16
Models after migration
down from 18
platform · 2025 · client unnamed
Inherited a TypeScript and SQL codebase, migrated the stack, and came out with fewer models and more capability.
16
Models after migration
down from 18
4
Content languages
4
Asset pipelines
image, GLB, USDZ, video
48
Route declarations
The problem
An existing platform needed to keep working while changing underneath. The original was TypeScript over SQL with eighteen models, and the brief was to take it over, migrate the stack, and extend it with multi-language content and 3D or AR vehicle assets.
The approach
Took ownership of the inherited codebase, understood what the eighteen models were actually doing, and rewrote it against a new stack with sixteen. Fewer models with more capability is the direction that matters in a migration; more models usually means the original structure was copied rather than understood. Added four-language content on every text field and four separate upload paths for images, GLB, USDZ and video, since 3D and AR assets have different handling requirements from photos.
Architecture
Complete rewrite rather than incremental port, model count reduced from eighteen to sixteen. Four distinct upload middlewares, because a GLB, a USDZ, an image and a video are not interchangeable.
Each with what it cost
A decision without a stated tradeoff is a preference.
01
Why
An incremental port of a schema you did not design tends to preserve its mistakes. Rewriting forced a decision about every model, and two turned out to be unnecessary.
What it costs
Higher upfront effort and a real cutover, against inheriting structure nobody could justify.
02
Why
Content is read constantly and written rarely, and every read wants one language. Variant fields mean no join on the hot path.
What it costs
The primary entity carries around fifty fields, half of them language variants, verbose in exchange for fast reads.
Outcome
A migrated platform with a smaller schema, four-language content throughout, and 3D and AR asset delivery. Documented with full endpoint tables rather than left as tribal knowledge.