Ingress with Istio
Ingress with Istio
Seldon Core can be used in conjunction with istio. Istio provides an ingress gateway which Seldon Core can automatically wire up new deployments to. The steps to using istio are described below.
Install Seldon Core Operator
Ensure when you install the seldon-core operator via Helm that you enabled istio. For example:
helm install seldon-core seldon-core-operator --set istio.enabled=true --repo https://storage.googleapis.com/seldon-charts --set usageMetrics.enabled=true
You need an istio gateway installed in the istio-system
namespace. By default we assume one called seldon-gateway. For example you can create this with the following yaml:
apiVersion: networking.istio.io/v1alpha3
kind: Gateway
metadata:
name: seldon-gateway
namespace: istio-system
spec:
selector:
istio: ingressgateway # use istio default controller
servers:
- port:
number: 80
name: http
protocol: HTTP
hosts:
- "*"
If you want to want to create SSL based gateway, create your signed certificate or actual signed certificate (for example named fullchain.pem), key (privkey.pem) and then run follwing commands to get SSL gateway. Assuming we're not using cert-manager then create self-signed certificate with
openssl req -nodes -x509 -newkey rsa:4096 -keyout privkey.pem -out fullchain.pem -days 365 -subj "/C=GB/ST=GreaterLondon/L=London/O=SeldonSerra/OU=MLOps/CN=localhost"
Import certificate and key as a secret into istio-system namespace
kubectl create -n istio-system secret tls seldon-ssl-cert --key=privkey.pem --cert=fullchain.pem
and create SSL Istio Gateway using following YAML file
apiVersion: networking.istio.io/v1alpha3
kind: Gateway
metadata:
name: seldon-gateway
namespace: istio-system
spec:
selector:
istio: ingressgateway # use istio default controller
servers:
- hosts:
- '*'
port:
name: https
number: 443
protocol: HTTPS
tls:
credentialName: seldon-ssl-cert
mode: SIMPLE
If you have your own gateway you will use then you can provide the name when installing the seldon operator. For example if your gateway is called mygateway
you can install the operator with:
helm install seldon-core seldon-core-operator --set istio.enabled=true --set istio.gateway=mygateway --repo https://storage.googleapis.com/seldon-charts --set usageMetrics.enabled=true
You can also provide the gateway on a per Seldon Deployment resource basis by providing it with the annotation seldon.io/istio-gateway
.
Istio Configuration Annotation Reference
seldon.io/istio-gateway:<gateway name>
istio-system/seldon-gateway
The gateway to use for this deployment. If no namespace prefix is applied it will refer to the namespace of the Seldon Deployment.
seldon.io/istio-retries
None
The number of istio retries
seldon.io/istio-retries-timeout
None
The per try timeout if istio retries is set
seldon.io/istio-host
*
The Host for istio Virtual Service
All annotations should be placed in spec.annotations
or metadata.annotations
. spec.annotations
will take precedence.
Traffic Routing
Istio has the capability for fine grained traffic routing to your deployments. This allows:
canary updates
green-blue deployments
A/B testing
shadow deployments
More information can be found in our examples, including canary updates.
Configuring Authentication/Authorization
To force clients to authenticate/authorize themselves in order to access the seldon model deployments, you can leverage Istio's RequestAuthentication
and AuthorizationPolicy
. This will deny or accept requests to the model depending on specified conditions that you designated in the policies. More information can be found here.
You can set the policies to target all the models belonging to a specific namespace, but you must be using istio sidecar proxy, and ensure your seldon operator configuration has the following:
istio:
enabled: true
tlsMode: STRICT
When you've set up an AuthorizationPolicy
, this will disrupt Prometheus from scraping metrics. Two proposed options to resolve this issue are:
You can specify that you want to allow GET requests to the prometheus endpoint in the
AuthorizationPolicy
Example:
- to:
- operation:
methods: ["GET"]
paths: ["/prometheus"]
ports: ["6000", "8000", "6001"]
You can also exclude ports in your Istio Operator configuration
proxy:
autoInject: enabled
clusterDomain: cluster.local
componentLogLevel: misc:error
enableCoreDump: false
excludeInboundPorts: ""
excludeOutboundPorts: "15021"
Troubleshoot
If you saw errors like Failed to generate bootstrap config: mkdir ./etc/istio/proxy: permission denied
, it's probably because you are running istio version <= 1.6. Istio proxy sidecar by default needs to run as root (This changed in version >= 1.7, non-root is the default) You can fix this by changing defaultUserID=0
in your helm chart, or add the following securityContext
to your istio proxy sidecar.
securityContext:
runAsUser: 0
Using the Istio Service Mesh
Istio can also be used to direct traffic internal to the cluster, rather than using it as an ingress (traffic from outside the cluster).
To do this, the Virutal Services Seldon will create need to be attached to the "special" Gateway named mesh
. This applies the routing rules to traffic inside the mesh without needing to route through a Gateway.
Due to limitations in Istio (as of v1.11.3), virtual services in the local mesh can only apply to one Host. (see their docs here). Therefor, a unique service is required for each Graph, which can be achieved by setting the seldon.io/svc-name
annotation in the main predictor.
Here's an example SeldonDeployment
that will utilize the internal mesh networking to split traffic between two predictors, 75% to the first, 25% to the second:
apiVersion: machinelearning.seldon.io/v1
kind: SeldonDeployment
metadata:
labels:
app: seldon
name: canary-example-1
namespace: my-ns
spec:
annotations:
seldon.io/istio-gateway: mesh # NOTE
seldon.io/istio-host: canary-example-1 # NOTE
name: canary-example-1
predictors:
- annotations:
seldon.io/svc-name: canary-example-1 # NOTE
componentSpecs:
- spec:
containers:
- image: seldonio/mock_classifier:1.11.0
imagePullPolicy: IfNotPresent
name: classifier
securityContext:
readOnlyRootFilesystem: false
terminationGracePeriodSeconds: 1
graph:
endpoint:
type: REST
name: classifier
type: MODEL
labels:
sidecar.istio.io/inject: "true"
name: main
replicas: 1
traffic: 75
- componentSpecs:
- spec:
containers:
- image: seldonio/mock_classifier:1.11.0
imagePullPolicy: IfNotPresent
name: classifier
terminationGracePeriodSeconds: 1
graph:
endpoint:
type: REST
name: classifier
type: MODEL
labels:
sidecar.istio.io/inject: "true"
name: canary
replicas: 1
traffic: 25
A few key things to point out:
A unique service is created for the main (first) predictor named
canary-example-1
. This service cannot collide with any other services in the namespace. This service could be a service not created via the SeldonDeployment, but also must match the necessary Istio routing rules.The above service is referenced in the annotations in
spec
by specify ing the host as follows:seldon.io/istio-host: canary-example-1
. This will set the host in the Istio Virutal Service to be the newly created service.The gateway is specified as
seldon.io/istio-gateway: mesh
to utilize this routing in the Istio Mesh. NOTE: In order to call this service, and have the appropriate routing take place, the Client must also be inside the mesh. This is accomplished by injecting the Istio Sidecar into the pod of the client.
From within the cluster, and inside a pod that is inside the mesh, a call like the following will work, as well as split traffic between the two predictors:
curl -X POST -H 'Content-Type: application/json' \
-d '{"data": { "names": ["a", "b"], "ndarray": [[1,2]]}}' \
http://mysvcname:8000/seldon/batest/canary-example-1/api/v1.0/predictions
Last updated
Was this helpful?