Using GenAI to Tackle Complex Prisma Model Migrations

TL;DR: Need to write complex Prisma migrations (renaming, splitting, merging models)? Instead of writing SQL by hand:

1. Save your old schema
2. Make changes in schema.prisma
3. Run prisma migrate dev --create-only
4. Let GenAI handle the SQL data transformations

Been using this approach while building Stems - saved hours on migrations while preserving data and relationships. Works great for anything from simple renames to complex model splits.

While building Stems, I needed to split our monolithic Track model into three separate models: OfficialTrack, Stem, and Remix. This better represented our domain - official tracks have stems, which users can download to create remixes. Here's how I handled this data model evolution using GenAI to save time on SQL migrations.

The change involved rougly the following:

// Before: Everything in one model
model Track {
  id          String   @id
  title       String
  artist      String
  previewPath String?
  stemPaths   String[] // All stems stored here
  isRemix     Boolean
  parentTrack Track?   @relation("RemixOf")
  remixes     Track[]  @relation("RemixOf")
}
// After: Split into domain-specific models
model OfficialTrack {
  id     String @id
  title  String
  artist String
  stems  Stem[]
  remixes Remix[]
}
model Stem {
  id        String       @id
  title     String      // e.g., "vocals", "drums"
  audioUrl  String
  track     OfficialTrack @relation(fields: [trackId], references: [id])
  trackId   String
}
model Remix {
  id        String        @id
  title     String
  artist    String
  audioUrl  String
  original  OfficialTrack @relation(fields: [trackId], references: [id])
  trackId   String
}

Instead of manually writing the data migration, here's what worked:

1. Save the current schema: cp schema.prisma schema-old.prisma
2. Update schema.prisma with the new models
3. Generate migration scaffold: npx prisma migrate dev --create-only
4. Prompt for AI assistance:

I'm splitting the Track model into OfficialTrack, Stem, and Remix models.
schema-old.prisma has the original Track model where:

• Non-remix tracks (isRemix=false) should become OfficialTrack records
• stemPaths array should be split into individual Stem records
• Remix tracks (isRemix=true) should become Remix records
• All relationships need to be preserved Please modify the SQL migration to handle this data transformation.

1. Review the generated SQL carefully - this kind of split is complex enough that you might need to tweak the migration or ask for adjustments

Key learnings:

• This approach works best when you can clearly describe the transformation rules
• For complex splits like this, I always test the migration on a copy of production data first
• Breaking down the migration into smaller steps (create new tables → migrate data → update relationships → drop old table) made it easier to verify correctness
• Having the old schema file is crucial for AI to understand the full context This method significantly reduced the time I would've spent writing and debugging SQL migrations, though I still needed to review and test thoroughly given the complexity of the model split.

Comments

[Werzy Dev]

This is quite interesting approach, thanks for sharing