Retiring an Aircraft That Still Flies Fine

Opening Scene

An older aircraft can still technically fly perfectly well and still get retired from active service — not because it’s suddenly broken, but because newer aircraft are meaningfully more efficient, the maintenance cost of an aging fleet keeps climbing, and continuing to operate it indefinitely stops making genuine operational sense well before it stops technically functioning. Retirement is a deliberate decision made proactively, based on genuine total cost and opportunity considerations, not a reaction to sudden failure.

That same deliberate, proactive decision is exactly what model decommissioning and deprecation require for a deployed model.

In Plain English

Model deprecation is the deliberate, planned process of phasing out a model — announcing the timeline, migrating dependent systems, and eventually decommissioning it entirely — rather than letting it quietly keep running indefinitely simply because nobody has actively decided to stop it. A model that’s still technically making reasonable predictions can genuinely deserve retirement: newer approaches may perform better, maintenance cost may have grown, or the underlying business need it served may have shifted enough that continuing to run it no longer makes sense.

The Old Way

Before formal model decommissioning was standard machine learning practice, the same drift toward indefinite, unexamined continuation, absent a deliberate decision to stop, showed up everywhere:

Legacy software systems kept running for years past their useful life, simply because nobody made an active decision to formally retire them.
Organizational processes continuing long after their original justification disappeared, purely from inertia rather than genuine ongoing value.
Equipment kept in service well past the point where replacement would genuinely be more cost-effective, absent a deliberate review process.

In each case, the absence of a deliberate, periodic “should this still exist” decision led to real, accumulating waste and risk that a proactive retirement process would have caught.

What’s Changing (and Why AI Is the Reason)

As organizations accumulate more deployed models over time, the practical cost of “model sprawl” — many old, quietly-still-running models nobody’s actively reviewing — has become a genuine, recognized organizational problem, connecting directly to this content library’s dedicated series on data platform cost and FinOps.
Model versioning and governance infrastructure, covered in Articles 8 and 11, makes it far easier to actually track which models are running and assess whether they still genuinely justify their operating cost.
Deliberate model lifecycle management — including planned decommissioning as a standard, expected phase, not a rare exception — is increasingly recognized as a core MLOps discipline, rather than deployment being treated as a one-way, permanent commitment.

The Metaphor, Fully Extended

Airport Operations	Model Decommissioning Concept
An older aircraft that still technically flies fine	A model that’s still technically making reasonable predictions
Rising maintenance costs on an aging fleet	Growing maintenance burden on an aging, unrefreshed model
A deliberate, planned retirement decision	A deliberate, planned deprecation and decommissioning process
An announced retirement timeline and transition plan	A communicated deprecation timeline and migration plan for dependent systems
An aircraft kept flying indefinitely absent a deliberate decision to retire it	A model left running indefinitely absent a deliberate decision to decommission it
A fleet management review assessing which aircraft to retire	A model portfolio review assessing which models still justify their operating cost

For Beginners: What to Actually Do

Understand that a model still “working” isn’t automatically a reason to keep it running indefinitely — genuine decommissioning decisions are a normal, healthy part of a model’s lifecycle.
Recognize deprecation as a deliberate, planned process, not something that should happen abruptly without warning to dependent systems and teams.
Get comfortable with the idea that “should we retire this” is a legitimate, recurring question worth asking periodically, not just when something breaks.

For Practitioners and Leaders: The Deeper Layer

Build periodic model portfolio reviews into standard organizational practice, assessing which deployed models still genuinely justify their ongoing operating cost and maintenance burden.
Treat deprecation as a planned process with a clear communicated timeline and migration path, not an abrupt, disruptive shutdown.
Recognize model sprawl — many old, unreviewed models quietly still running — as a genuine, growing organizational cost worth actively managing, directly connecting to broader cost management priorities.

Quick Recap

Model deprecation and decommissioning are deliberate, planned processes for retiring a model, even one that’s still technically functioning reasonably well.
This mirrors aviation’s deliberate aircraft retirement decisions, made proactively based on genuine cost and opportunity considerations, not reactive failure.
Growing model sprawl has made deliberate lifecycle management, including planned decommissioning, a recognized, essential organizational discipline.
Periodic portfolio reviews and clear, communicated deprecation timelines are the practical tools for making this discipline real.

Where This Fits in the Series

Article 13 covered handling genuinely unexpected input; this article covered deliberately retiring a model that’s run its course. Article 15 looks at how AI itself is now helping watch over every deployed model at once.