svcforge/tests/e2e/test_provision_real.py

"""The real thing: a real helm, against a real cluster, installing a real chart.

Everything else in this suite runs against a FakeProvisioner, which is what makes the
worker tests take milliseconds. That trade has one cost, and this file is the payment: the
fake proves the *worker* is correct, and cannot prove the *adapter* is. Argv order, the
`--wait` flag, chart resolution, RBAC, whether `upgrade --install` is genuinely idempotent
against a live release — none of it is exercised by a fake that returns None.

So one test does it for real, exactly once. It is marked `e2e` and excluded from `make test`
and from every pre-commit run; CI runs `-m e2e` in a job that has a cluster.

    kind create cluster --name svcforge
    uv run pytest -m e2e -q

It skips (does not fail) with no cluster, because a laptop without kubectl is not a
regression.
"""

from __future__ import annotations

import asyncio
import shutil
import subprocess
import uuid

import pytest

from svcforge_core.adapters.helm import HelmProvisioner
from svcforge_core.domain.models import CatalogEntry, SizeSpec

pytestmark = [pytest.mark.e2e, pytest.mark.slow]

NAMESPACE = "svcforge-e2e"

# A chart with no dependencies, no PVCs and no image pulls worth waiting on. The point is
# to exercise the adapter, not to wait four minutes for Elasticsearch.
ENTRY = CatalogEntry(
    service_type="podinfo",
    chart="oci://ghcr.io/stefanprodan/charts/podinfo",
    chart_version="6.7.1",
    sizes={"small": SizeSpec(replicas=1, resources={})},
)


def _cluster_reachable() -> bool:
    if not shutil.which("helm") or not shutil.which("kubectl"):
        return False
    out = subprocess.run(
        ["kubectl", "cluster-info"],
        capture_output=True,
        timeout=15,
    )
    return out.returncode == 0


requires_cluster = pytest.mark.skipif(
    not _cluster_reachable(),
    reason="no reachable cluster: `kind create cluster --name svcforge`",
)


@pytest.fixture(scope="module")
def namespace() -> str:
    subprocess.run(
        ["kubectl", "create", "namespace", NAMESPACE],
        capture_output=True,
        check=False,  # already exists is fine — the whole system is idempotent or it is broken
    )
    return NAMESPACE


@requires_cluster
async def test_install_is_idempotent_against_a_real_cluster(namespace: str) -> None:
    """Install twice. Get one release. This is the claim the fake cannot make for us."""
    release = f"e2e-podinfo-{uuid.uuid4().hex[:8]}"
    prov = HelmProvisioner(timeout_s=300)

    try:
        await prov.install(release=release, ns=namespace, entry=ENTRY, values={"replicaCount": 1})

        releases = [r for r in await prov.list_releases() if r.name == release]
        assert len(releases) == 1, f"expected exactly one release, got {releases}"

        # The redelivery, for real: same task, same deterministic release name, run again.
        # `helm upgrade --install` must converge, not duplicate and not error.
        await prov.install(release=release, ns=namespace, entry=ENTRY, values={"replicaCount": 1})

        releases = [r for r in await prov.list_releases() if r.name == release]
        assert len(releases) == 1, "second install created a second release: not idempotent"

        # --wait means the DB writing `ready` is telling the truth.
        out = subprocess.run(
            [
                "kubectl",
                "get",
                "deploy",
                "-n",
                namespace,
                "-l",
                f"app.kubernetes.io/instance={release}",
                "-o",
                "jsonpath={.items[*].status.readyReplicas}",
            ],
            capture_output=True,
            text=True,
            timeout=30,
        )
        assert out.stdout.strip() == "1", f"--wait returned before the pod was ready: {out.stdout!r}"
    finally:
        await prov.uninstall(release=release, ns=namespace)


@requires_cluster
async def test_uninstall_of_a_missing_release_is_not_an_error(namespace: str) -> None:
    """The desired state — no release — is already true. That is success, not failure."""
    prov = HelmProvisioner(timeout_s=120)
    await prov.uninstall(release=f"never-existed-{uuid.uuid4().hex[:8]}", ns=namespace)


@requires_cluster
async def test_real_helm_timeout_leaves_no_helm_behind(namespace: str) -> None:
    """A deadline that does not kill helm is a deadline that lets helm keep mutating."""
    prov = HelmProvisioner(timeout_s=1)  # cannot possibly finish
    release = f"e2e-timeout-{uuid.uuid4().hex[:8]}"

    with pytest.raises((TimeoutError, Exception)):
        await prov.install(release=release, ns=namespace, entry=ENTRY, values={})

    await asyncio.sleep(0.5)
    out = subprocess.run(["pgrep", "-f", f"helm.*{release}"], capture_output=True, text=True)
    assert out.stdout.strip() == "", "helm survived its own timeout and is still touching the cluster"

    await prov.uninstall(release=release, ns=namespace)