CodeShare Private notes
Sign in Create account

Architecture

Status mix: CURRENT PHP/MySQL CodeShare. PLANNED experiment platform. FUTURE vision workers and router.

CURRENT runtime

LayerChoiceWhy it is what it is
AppPHP 8 in app/pastesv2Same host as Reversal apps; fastest integration
HTTPApache root .htaccessClean URLs on pastes.reversal.org
DBMySQL pastes_dbExisting CodeShare + legacy Paste table
AuthEE Holdings session + local users + GoogleShared identity
Keyspvsk_ hashed SHA-256Agent ingest without OAuth in Cursor
CommandEncrypted bind + HTTP forwardPer-project notes without exposing pvsk_ to agents

Request path today:

Browser or agent -> TLS -> Apache rewrite -> PHP index.php or api/v1/*/index.php -> PDO -> pastesv2_* tables.

There is no job queue, no object store, no Playwright, no separate FastAPI service.

Target platform (PLANNED / FUTURE)

The strategy paper recommends:

  • Frontend: Next.js / React / TypeScript when the experiment UI outgrows PHP pages (Phase 4+). Until then, CodeShare + Command PHP UIs.
  • Core API: keep PHP for ingest and Command integration; add Python workers for eval/vision (Phase 3). FastAPI is optional, not a rewrite mandate.
  • Database: PostgreSQL when version/experiment volume and JSON querying justify a migration. MySQL can carry Phase 1-2 if tables stay normalized.
  • Queue: Redis + workers (Celery or PHP workers). Required before Playwright and LLM-as-judge.
  • Object storage: S3 / R2 / MinIO for screenshots. Not LONGBLOB.
  • Auth: continue EE identity + scoped API keys.
  • Observability: OpenTelemetry + structured logs (PLANNED).
  • Browser: Playwright (Phase 3).
  • Vision: pluggable providers behind EE VIZN (FUTURE).
  • ML: Python, scikit-learn / XGBoost, PyTorch only when rankers exist (Phase 7).
  • Deploy: Docker when workers exist. Kubernetes only when queue depth and multi-region require it.

Scaling strategy

  1. Vertical MySQL + PHP for notes (CURRENT).
  2. Add paste_versions + ratings on MySQL (Phase 1-2).
  3. Offload assets to object storage; add Redis workers (Phase 3).
  4. Split read models for leaderboards (Phase 4).
  5. Migrate experiment OLTP to PostgreSQL if JSON and advisory locks become painful (Phase 4-5).
  6. Kubernetes last.

Do not start at step 6. A paste bin that cannot record a score does not need a cluster.

Diagram

See diagrams/architecture.mmd and the rendered copy in the main paper.