Enable safe NPM replica convergence

This commit is contained in:
Lyncoln
2026-09-20 10:37:11 -04:00
parent 291f4f492e
commit e4ca80d4d5
7 changed files with 351 additions and 16 deletions
+17
View File
@@ -314,6 +314,17 @@ def _required_capability(
)
def _required_retirement_support(
adapter: ExecutionAdapter,
) -> None:
retire = getattr(adapter, "retire", None)
if not callable(retire):
raise RuntimeError(
f"{adapter.endpoint.identity} adapter does not support "
"deferred retirement"
)
def _preflight(
adapters: Iterable[ExecutionAdapter],
plans: Iterable[EndpointPlan],
@@ -345,6 +356,9 @@ def _preflight(
f"No adapter for endpoint plan {plan.endpoint_id}"
)
if plan.retirements:
_required_retirement_support(adapter)
for retirement in plan.retirements:
planned_retirements.append(
(
@@ -381,6 +395,9 @@ def _preflight(
"Executable plan action requires desired state"
)
if action.operation == PlanOperation.REPLACE:
_required_retirement_support(adapter)
object_type = action.desired.object_type
_required_capability(
adapter,
+12
View File
@@ -21,6 +21,7 @@ from .persistence import SQLiteStateStore
from .projections import project_proxy_hosts_to_dns
from .reconciliation import (
build_endpoint_plans,
merge_desired_with_previous,
reconcile_observations,
)
@@ -315,6 +316,17 @@ def discover_reconcile_plan(
else []
)
if certificate_authoritative:
previous_certificates = [
row.object
for row in store.load_desired_state(CERTIFICATES)
]
npm_desired = merge_desired_with_previous(
npm_desired,
previous_certificates,
)
proxy_observations = _authoritative_observations(
observations,
live_pairs,
+8
View File
@@ -43,6 +43,10 @@ class Plugin:
PROXY_HOSTS,
CapabilityOperation.DISCOVER,
),
Capability(
PROXY_HOSTS,
CapabilityOperation.APPLY,
),
Capability(
PROXY_HOSTS,
CapabilityOperation.VERIFY,
@@ -51,6 +55,10 @@ class Plugin:
CERTIFICATES,
CapabilityOperation.DISCOVER,
),
Capability(
CERTIFICATES,
CapabilityOperation.APPLY,
),
Capability(
CERTIFICATES,
CapabilityOperation.VERIFY,
+60
View File
@@ -54,6 +54,46 @@ class ReconciliationPolicy(Protocol):
) -> bool: ...
def merge_desired_with_previous(
fresh: Iterable[NormalizedObject],
previous: Iterable[NormalizedObject],
policies: Mapping[str, ReconciliationPolicy] | None = None,
) -> list[NormalizedObject]:
selected_policies = (
DEFAULT_POLICIES
if policies is None
else policies
)
previous_by_identity = {
(object.object_type, object.logical_key): object
for object in previous
}
merged: list[NormalizedObject] = []
for candidate in fresh:
identity = (
candidate.object_type,
candidate.logical_key,
)
prior = previous_by_identity.get(identity)
policy = selected_policies.get(candidate.object_type)
merger = (
getattr(policy, "merge_desired", None)
if policy is not None
else None
)
if prior is not None and callable(merger):
candidate = merger(prior, candidate)
merged.append(candidate)
return merged
def _observation_sort_key(
observation: Observation,
) -> tuple[str, str, str]:
@@ -343,6 +383,26 @@ class CertificateNewestPolicy:
== self._portable_state(desired)
)
def merge_desired(
self,
previous: NormalizedObject,
fresh: NormalizedObject,
) -> NormalizedObject:
if (
previous.object_type != CERTIFICATES
or fresh.object_type != CERTIFICATES
or previous.logical_key != fresh.logical_key
):
raise ReconciliationConflict(
"Certificate desired-state merge requires "
"matching certificate identities"
)
if self._newest_key(previous) > self._newest_key(fresh):
return previous
return fresh
@staticmethod
def _equivalent_selection_key(
object: NormalizedObject,
+142
View File
@@ -655,6 +655,148 @@ class ExecutionTests(unittest.TestCase):
)
self.assertEqual(store.load_applied_state(), [])
def test_preflight_requires_retire_for_planned_retirement_before_mutation(
self,
):
endpoint = Endpoint("npm", "192.168.1.4:81")
inner = FakeAdapter(
endpoint,
capabilities(CERTIFICATES),
)
class NoRetireAdapter:
def __init__(self, wrapped):
self.endpoint = wrapped.endpoint
self.capabilities = wrapped.capabilities
self.wrapped = wrapped
def apply(self, plan_action):
return self.wrapped.apply(plan_action)
def verify(self, object):
return self.wrapped.verify(object)
adapter = NoRetireAdapter(inner)
old_certificate = NormalizedObject(
object_type=CERTIFICATES,
logical_key="app.example.com",
local_key="67",
payload={
"domain_names": ["app.example.com"],
},
)
plan = EndpointPlan(
endpoint.identity,
(
action(
endpoint.identity,
PlanOperation.CREATE,
object_type=CERTIFICATES,
logical_key="other.example.com",
),
),
retirements=(
RetireRequest(
object=old_certificate,
),
),
)
with SQLiteStateStore(":memory:") as store:
with self.assertRaisesRegex(
RuntimeError,
"does not support deferred retirement",
