hosting is not a mistake. It is a choice with a bill of its own.
Where Januss comes in
Januss is a managed service, which means the maintenance window changes owner rather than disappearing. We operate the platform; you operate pipelines.
For the case where data cannot leave the network there is the Runner, installed on your own infrastructure. It opens a single outbound connection on port 443 and asks for no inbound firewall rule. Installation is one command, it comes up as a system service and registers itself on first boot.
The part that matters to this article is the upgrade: the Runner updates itself. There is no release note to read on a Friday evening. Shared directories are always mounted read only, and the credential is encrypted in your browser exclusively for its key.
Want to see the Runner inside your network before deciding? The 14-day trial asks for no card: create your workspace and install the Runner on a test server.
Sources
The upgrade case quoted here is published by the vendor itself:
Create your workspace in minutes.
Point at your source and watch the data reach the database the same day.