AI summary by TwinMind
Flaws of the T3 Stack
Theo - t3․gg
37m 39s
Introduction and Stack Overview
- Creator of T3 stack discussing worst parts/limitations (not good parts today)
- Stack includes: Prisma, Next.js, React, Vercel, Tailwind, NextAuth, tRPC
- Acknowledges good parts: full stack type safety, great scalability, horizontal scaling, good standards, hot swap ability, bleeding edge responsibly
- Warning: don't use for blogs
Prisma Limitations
- Edge Runtime Issues: slow on edge, gigantic bundle size, includes rust binary that won't run on edge
- Performance: rough cold starts in serverless environments, not fastest for heavy/nested queries
- Node Modules: very opinionated, overrides node modules as npm install post install step to create types
- Code Generation: barrier compared to tRPC (no type gen needed), requires generating types to use properly
- General: not the fastest query resolution, especially with big heavy queries
tRPC Limitations
- Scope Limitations: sucks outside TS mono repos, don't use for third-party APIs
- Coupling: tightly couples backend to frontend (creator considers this compliment, not insult)
- Performance: TS dev performance limiting with lots of endpoints (couple hundred queries/mutations slow down TS server, V10 improved to few thousand)
- Client Support: support for non-JS clients inconsistent, React bindings most mature (wrapping React Query)
- Language Support: limited to TS environments
Tailwind Limitations
- Build Process: needs to parse all files that might use Tailwind classes, can get rough for big codebases
- CSS Knowledge: requires learning CSS (creator considers positive, others don't)
- Code Splitting: no real code splitting - all CSS goes to one file for whole app, not per-page optimization
- Bundle Size: individual pages won't have minimal CSS, but total app CSS will be small
- Styling Approach: some consider inlining messy (creator disagrees)
NextAuth Limitations
- Philosophy: too generic but also too specific, tries to solve all use cases rather than prescribing best practices
- User Management: no user ID by default, getting user ID is annoying
- JWT Support: half-baked JWT implementation
- TypeScript: requires many type overrides for proper usage
- Edge Runtime: buggy at best, creator hasn't gotten it working
- Session Management: useSession doesn't have meaningful states, rough data flow control
- Server Integration: requires getServerSideProps with session fetching, complex page props pass-through
- Token Management: no easy way to refresh tokens, not really implemented
- Platform Support: React Native support not there yet, non-React support even further behind
- Learning Curve: need to know a lot about NextAuth to use effectively
Next.js Limitations
- Bundle Size: ships a lot of JS, can't ship very small bundles
- Routing: mostly locked into file-based routing, bringing own router difficult with SSR
- Build System: locked into their builder (webpack/SWC), can't use Vite or Bun realistically
- React Dependency: React pretty mandatory, difficult to use without React
- Data Loading: patterns are "trash" and dated, getServerSideProps very dated
- Edge Support: really early, breaks in unexpected ways, not recommended yet
- Data Flow: unintuitive and limited, page props particularly problematic
- Imports: most Next imports "cursed", even head component problematic
- Router Issues: useRouter sucks at query params (undefined on first render)
- Use Case: not best for static sites, better for dynamic applications
React Limitations
- Static Sites: not good for static content, loads entire DOM into virtual DOM unnecessarily
- Prescriptive Nature: doesn't prescribe data, styles, routing - can be problem for some teams
- Solution Space: absurdly large variety in problem-solving approaches, inconsistent for teams wanting uniformity
- Documentation: new docs incredible but not localized yet, non-English support lagging behind
- Agency vs Consistency: gives developers lots of agency, but some teams prefer more consistency (like Angular)
Vercel Limitations
- Stateless: entirely stateless, need external database/caching solutions
- Serverless: cold starts if not using edge, Lambda vs Edge almost different products
- Multi-region: rough for pro users, only available in Enterprise plan for non-edge
- Edge Support: still early stage
- Framework Support: Next.js happiest path, other frameworks less supported
- Pricing: can scale weird for some features (image optimization, invocations)
- WebSockets: can't do websockets (implicit in serverless)
T3 Stack Opinions with Rough Edges
- Modularity: lot of flexibility between parts, developers can make poor decisions
- Learning Curve: lot of pieces to learn (tRPC implies React Query, Next.js implies server rendering, etc.)
- Developer Agency: assumes developers good enough to work around assumptions and make good decisions
- Testing: no prescribed testing solution, developers might pick poorly
- Not Rails Killer: not trying to build Rails replacement, don't care enough about testing to build it in
SSR and Data Loading Philosophy
- No First Paint Data: don't fetch data on first paint by default
- Loading States: show loading states rather than blocking server response
- Type Safety: using getServerSideProps loses type safety with tRPC
- SPA Plus Plus: single page app with good loading paths, data fetching, behaviors
- Cold Start Performance: blocking data from server to client on cold start is rough experience
- SEO Considerations: HTML coming down wire not fully correct with all data, potential SEO issues
TypeScript and Type Safety Issues
- Database Type Safety: databases might not be type safe, need to write "lies" somewhere
- Performance: TS dev performance issues with large codebases (tens of thousands of files)
- Runtime Performance: JavaScript runtime performance not optimal (though horizontally scalable)
- Editor Performance: might slow down editor due to type inference work
Documentation and Learning
- Minimal Docs: intentionally minimal documentation, defer to individual component docs
- Learning Curve: depends heavily on existing knowledge of stack components
- Stack vs Framework: not their responsibility to over-document since it's a stack, not framework
- Version Changes: heavy documentation would become outdated with component updates (e.g., tRPC V10)
Use Case Recommendations
- Not for Every Project: creator using it less recently due to more static work
- Best for Apps: ideal for data-changing applications where users keep pages open for minutes/hours
- Not for Blogs: repeatedly emphasizes not for blogs or mostly static content
- Fastest Full Stack: believes it's fastest way to build and scale full stack solutions
- Team Considerations: teams should evaluate limitations against their specific needs
