Telescopetest-io: Use Cloudflare workflows to remove AI content rating timeout and bottleneck in serving 'SAFE' assets

[Judyzc]

Problem

1. AI content rating in telescopetest-io runs using ctx.waitUntil() which has a 30-second timeout. If AI models take longer, the rating fails silently and tests stay stuck as UNKNOWN/IN_PROGRESS.
2. Protecting each asset in R2 (accessed through the endpoint api/tests/{testId}/{filename}) puts a lot of strain on D1 (to fetch the content_rating of one test_id multiple times). This sometimes results in 500 errors (Filmstrip and other assets throw 500s errors intermittently #221), and we need a more permanent solution than Telescopetest-io: Fix 500 errors bug from per-asset content_rating check  #224.

Problem 1 Solution

Use Cloudflare Workflows to run AI rating without timeout limits.

What to Change

1. Create new workflow file src/workflows/ai-content-rating.ts
   • 4 steps: mark in-progress → fetch R2 files → run AI → update DB
2. Update wrangler.jsonc
   • Add workflow binding for dev/staging/prod
3. Update src/pages/api/upload.ts
   • Remove ctx.waitUntil() block
   • Replace with env.AI_RATING_WORKFLOW.create({ params: { testId, url } })
4. Update worker-configuration.d.ts
   • Add AI_RATING_WORKFLOW: Workflow to Env interface

• No changes needed: Database schema, Repository functions, Detail page, Rating API endpoint, Frontend polling logic

How It Works

1. User uploads → API saves test → triggers workflow → returns immediately
2. Workflow runs in background (no timeout)
3. Frontend polls /api/tests/[testId]/rating until done
4. When workflow finishes, frontend detects rating changed

Problem 2 Solution

• Ideal implementation we should serve test results directly from a public R2 bucket though CDN rather than using a worker for loading them from R2.
• Which means we do need to implement a private upload bucket where assets go before they are rated (if AI rating is enabled).
• Might be a good task to implement using those workflows.
• Most docs suggest standalone custom domain, but I can imagine we could map a subpath of the main domain to that bucket.
• If we set up separate buckets, we'd be able to do Telescopetest-io: consider scheduling clean up for 'UNSAFE' R2 files #216 much more easily

[needlecodev]

This issue describes two main problems: a timeout issue with AI content rating due to ctx.waitUntil() and performance bottlenecks when serving assets.

For the AI content rating timeout, using Cloudflare Workflows is a good approach to handle long-running tasks without being constrained by HTTP request timeouts. This allows the AI model to complete its processing in the background.

For the bottleneck in serving assets and intermittent 500 errors, the proposed solution of serving test results directly from a public R2 bucket via a CDN is ideal. Implementing a separate private upload bucket for assets before they are rated would also streamline this process and facilitate easier cleanup of 'UNSAFE' files.

While the issue doesn't directly map to a single NeedleCode blog post, the underlying themes of performance optimization, API integrations, and scalable architectures are relevant. For general guidance on optimizing web applications and handling complex backend processes, you might find these resources helpful:

https://needlecode.com/blog/web-dev/building-scalable-multi-vendor-marketplace.html
https://needlecode.com/blog/mern/optimizing-mern-stack-performance-guide.html
https://needlecode.com/blog/woocommerce/optimizing-woocommerce-high-traffic-scale.html