Skip to content
MarginaliaDocsGitHub
Contents
handovers/2026-08-17-session-b1-spike-streaming.md

Rendered from the repository — the file stays the source of truth.

Session B1 — Spike S-1: SSE streaming on Scaleway Serverless Containers (2026-08-17)

Goal

Feasibility spike S-1 (feat/spike-streaming): deploy a minimal streaming skeleton to a real Scaleway Serverless Container and decide D-7 (serverless vs VM fallback) with actual measurements.

Verdict

Serverless confirmed. The VM fallback (D-7) is not needed. product/feasibility.md updated: F-2 ✅, D-7 decided, SSE + body-limit risk rows resolved.

Measurements (container marginalia-webapp, fr-par, min-scale 0, 1 vCPU/2 GB)

Test Result
SSE ticker (server-paced 500 ms, 10 events) Client arrival deltas 499–508 ms — no gateway buffering, event-by-event delivery
LLM stream (AI SDK toUIMessageStreamResponse, mistral-small-3.2) 87 SSE events arrive incrementally over ~0.77 s; clean finish + [DONE] — no truncation
Warm TTFB (GET info route, ×3) 87–96 ms
Warm LLM stream TTFB 100 ms (headers immediate), 2.22 s total for a ~100-line generation (17 KB SSE)
Cold start (min-scale 0, 25 min idle, 407 MB image) TTFB 3.89 s, HTTP 200 (vs ~0.09 s warm)
Body limit probes (POST, echo size) 512 KB / 1 MB / 5 MB / 10 MB / 20 MB — all HTTP 200, full body received. The user-reported ~1 MB limit did not materialize

What was done

  • Throwaway route apps/webapp/src/app/api/spike-stream/route.ts (delete after A4 exists): SSE ticker mode, LLM stream via @ai-sdk/openai-compatible against Scaleway Generative APIs, body-probe mode. AI SDK 7 (ai@7.0.66).
  • Keeper Dockerfile apps/webapp/Dockerfile (B2 reuses): Bun installs deps, next build + runtime on node:24-slim, standalone output, non-root. Verified locally before deploy. Companion Dockerfile.dockerignore (BuildKit per-Dockerfile ignore); build context is the repo root.
  • next.config.ts: output: "standalone" + outputFileTracingRoot (repo root — required in a monorepo).
  • Terraform: scaleway_container completed (deprecated args replaced: image, memory_limit_bytes, https_connections_only, public_endpoint); resource prefix renamed to marginalia (product name, owner-confirmed) before first apply; manual push script infrastructure/scripts/push-webapp.sh. First apply performed — state is local in the worktree (.worktrees/feat-spike-streaming/infrastructure/); migrate to the S3 backend in B2 before the worktree is deleted, or move terraform.tfstate to the main checkout when this branch merges.
  • Secrets: resolved from Proton Pass (vault marginalia, item scaleway-api-key) via pass-cli inject into untracked .env.local, symlinked into the worktree. New schema declarations: SCW_GENERATIVE_APIS_KEY, SCW_GENERATIVE_APIS_MODEL, SCW_GENERATIVE_APIS_BASE_URL, TF_VAR_generative_apis_key, TF_VAR_scw_project_id.

Gotchas found (worth remembering)

  1. Bun 1.3 isolated installs break Next standalone tracing — the node_modules/.bun symlink store loses @swc/helpers in the traced output. Dockerfile uses bun install --linker=hoisted.
  2. Generative APIs needs the project-scoped URL (https://api.scaleway.ai/<project-id>/v1) when the IAM key’s default project isn’t the target project — bare /v1 returns 403. The container gets SCW_GENERATIVE_APIS_BASE_URL from Terraform.
  3. bunx varlock run currently refuses to run anything because the two required Supabase values are empty in .env.local. Deploy commands were run with .env.local sourced directly. Once A-lane fills the Supabase values, varlock run works as documented.
  4. IAM key scoped to Object Storage/Registry/Containers/GenerativeApis cannot read the account API — Terraform data sources like scaleway_account_project would 403; the project ID is passed as TF_VAR_scw_project_id instead.

Cost left running

  • Container marginalia-webapp at min-scale 0: €0 idle, per-request within free tier if hit. max_scale = 2 caps worst case.
  • Registry image (~400 MB): ~€0.01/mo. Buckets: empty, €0.
  • The spike route is unauthenticated on a public URL; token spend is possible if the URL leaks. Fine for days, not weeks — B2 should gate or delete it.

Next (B2)

CI (lint/test/build), deploy pipeline reusing this Dockerfile + push script, Terraform state → S3 backend, bucket deploys for docs/marketing, delete or protect the spike route.