Large-video Rendering Pattern

Workers have strict memory/runtime limits, so run heavy render/extract work in E2B and stream outputs to R2.

Render Flow

1. Fetch script + assets metadata from D1.
2. Send render input to E2B sandbox (Remotion or Python path).
3. Upload output to R2 using multipart/chunked transfer.
4. Persist rendered_video asset row.
5. Post-process highlight clips and persist shorts rows.

Memory-safe Rules

• Avoid whole-file base64 for large outputs.
• Prefer 15MB multipart chunks for R2 uploads.
• For split image files, read in binary mode (format: 'bytes') to avoid corruption.

Highlight Post-processing

• Build slide timeline from measured audio durations.
• Resolve highlight windows from slide indexes.
• Extract portrait clips in E2B.
• Seed youtube_shorts_info rows as pending or error.

Failure Isolation

Render can still complete even if highlight extraction fails. Record short-level failures instead of failing the entire render stage.