Start implementing a backend with code-first schemas

[bbenoist]

Work performed

I've added a typescript/backend project which includes a backend based on NestJS and the following packages:

• type-graphql
• typeorm
• class-validator
• dataloader

Results

• Backend schema is code-first. You only declare once your entities.
• You can switch the app to the new backend by defining the BACKEND_URL environment variable.
• We can directly reuse the entities as types in our future Workflow system. Benefiting from all of what's been already declared in the backend schema.
• DataLoader is supported! Most queries with children are batched 🎉

Problems encountered

• I had to reproduce the NextJS JWT token decryption functions since the next-auth package depends on unwanted dependencies (i.e. React).
• They are some behaviors in the current backend which I had to reproduce before we can change them:
  • LabelClasses must be ordered by index when queried to the DB
  • Some entities use slugs instead of ID to reference other entities in DB.
  • Nearly all mutations returns entities which we have to fetch in DB.

Caveats

• It can be a long debate but code-first designs have their own advantages and drawbacks compared to schema-first
• The backend is not loaded by NextJS. It's possible but not done yet as I wanted the old code to be still working on this branch and because of Deploy backend on Kubernetes #837
• Most of the relation queries have been batched with DataLoader but:
  • The aggregates counts can't be batched in a single high-level repository call. We must use QueryBuilder.
  • Pagination is applied on the global batched query but it should be per-parent-id

Resolved issues

Closes #780

Newly raised issues

Here are the items I see so far:

• Deploy backend on Kubernetes #837
• Move the prototype to our repository
• From Nest, decrypt JWT tokens signed and encrypted by next-auth
• Forward /api/graphql queries to the backend when BACKEND_URL is set.
• Make all queries work
• Make all mutations work
• Make sure that the entities are up-to-date with the latest GraphQL and Prisma schema revisions
• Move the remaining bits (s3, exports, image uploads, emails?, etc) into the backend (to be refined)
• Don't output entities in every mutation. They most often involve new DB calls and we have refetchQueries for that.
• Add generic CRUD unit, acceptance and integration tests which runs on every entity types

Bonus:

• In DB, replace relations referencing slug with id (e.g. membership.workspaceSlug => membership.workspaceId)
• Rename postgres db tables to snake_case to fit the default naming convention (⚠️ metabase updates needed)
• Make the repositories don't output deletedAt entities or entities whose parent has been deleted (already problematic on main)
• Add generic where, limit, order and skip types (included in filter types). Example implementation:
  • https://loopback.io/doc/en/lb4/Querying-data.html
  • https://github.com/loopbackio/loopback-next/blob/master/packages/filter/src/query.ts#L110
• Use GraphQL CLI plugin if possible
• See if our current pagination matches GraphQL recommendations and GraphQL Cursor Connections Specification then make changes it trivial

[vercel]

This pull request is being automatically deployed with Vercel (learn more).
To see the status of your deployment, click below or on the icon next to each commit.

🔍 Inspect: https://vercel.com/labelflow/labelflow/2hY8ec9oSyDYe4zDS2NznsJWR8fR
✅ Preview: https://labelflow-git-nest-backend-main-labelflow.vercel.app