mirror of
https://github.com/hanzoai/compute.git
synced 2026-10-30 14:50:41 +00:00
zeekayandHanzo Dev
bc02f7fe26
Two things visor could not do. It bound no socket, so a sibling reaching it
through zip.SocketPath got ErrNoPeer and cloud kept talking HTTP to
visor.hanzo.svc:19000 — there was nothing on the plane to convert to. And it
declared no typed op, so it was in no registry: no OpenAPI schema, no MCP tool,
no CLI verb, nothing zip.Here could call.
--zap now defaults to zip.SocketPath(visor.Name), so the canonical socket comes
up beside the HTTP edge and zip.Serving("visor") resolves. Name is stated once,
in pkg/visor, and is both the app name and the socket's — a second spelling of
a name is a peer nobody can find. --zap="" opts out.
/v1/health is the first typed op, and it fixes a probe that could not fail. The
liveness and readiness probes requested /api/health, which is not a route:
outside /v1/ it fell to TransparentStatic, which serves index.html as a 200. So
both probes measured the static file server, and a pod with its database face
down stayed in the Service. /v1/health answers 503 when object.Ready() cannot
reach the store — Ready lives in object because `store` is unexported there and
anywhere else would be a second copy of the backend-selection rule. It pings the
shared engine, not every org's: a check whose cost grows with the customer list
is an outage waiting for a big enough customer list.
Health registers AHEAD of the filter chain, and the position is the design.
ApiFilter would judge a credential-less probe as "anonymous" and authz.IsAllowed
panics on a nil Enforcer — so an authz mistake would read to kubelet as every
pod being sick. RecordMessage is the other half: an audit row per request, twice
every ten seconds, describing nothing. The routers test proves the bypass with a
control case, because an assertion about a bypass is only meaningful against a
route that is still gated.
The contract test is retargeted rather than retired. It was written as the
migration's acceptance criterion; the property it actually pins outlives the
migration — what visor serves equals what visor says it serves — so it now reads
that way and covers both registration functions. 73 routes, GET 32.
Co-authored-by: Hanzo Dev <[email protected]>
68 lines
1.5 KiB
YAML
68 lines
1.5 KiB
YAML
apiVersion: apps/v1
|
|
kind: Deployment
|
|
metadata:
|
|
name: visor
|
|
namespace: hanzo
|
|
labels:
|
|
app: visor
|
|
spec:
|
|
replicas: 2
|
|
strategy:
|
|
type: RollingUpdate
|
|
rollingUpdate:
|
|
maxSurge: 1
|
|
maxUnavailable: 0
|
|
selector:
|
|
matchLabels:
|
|
app: visor
|
|
template:
|
|
metadata:
|
|
labels:
|
|
app: visor
|
|
spec:
|
|
serviceAccountName: visor
|
|
imagePullSecrets:
|
|
- name: ghcr-secret
|
|
containers:
|
|
- name: visor
|
|
image: ghcr.io/hanzoai/visor:latest
|
|
ports:
|
|
- containerPort: 19000
|
|
protocol: TCP
|
|
env:
|
|
- name: RUNNING_IN_DOCKER
|
|
value: "true"
|
|
envFrom:
|
|
- secretRef:
|
|
name: visor-env
|
|
optional: true
|
|
volumeMounts:
|
|
- name: config
|
|
mountPath: /conf/app.conf
|
|
subPath: app.conf
|
|
resources:
|
|
requests:
|
|
cpu: 100m
|
|
memory: 256Mi
|
|
limits:
|
|
cpu: "1"
|
|
memory: 1Gi
|
|
livenessProbe:
|
|
httpGet:
|
|
path: /v1/health
|
|
port: 19000
|
|
initialDelaySeconds: 15
|
|
periodSeconds: 30
|
|
timeoutSeconds: 5
|
|
readinessProbe:
|
|
httpGet:
|
|
path: /v1/health
|
|
port: 19000
|
|
initialDelaySeconds: 10
|
|
periodSeconds: 10
|
|
timeoutSeconds: 5
|
|
volumes:
|
|
- name: config
|
|
secret:
|
|
secretName: visor-config
|