Backend Patterns Skill

Degree of freedom: MIXED. Which pattern to apply [HIGH freedom]; existing-architecture probes and the validation list [LOW freedom — run exactly].

How to reason

1. Observe — existing server/api/actions, ORM, queues, cache
2. Interpret — missing pattern vs duplicate vs wrong layer
3. Classify — reuse / add-queue / add-cache / add-rate-limit / edge-fn
4. Severity — unauthenticated mutation outranks a missing cache

Worked example

Observe: createOrder sends email inline; no Inngest/queue; checkout p95 4s. Interpret: confirmation is post-response work, not request-path. Classify: after() or an order/created job; keep the write transactional. Verify: order row commits; email/inventory run after response; retries do not double-charge.

Self-critique before reporting

• Existing first — searched server/api/actions before adding a second pattern
• Auth + validate — every mutation checks session and Zod
• Idempotent — retries on the new path do not double-apply
• Right owner — which architecture to pick → audit-backend-architecture

Design scalable, maintainable backend architectures using modern patterns and best practices.

Code examples lean on Next.js App Router + Supabase/Prisma. The patterns are stack-agnostic — adapt ORMs, client libraries, and deploy targets to your detected ecosystem.

Check existing first [LOW freedom — run exactly]

Before implementing ANY backend pattern, verify:

1. Check existing architecture:

ls -la src/server/ src/api/ app/api/ supabase/functions/ 2>/dev/null
cat package.json | grep -i "prisma\|drizzle\|supabase\|trpc"

2. Check existing patterns:

rg "createTRPCRouter|publicProcedure" --type ts -l
rg "'use server'" --type ts -l
ls -la supabase/migrations/*.sql 2>/dev/null | tail -5

3. Check database setup:

cat prisma/schema.prisma 2>/dev/null | head -50
cat supabase/config.toml 2>/dev/null

Why: Backend changes have wide impact. Understand existing architecture first.

Server Actions (Next.js 16+) [HIGH freedom]

Next.js 16: Turbopack default; 'use cache' + cacheComponents; reactCompiler: true; middleware.ts → proxy.ts (grep both). Instant navigations → enhance-web-instant-nav.

• Order inside every action: auth check → Zod safeParse → execute → map known errors (Prisma P2002) → generic failure message.
• Return ActionResult<T>: { success: true, data } or { success: false, error, fieldErrors? }; never throw to the client.
• revalidatePath after a successful write.
• Post-response work (confirmation email, inventory, notifications) goes in after() from next/server, keeping the write transactional.

'use server'
import { after } from 'next/server'

export async function createOrder(formData: FormData) {
  const order = await db.order.create({ data: { /* ... */ } })
  after(async () => {                      // runs after the response is sent
    await sendOrderConfirmation(order.id)
    await updateInventory(order.items)
  })
  revalidatePath('/orders')
  return { success: true, data: order }
}

Full createUser action and the after() example: references/request-patterns.md §Server Actions.

tRPC Setup [HIGH freedom]

• publicProcedure for reads anyone may do; protectedProcedure for anything that uses ctx.session.
• Every procedure has a Zod .input(); mutations stamp createdById from the session, not the input.
• List endpoints paginate by cursor: fetch limit + 1, pop the extra row as nextCursor.

Full usersRouter (getById, create, cursor-paginated list): references/request-patterns.md §tRPC router definition.

Supabase Edge Functions [HIGH freedom]

• Deno.serve; answer OPTIONS with the CORS headers first.
• Verify the webhook signature against the raw body before parsing JSON; 401 on mismatch.
• Admin client from SUPABASE_SERVICE_ROLE_KEY bypasses RLS; use it only inside the function.
• Catch everything; log server-side, return a generic 500 JSON body with CORS headers.

Full process-webhook function: references/request-patterns.md §Supabase Edge Function.

Database Patterns [HIGH freedom]

• Optimistic locking — version INT column; UPDATE ... WHERE id = $1 AND version = $2 and treat 0 rows as a concurrent update.
• Soft deletes — nullable deletedAt with an index; every read filters deletedAt: null; delete sets the timestamp.
• Audit logging — audit_logs(table_name, record_id, action, old_data, new_data, user_id) filled by a SECURITY DEFINER trigger using TG_OP, row_to_json, auth.uid().

UPDATE orders SET status = 'shipped', version = version + 1
WHERE id = $1 AND version = $2;   -- 0 rows → concurrent update, retry or surface conflict

Soft-delete Prisma model and the full audit trigger: references/data-and-jobs.md §Soft deletes, §Audit logging.

Caching Patterns [HIGH freedom]

• Next.js: fetch(url, { next: { revalidate, tags } }); revalidateTag on demand; unstable_cache around DB queries with tags.
• Redis (Upstash): one getCachedData(key, fetcher, ttl) helper, cache-aside; key by entity and id (user:${id}).
• Pick TTL by staleness tolerance, and always tag so a write can bust the cache.

const getCachedUser = unstable_cache(
  async (id: string) => db.user.findUnique({ where: { id } }),
  ['user'],
  { revalidate: 3600, tags: ['users'] }
)

Next.js cache calls and the Redis helper: references/data-and-jobs.md §Next.js cache, §Redis caching.

Background Jobs [HIGH freedom]

• Default: Inngest. inngest.createFunction({ id }, { event }, async ({ event, step }) => ...); each side effect in its own step.run so retries resume, not restart.
• Trigger with inngest.send({ name: 'order/created', data }) from the Server Action after the write commits.
• Early exit (out of stock) is a step.run + return { status: 'cancelled' }, not a throw.

export const processOrder = inngest.createFunction(
  { id: 'process-order' }, { event: 'order/created' },
  async ({ event, step }) => {
    const payment = await step.run('charge-payment', () => chargeCustomer(event.data.paymentMethod))
    await step.run('send-confirmation', () => sendOrderConfirmation(event.data.orderId))
    return { status: 'completed', paymentId: payment.id }
  }
)

Full Inngest function and the Trigger.dev scheduled-task alternative: references/data-and-jobs.md §Background jobs, §Alternatives.

Rate Limiting [HIGH freedom]

• @upstash/ratelimit with Ratelimit.slidingWindow(10, '10 s') over Redis.fromEnv().
• Key by user id (or IP for anonymous); on !success return Too many requests with retryAfter seconds from reset.

Full rateLimitedAction: references/request-patterns.md §Rate limiting.

Architecture patterns (distributed systems) [HIGH freedom]

Gateway, BFF, bulkhead, circuit breaker, outbox+CDC, saga, hexagonal, ACL, and strangler-fig → references/architecture-patterns.md. Pick the pattern for the topology (no mesh on a monolith; no CQRS unless reads/writes diverge). Timeouts/retries/idempotency → audit-resilience; structural gap report → audit-backend-architecture.

Validation [LOW freedom — do not skip]

After implementing backend patterns:

1. Error handling → All errors caught, logged, safe response returned
2. Auth checks → Every mutation verifies authentication
3. Input validation → Zod schema on all inputs
4. Rate limiting → Sensitive endpoints protected
5. Idempotency → Critical operations handle retries
6. Logging → Structured logs without sensitive data
7. Testing → Unit tests for business logic, integration for APIs