Milestone · Boson AI · Modal overflow

Build & validation report

← Back to plan & tracker

Milestone 9 — Control-plane server (API / Redis / local MP4) ✅

Goal: the pure-CPU control plane — the Boson video API, job status, MP4 storage, render dispatch, and spillover. Volume-only by default: the MP4 is written to the control plane's local disk and TTL-cleaned; S3 is optional. Code: ~/modal_examples/milestones/m9_control_plane.py.

What was built

A server (stdlib-only, so it runs and self-tests with no external services) implementing:

Two-sided cleanup: the renderer caches the MP4 to its own disk, transmits to the control plane, then deletes its local copy (per-job, so warm containers don't accumulate); the control plane TTL-cleans served files.

Production mapping: StatusStore → Redis, LocalBlobStore → control-plane disk (or shared NFS/EFS), render_fn → Modal spawn(). S3 is optional (CDN/presigned downloads, or a multi-replica control plane without shared disk).

How it was validated (python3 m9_control_plane.py)

submit -> video_xxx queued
status -> completed, progress=100
content -> 54 bytes, video/mp4 (served from local disk)
cleanup -> stale MP4 removed, fresh kept
back-pressure -> ['queued','queued','failed']  (1 shed at cap)
failed render -> /content 409 (no infinite poll)
VALIDATION PASSED

Code review (separate subagent) — CHANGES NEEDED → fixed

Finding (severity) Fix applied
/content couldn't distinguish failed from in_progress → clients poll forever (HIGH) /content now returns 409 + error for failed jobs; added a failed-path test over HTTP
Shared SpilloverController unlocked → "thread-safe" was false (MED) Fixed in M8 (added a lock); verified under the multi-threaded control plane
In-memory MP4 hides a streaming/OOM problem (MED) Switched default to a LocalBlobStore (disk-backed) with TTL cleanup; /content streams from disk. No S3.
Queued/worker race Verified already fixed — POST returns the queued snapshot, not mutable state

The POST→queued→in_progress→completed→content path and the shed path are both exercised end-to-end over real HTTP.

Validation also covers local-disk MP4 cleanup (stale file removed by TTL, fresh kept) — verified in python3 m9_control_plane.py.

Status: ✅ validated (API + status + local-disk MP4 + TTL cleanup + spillover + failed-path; no S3).