Refueling Mid-Route

Opening Scene

Some specialized long-haul flights use mid-air refueling — taking on fuel from a tanker while still in flight, without ever landing or interrupting the mission. It’s a genuinely different operational approach from the standard “land, refuel fully at the gate, take off again” cycle most flights use, built specifically for situations where staying continuously airborne actually matters more than the convenience of a full ground stop.

That distinction — continuous, in-flight updating versus a full stop-and-restart cycle — maps directly onto the difference between online (or incremental) retraining and traditional, full batch retraining for a deployed model.

In Plain English

Batch retraining periodically retrains a model completely from scratch (or from a substantial checkpoint) on an updated dataset, then redeploys it — the “land, refuel fully, take off” approach. Online or incremental retraining updates a model continuously or in small, frequent increments as new data arrives, without a full retraining cycle each time. The right choice depends on how quickly the underlying patterns genuinely change, how costly full retraining is, and how much latency a team can tolerate between new data arriving and the model actually reflecting it.

The Old Way

Before online retraining approaches matured, batch retraining was essentially the only practical option, the same way mid-air refueling wasn’t an option before the specific engineering that makes it possible existed:

A textbook fully rewritten and reprinted periodically, rather than continuously updated in small increments.
A store’s full seasonal inventory refresh, rather than continuous, item-by-item restocking.
A software application fully rebuilt and released periodically, rather than continuously deployed in small increments.

In each case, batch updating was the practical default, simply because continuous, incremental updating wasn’t yet a well-developed, practical alternative.

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

Techniques for updating a model incrementally, without the full cost and disruption of complete retraining, have matured considerably, making online retraining a genuinely practical option for more situations than it used to be.
As real-world conditions change faster in many domains — connecting directly to the distribution shift and feature drift concepts covered elsewhere in this content library — the gap between when data becomes available and when a model actually reflects it matters more than it used to.
Infrastructure for safely validating incremental updates before they fully take effect, echoing the canary deployment approach from Article 4, has developed alongside online retraining techniques, reducing the risk of a bad incremental update degrading a live model.

The Metaphor, Fully Extended

Airport Operations	Retraining Concept
The standard “land, refuel fully, take off” cycle	Batch retraining — a complete retraining cycle, then redeployment
Mid-air refueling without ever landing	Online or incremental retraining, updating continuously
Choosing refueling approach based on the mission’s real requirements	Choosing a retraining approach based on how fast conditions genuinely change
The specialized engineering that makes mid-air refueling possible	The specialized techniques that make safe incremental updating possible
Validating a mid-air refueling procedure carefully before relying on it	Validating incremental updates carefully before they fully take effect
A standard flight that doesn’t actually need mid-air refueling	A model whose underlying patterns don’t change fast enough to need online retraining

For Beginners: What to Actually Do

Understand batch and online retraining as two genuinely different approaches with different tradeoffs, not one being simply “more advanced” than the other — the right choice depends on the specific situation.
Ask, for any model you’re working with, how quickly its underlying patterns actually change, since that directly informs whether batch retraining is genuinely sufficient or online retraining is worth the added complexity.
Recognize online retraining as connecting directly to the drift detection concepts covered earlier in this series — faster detection often pairs naturally with faster updating.

For Practitioners and Leaders: The Deeper Layer

Choose between batch and online retraining deliberately, based on the genuine rate of change in your specific domain and the real cost of retraining delay, not by default habit or unnecessary technical ambition.
Online retraining introduces genuine additional complexity and risk — validate incremental updates carefully before assuming faster is automatically better for your specific situation.
Recognize that many domains genuinely don’t need online retraining’s added complexity — batch retraining remains the right, sufficient choice for a large share of real-world use cases.

Quick Recap

Batch retraining fully retrains a model periodically; online or incremental retraining updates it continuously or in small increments as new data arrives.
This mirrors the distinction between a standard “land and refuel” cycle and specialized mid-air refueling, each suited to different operational needs.
Techniques for safe incremental updating have matured, making online retraining a genuinely practical option for more situations than before.
The right choice depends on how quickly underlying patterns genuinely change and how costly retraining delay actually is, not a default assumption either way.

Where This Fits in the Series

Article 8 covered maintaining a traceable version history; this article covered keeping a model’s knowledge current through different retraining approaches. Article 10 looks at what happens when something goes seriously wrong — grounding a fleet after one bad flight.