When Walter, a developer at one of two companies vying for the odd market of celebratory poems and TV dramas inspired by them, had to migrate from Python 2 to Python 3, the mood in the office was anything but festive. The latest round of layoffs had spared him but crushed the confidence of those who remained. «No one was willing to pull the all-nighters needed to fix the 20 percent of machines that refused to cooperate,» Walter told The Register. So, on a morning fueled by terrible coffee and another delayed job interview, he had what he himself calls a terrible idea: point Python 3 straight at Python 2 with a simple symbolic link.

ln -s /usr/bin/python2 /usr/bin/python3

He expected fire, alarms, angry emails. Instead, it worked. Silently. So well that nobody noticed. And Walter, with a mix of relief and self-deprecation, left behind a comment that is a monument to impostor syndrome: «Yeah, yeah, yeah. I know, I know. At least I was mostly sober so it can’t be THAT dumb!»

The story would have ended there, except that eight years later, now working elsewhere, Walter received a LinkedIn message from a former colleague: the last Linux upgrade was finally complete. And with it, they had unearthed the symlink, which had survived the migration from a homemade script to an Ansible playbook. A relic buried in the infrastructure for almost a decade.

This episode, though humorous, holds an uncomfortable truth about any on-premise architecture, especially when managed by understaffed and stressed teams. Technical shortcuts — the «it works, don’t touch it» kind — accumulate like geological layers. As long as the system holds, they become invisible. But automation (Ansible, in this case) does not eliminate them: it inherits them, reproduces them, and sometimes makes them harder to find, because they are embedded in declarative configurations that nobody questions anymore.

Those orchestrating self-hosted LLM pipelines today know the temptation well. A library version mismatch, a CUDA driver that doesn’t align with the containerized runtime, a model quantization that works only if we keep a deprecated flag active. The quick fix is just a commit away: a symlink, an alias, a half-documented workaround. In environments where you have total control — and therefore total responsibility — technical debt never pays itself off. It survives until someone, perhaps during a major migration, digs it up.

Walter’s case is emblematic for another reason: the line between what is sustainable and what is precarious is not always clear. His trick held up on a heterogeneous fleet of machines for eight years, without incident. That does not make it a good pattern, but it signals something deeper about the resilience of on-premise systems: when control is complete, even abuses can survive for a long time, precisely because the external orchestration typical of the cloud — which tends to invalidate non-standard configurations through forced updates — is absent. In the cloud, the symlink would have likely broken at the first managed environment update. In a private datacenter, your anomaly can become an institution.

Thus the lesson widens: automation for its own sake — «we put everything on Ansible» — is not enough to guarantee architectural hygiene. You need continuous review practices, integration tests that simulate changed conditions, and an almost obsessive attention to removing exceptions. For those evaluating on-premise deployment of LLMs, this is a critical reflection: the price of sovereignty and control is paid, in part, through an internal discipline that no vendor enforces. Technical debt, in these contexts, is not just a delay: it is a systemic risk that can lie dormant until someone, via a LinkedIn message, reminds you it exists.