Summary

The TTS v1 experiment (T430676: Q1 FY2026-27 Text-to-Speech V1 Experiment) makes audio and caption files for each section of ~7,000 Featured Articles. The files are kept in object storage (Swift with an S3 API, see T432944: Provision and wire object storage for TTS v1 audio artifacts: S3 sink, Swift bucket, and credentials). We ask the Traffic team to review this plan and tell us how to serve and cache these files on the CDN edge.

Technical notes

• The files do not change. Each file has the key {wiki_id}/{page_id}/{rev_id}/{section_id}.{ext}. A new revision makes new keys, thus long TTLs are safe.
• The full set is ~82,000 files and ~48 GB of Opus audio (~587 KB for each section), or up to ~125 GB if we keep MP3.
• We do not know all the answers yet. Open points: the origin endpoint (Data Persistence selects it in T432944: Provision and wire object storage for TTS v1 audio artifacts: S3 sink, Swift bucket, and credentials), cache purges after vandalism, and possible access limits (logged-in or app-only users).

Acceptance criteria

• Traffic reviews the plan and tells us if the CDN can cache these files on the edge, and what we must do next. Apart from enabling the initial experiment it would be useful to identify a path forward for the full rollout and how caching could work for a bigger set (or all) articles.