Refuse session windows beyond the server's 4-hour maximum in deploy policy
Foundation / checks (client) (push) Successful in 4s
Foundation / checks (checks) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 5s
Foundation / checks (network) (push) Successful in 5s
Foundation / checks (postgres) (push) Successful in 5s
Foundation / checks (tsan) (push) Successful in 4s
Foundation / checks (deploy) (push) Successful in 3m31s
Foundation / postgres-tools (push) Successful in 5s
Foundation / server-image (push) Successful in 5s
Foundation / image (push) Successful in 5s

A window ending 5.2 h after start (b2f224d) put evo-region into a
crashloop (error m10-window-too-long); the server only sees it at
startup. The deploy policy now requires REGION_SESSION_WINDOW to name its
UTC activation and end within 4 h of it, with two selftest cases (one is
the b2f224d value). DEPLOYMENT.md documents the follow-on quarantine:
syncs recorded as Deployment-only run no hooks, so a promoted build is
never retained; the provision hook is re-run from the rendered manifest.
VALIDATION records the incident and the recovery.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
dingo
2026-10-08 07:41:06 -04:00
co-authored by Claude Opus 5.5
parent 4a72897679
commit 47b84bce74
4 changed files with 76 additions and 0 deletions
+21
View File
@@ -134,6 +134,27 @@ exit 4
That is a build-retention failure, not a reason to reset the region. Re-run the
supported sync hook using the current rendered manifest.
A sync that Argo records with `resources: [Deployment/evo-region]` runs no hooks. On
2026-10-08 syncs requested with the `aevolution-dev-agent` token were recorded that way
even when no resource filter was requested, so a promotion never retained its build.
Check the operation with:
```bash
kubectl -n argocd get application aevolution-dev -o jsonpath='{.status.operationState.operation.sync}'
```
If it was limited to the Deployment, re-run the provision hook from the rendered manifest:
1. Render `deploy/kubernetes/dev` with the pinned kustomize and take
`Job/aevolution-region-provision`. Its image and `REGION_BUILD` must equal the
Deployment's.
2. Delete the old Job, create the rendered one and wait for `Complete`. The log shows
`already-provisioned`; the build was retained first.
3. `kubectl -n evolution-dev rollout restart deployment/evo-region`, then verify below.
`REGION_SESSION_WINDOW` must name its UTC activation (`dev-YYYYMMDDTHHMMSS-...`) and end
at most 4 hours after it. evo-server refuses a longer window at startup
(`error m10-window-too-long`), and the deploy policy check refuses it before push.
## 7. Verify the region
All of these must be true:
+33
View File
@@ -5085,3 +5085,36 @@ Executed:
- **Linux GCC 14:** `postgres-check.sh` with QUIC, all suites.
- **Linux Clang 19:** clang-tidy 19 on the M10 process tests (clean); clang-format 19.
- **Windows:** not applicable (the PostgreSQL process tests are Linux only).
## 2026-10-08 — evolution-dev crashloop and recovery
Incident:
- `b2f224d` set `REGION_SESSION_WINDOW` to end about 5.2 h after the pod started. evo-server
refuses windows longer than 4 h (`SessionWindow::max_duration_seconds`), so
`evo-region` exited with `error m10-window-too-long` in a loop.
- `4a72897` set a window of 3 h 50 min. The region then reached
`phase quarantined epoch=0 … exit 4`: the bot promotion `c3b40f4` (image `56647b49…`)
had never had its build retained. The provision hook still ran the previous image
(`2ccf6074…`, created 01:58Z).
- The two syncs requested afterwards with the agent token were both recorded by Argo
as limited to `Deployment/evo-region`, even though the second request carried no
resource filter. So no hooks ran. The cause is not visible with the agent's access
(possibly a guard on that account).
Recovery, within the agent contract (no reset, no SQL):
- Rendered `Job/aevolution-region-provision` from `deploy/kubernetes/dev` at
`4a72897` (image and `REGION_BUILD` `56647b49…`), deleted the stale Job and created it.
It completed with `already-provisioned … sequence=578 tick=12084`, the build retained
first.
- `rollout restart deployment/evo-region`. The region came up `phase running epoch=54
committed=12084`, `ready 1`, with history intact. Argo reports Synced and Healthy, and a
public-endpoint login attempt with an untrusted key was answered `unauthenticated`
(TLS and protocol fine).
Prevention:
- The deploy policy now refuses a `REGION_SESSION_WINDOW` that does not name its UTC
activation or that ends more than 4 h after it (two new selftest cases, one with the
`b2f224d` value).
- DEPLOYMENT.md documents the Deployment-only sync symptom and the hook re-run.
Executed: deploy policy check (0 failures) and selftest (46 cases, 0 failures).
+14
View File
@@ -687,6 +687,20 @@ for overlay, docs in renders.items():
check(text.count('--m10-listen') == 1 and '--m10-listen "$REGION_SESSION_LISTEN"' in text
and '--session-window "$REGION_SESSION_WINDOW"' in text,
f'{where}: one M10 listener and absolute session window, addressed only by GitOps env')
# The server refuses a window ending more than 4 h after it starts
# (SessionWindow::max_duration_seconds) and exits in a loop. The id carries
# its UTC activation (dev-YYYYMMDDTHHMMSS-...); the expiry must fall within
# 4 h of it. Checked here because the server sees it only at startup.
window = env.get('REGION_SESSION_WINDOW') or ''
if route_active and window:
import calendar, re, time
m = re.fullmatch(r'[0-9A-Za-z._-]*?(\d{8}T\d{6})Z?[0-9A-Za-z._-]*:(\d+)', window)
ok = False
if m:
start = calendar.timegm(time.strptime(m.group(1), '%Y%m%dT%H%M%S'))
ok = start < int(m.group(2)) <= start + 4 * 60 * 60
check(ok, f'{where}: REGION_SESSION_WINDOW must name its UTC activation and end '
'within 4 h of it (the server maximum)')
else:
check('--m10-listen' not in text and '--synthetic-inputs' not in text,
f'{where}: only evo-server may listen for sessions')
+8
View File
@@ -106,6 +106,14 @@ CASES = [
('non-RFC 1918 VIP without approval', lambda t: edit(t, ROUTE + 'service.yaml', '172.10.1.202',
'172.11.1.210'), 'internal address'),
# Review M9.8f P1: inline/JSON patches and namespace changes, seen only when rendered.
('session window longer than the server maximum', lambda t: inline_patch(
t, 'deploy/kubernetes/dev/kustomization.yaml',
env_patch('REGION_SESSION_WINDOW', 'dev-20261008T111342-32432:1791477222')),
'end within 4 h of it'),
('session window without an activation time', lambda t: inline_patch(
t, 'deploy/kubernetes/dev/kustomization.yaml',
env_patch('REGION_SESSION_WINDOW', 'dev-manual:1791477222')),
'end within 4 h of it'),
('inline patch: synthetic inputs on', lambda t: inline_patch(
t, 'deploy/kubernetes/dev/kustomization.yaml', env_patch('REGION_INPUTS', 'development-v1')),
'sessions must be dev only with the route'),