svcforge — reference implementation

A complete, working, verified build of the system ../learn-python/ teaches you to build.

Read this part first

This folder can ruin the course, and it will if you let it.

learn-python/AGENTS.md opens with a rule: "Do not write his implementation code for him. domain/, repo/, the claim loop, the handlers — those are the course." This repo is exactly that code. You asked for it deliberately, and the rule allows you to overrule it — but the reason for the rule did not go away when you did.

Module 4 is five evenings of getting the claim query wrong, and those five evenings are the learning. The evening you spend watching two workers grab the same task is the evening FOR UPDATE SKIP LOCKED stops being a phrase and becomes something you understand. Reading repo/tasks.py here takes ninety seconds, teaches you close to nothing, and feels exactly like learning. That feeling is the trap.

So:

Use it like this	Not like this
Attempt the module. Get stuck. Stay stuck 30 minutes. Then diff your version against this one.	Open this first "just to see the shape".
Steal the scaffolding — pyproject.toml, Dockerfiles, ci.yaml, the chart. Nothing is learned by fighting hatchling.	Copy domain/, repo/, handlers.py, or the claim loop.
Read the comments. They explain why, which is the part that transfers.	Read the code. It is the part that doesn't.
Use it to check an answer you already produced.	Use it to produce an answer.

The scaffolding is where struggling teaches you nothing. The domain and the queue are the whole point. Know which one you're reading.

What is actually verified

Everything below was executed, not asserted:

Claim	How it was proven
The claim query never double-claims	50 concurrent workers, 50 tasks, real Postgres. Each claimed exactly once.
The transaction story is real	Instance + task roll back together on an abort.
SIGTERM drains in flight	Worker finishes a 2s provision after stop is set, exits 0.
Handlers are idempotent	Re-running handle_provision installs once, not twice.
helm timeouts kill the process group	A negative control with proc.kill() leaks two sleep 300s; the real _run leaves zero.
The state machine is enforced in SQL too	Guard test fails when the guard is removed (control-tested).
Redis degrades safely	Rate limit fails open, cache falls through, platform stays up with Redis dead.
The images ship a wheel, not an editable	Negative control proved uv sync alone ships /app/libs; --no-editable fixed it.
The chart is valid	helm lint, helm template, hadolint — all clean.

Numbers from a real drain (400 tasks, local Postgres, FakeProvisioner) are in RUNBOOK.md.

Running it

No Docker on this host, deliberately. This box's containerd belongs to a Kubernetes kubelet; installing docker.io would evict the runtime and take every pod with it. So the integration tests take a DSN from the environment and only fall back to testcontainers (the CI path) when it is absent:

sudo -u postgres psql -c "create role svcforge login password 'svcforge' superuser;"
sudo -u postgres psql -c "create database svcforge_test owner svcforge;"

export PATH="$HOME/.local/bin:$PATH"
export SVCFORGE_TEST_DSN="postgresql://svcforge:svcforge@127.0.0.1:5432/svcforge_test"

make dev                              # uv sync
make lint                             # ruff + ruff format + mypy --strict
uv run pytest -q -m "not e2e and not slow"

Each pytest process clones its own database, so concurrent runs don't truncate each other.

Redis tests marked slow hit real Upstash. Mind the budget: the free tier is 500K commands/month = 0.19/sec sustained. The whole suite spends about 35.

set -a; . ~/.config/svcforge/secrets.env; set +a   # never commit these
uv run pytest -q -m slow
uv run python -m scripts.redis_budget            # projects month-end burn, exits 1 if over

Where things live

Module	Teaches	Read here
0	toolchain	pyproject.toml, Makefile
1	domain, pydantic, Protocol	libs/svcforge_core/svcforge_core/domain/
2	psycopg3, repos, migrations	repo/{db,instances}.py, migrations/001_init.sql
3	FastAPI, JWT, one transaction	services/api/
4	the queue — the core	repo/tasks.py, services/worker/
5	subprocess, timeouts, Protocols	adapters/helm.py (read _run twice)
6	day-2 fleet upgrades	domain/windows.py, InstanceRepo.list_upgradable
7	reconciler, obs	services/reconciler/, obs.py
8	packaging, CI, ArgoCD	services/*/Dockerfile, .gitea/workflows/ci.yaml, deploy/
9	chaos, load, CLI, runbook	scripts/load.py, services/cli/, RUNBOOK.md
10	Redis as a shortcut	adapters/redis.py, scripts/redis_budget.py

Where this deviates from the spec, and why

Honest list. Each was a real conflict, not a shortcut.

1. TaskRepo.enqueue exists twice. Module 2 specs enqueue(conn, ...); Module 4 specs enqueue(instance_id, kind) -> int. Both callers are real, so both exist: enqueue(conn, ...) and enqueue_standalone(...). Making conn optional would have hidden the transaction question, which is the one thing that module is about.
2. The claim query has a CTE wrapper. The spec calls it byte-identical. The update ... for update skip locked shape is untouched; an outer select joins instances.team so the worker can bind team to its logs at claim time. The alternative was a second round trip per task for a column the DB already had.
3. The day-2 work-list query gained service_type and a halted check. As printed it compares every service type against one version, and never consults catalog_versions despite the same module requiring halted → 0 rows.
4. Role → ClusterRole. The specced rules grant create on namespaces, which is cluster-scoped. A namespaced Role cannot express it. Rules verbatim otherwise.
5. DELETE is not one transaction. InstanceRepo.update_state owns its connection, so the CAS and the enqueue can't share one without reaching around the repo. Ordered for the failure mode instead: CAS first, enqueue second — a crash between leaves deleting with no task, which the reconciler sweeps up. The reverse would tear down a live service.
6. Traceparent is -03, not the spec's -01. Current SDKs set the random-trace-id bit alongside sampled. The spec's acceptance line is stale.
7. make_redis returns Redis | None. "No Redis configured" has to be runnable.

What was deliberately NOT built

Because Module 6 says so, and the restraint is the lesson: no resize, no backup/restore, no helm rollback automation, no deprecation timers, no rollouts table, no pause/resume CLI. A halted rollout is one column, cleared by hand with SQL.