):
execute_endpoint_plans(
[adapter],
[plan],
store,
executed_at=NOW,
)
self.assertEqual(
store.load_applied_state(),
[],
)
self.assertEqual(inner.events, [])
self.assertEqual(inner.applied_actions, [])
def test_preflight_requires_retire_for_replace_before_mutation(
self,
):
endpoint = Endpoint("npm", "192.168.1.4:81")
inner = FakeAdapter(
endpoint,
capabilities(CERTIFICATES),
)
class NoRetireAdapter:
def __init__(self, wrapped):
self.endpoint = wrapped.endpoint
self.capabilities = wrapped.capabilities
self.wrapped = wrapped
def apply(self, plan_action):
return self.wrapped.apply(plan_action)
def verify(self, object):
return self.wrapped.verify(object)
adapter = NoRetireAdapter(inner)
current = NormalizedObject(
object_type=CERTIFICATES,
logical_key="app.example.com",
local_key="67",
payload={
"domain_names": ["app.example.com"],
"expires_on": "2026-11-01T00:00:00+00:00",
},
)
desired = NormalizedObject(
object_type=CERTIFICATES,
logical_key="app.example.com",
payload={
"domain_names": ["app.example.com"],
"expires_on": "2026-12-01T00:00:00+00:00",
},
)
plan = EndpointPlan(
endpoint.identity,
(
PlanAction(
endpoint_id=endpoint.identity,
operation=PlanOperation.REPLACE,
desired=desired,
current=current,
),
),
)
with SQLiteStateStore(":memory:") as store:
with self.assertRaisesRegex(
RuntimeError,
"does not support deferred retirement",
):
execute_endpoint_plans(
[adapter],
[plan],
store,
executed_at=NOW,
)
self.assertEqual(
store.load_applied_state(),
[],
)
self.assertEqual(inner.events, [])
self.assertEqual(inner.applied_actions, [])
def test_preflight_requires_verify_before_any_mutation(self):
endpoint = Endpoint("verizon", "192.168.1.1:443")
adapter = FakeAdapter(
+9 -16
View File
@@ -94,7 +94,7 @@ class NpmAdapterTests(unittest.TestCase):
self.assertIs(adapters[0].client, first_client)
self.assertIs(adapters[1].client, second_client)
def test_capabilities_cover_discovery_and_verify_only(self):
def test_capabilities_cover_discovery_apply_and_verify(self):
adapter = self.adapter(Mock())
self.assertEqual(
@@ -105,6 +105,10 @@ class NpmAdapterTests(unittest.TestCase):
PROXY_HOSTS,
CapabilityOperation.DISCOVER,
),
Capability(
PROXY_HOSTS,
CapabilityOperation.APPLY,
),
Capability(
PROXY_HOSTS,
CapabilityOperation.VERIFY,
@@ -113,6 +117,10 @@ class NpmAdapterTests(unittest.TestCase):
CERTIFICATES,
CapabilityOperation.DISCOVER,
),
Capability(
CERTIFICATES,
CapabilityOperation.APPLY,
),
Capability(
CERTIFICATES,
CapabilityOperation.VERIFY,
@@ -121,21 +129,6 @@ class NpmAdapterTests(unittest.TestCase):
),
)
self.assertNotIn(
Capability(
PROXY_HOSTS,
CapabilityOperation.APPLY,
),
adapter.capabilities,
)
self.assertNotIn(
Capability(
CERTIFICATES,
CapabilityOperation.APPLY,
),
adapter.capabilities,
)
def test_discovers_proxy_hosts_with_endpoint_local_key(self):
client = Mock()
client.proxy_hosts.return_value = [
+103
View File
@@ -240,6 +240,109 @@ class ObservationAuthorityTests(unittest.TestCase):
)
)
def test_persisted_newer_certificate_cannot_regress_to_older_live_certificate(
self,
):
endpoint = Endpoint(
"npm",
"192.168.1.3:81",
)
live = FakeAdapter(
endpoint=endpoint,
capabilities=capability_set(
CERTIFICATES,
apply=True,
),
discovered={
CERTIFICATES: [
discovered_certificate(
"20",
valid_from=(
"2026-09-15T00:00:00+00:00"
),
expires_on=(
"2026-12-15T00:00:00+00:00"
),
)
]
},
)
persisted = normalized_certificate(
"99",
valid_from=(
"2026-09-19T00:00:00+00:00"
),
expires_on=(
"2026-12-19T00:00:00+00:00"
),
)
persisted = NormalizedObject(
object_type=persisted.object_type,
logical_key=persisted.logical_key,
payload=dict(persisted.payload),
)
with SQLiteStateStore(":memory:") as store:
store.replace_desired_state(
CERTIFICATES,
[persisted],
T1,
)
result = discover_reconcile_plan(
[live],
store,
target_ip="192.168.1.101",
run_at=T2,
)
stored = store.load_desired_state(
CERTIFICATES
)
certificates = [
object
for object in result.desired
if object.object_type == CERTIFICATES
]
self.assertEqual(len(certificates), 1)
self.assertEqual(
certificates[0].payload["valid_from"],
"2026-09-19T00:00:00+00:00",
)
self.assertEqual(
certificates[0].payload["expires_on"],
"2026-12-19T00:00:00+00:00",
)
self.assertEqual(len(stored), 1)
self.assertEqual(
stored[0].object,
persisted,
)
self.assertEqual(len(result.plans), 1)
self.assertEqual(
result.plans[0].endpoint_id,
endpoint.identity,
)
self.assertEqual(
len(result.plans[0].actions),
1,
)
self.assertEqual(
result.plans[0].actions[0].operation.value,
"replace",
)
self.assertEqual(
result.plans[0].actions[0].desired,
persisted,
)
def test_all_failed_sources_preserve_persisted_desired_without_planning(self):
npm_endpoint = Endpoint(
"npm",