Istio installer is a modular, 'a-la-carte' installer for Istio. It is based on a fork of istio helm templates, refactored to increase modularity and isolation.
Goals:
-
Improve upgrade experience: users should be able to gradually roll upgrades, with proper canary for Istio components. It should be possible to deploy a new version while keeping the stable version in place and gradually migrate apps to the new version.
-
More flexibility: the new installer allows multiple 'environments', allowing applications to select a set of control plane settings and components. While the entire mesh respects the same APIs and config, apps may target different instances and variants of Istio.
-
Better security: separate istio components in different namespaces, allowing different teams or roles to manage different parts of Istio. For example a security team would maintain the root CA and policy, a telemetry team may only have access to Mixer-telemetry and Prometheus, and a different team may maintain the control plane components (which are very security sensitive).
The install is organized in 'environments' - each environment consists of a set of components in different namespaces that are configured to work together. Regardless of 'environment', workloads can talk with each other and follow the Istio configs, but each environment can use different versions and different defaults.
Kube-inject or the automatic injector are used to select the environment. The the later, the namespace label 'istio-env:NAME_OF_ENV' is used instead of 'istio-injected:true'. The name of the environment is defined as the namespace where the corresponding control plane is running. Pod annotations can also select a different control plane.
The new installer is intended to be modular and very explicit about what is installed. It has far more steps then the Istio installer - but each step is smaller and focused on a specific feature, and can be performed by different people/teams at different times.
It is strongly recommended that different namespaces are used, with different service accounts. In particular access to the security-critical production components (root CA, policy, control) should be locked down and restricted. The new installer allows multiple instances of policy/control/telemetry - so testing/staging of new settings and versions can be performed by a different role than the prod version.
The target is production users who want to select, tune and understand each binary that gets deployed, and select which combination to use.
Note that each component can be installed in parallel with an existing Istio 1.0 or 1.1 install in Istio-system. The new components will not interfere with existing apps, but can interoperate and it is possible to gradually move apps from Istio 1.0/1.1 to the new environments and across environments ( for example canary -> prod )
Note - there are still some cluster roles that may need to be fixed, most likely cluster permissions will need to move to the security component.
Each component in the new installer is optional. User can install the component defined in the new installer, use the equivalent component in istio-system, configured with the official installer, or use a different version or implementation.
For example you may use your own Prometheus and Graphana installs, or you may use a specialized/custom certificate provisioning tool, or use components that are centrally managed and running in a different cluster.
This is work in progress - building on top of the multi-cluster installer.
As an extreme, the goal is to be possible to run Istio workloads in a cluster without installing any Istio component in that cluster. Currently the minimum we require is the security provider (node agent or citadel).
The new installer recommends isolating components in different namespaces with different service accounts and access.
Recommended mode:
Singleton:
- istio-system: root CA and cert provisioning components.
- istio-cni: optional CNI (avoids requiring root/netadmin from workload pods)
Multi-environment components:
- istio-control: config, discovery, auto-inject. All impact the generated config including enforcement of policies and secure naming.
- istio-telemetry: mixer, kiali, tracing providers, graphana, prometheus. Custom install of prometheus, graphana can be used instead in dedicated namespaces.
- istio-policy
- istio-gateways - production domains should be in a separate namespace, to restrict access. It is possible to segregate gateways by the team that control access to the domain. Access to the gateway namespace provides access to certificates and control over domain delegation. The optional egress gateway provides control over outbound traffic.
In addition, it is recommended to have a second set of the multi-environment components to use for canary/testing new versions. In this doc I'll use a "istio-master" based set:
- istio-master: config, discovery, etc
- istio-telemetry-master
- istio-gateway-master
- istio-policy-master ...
For each component, there are 2 styles of installing, using 'helm + tiller' or 'helm template + kubectl apply --prune'.
Using 'kubectl --prune' is recommended:
helm template --namespace $NAMESPACE -n $COMPONENT $CONFIGDIR -f global.yaml | \
kubectl apply -n $NAMESPACE --prune -l release=$COMPONENT -f -
Using helm:
helm upgrade --namespace $NAMESPACE -n $COMPONENT $CONFIGDIR -f global.yaml The doc will use the "iop $NAMESPACE $COMPONENT $CONFIGDIR" helper from env.sh - which is the equivalent with the commands above.
TODO: replicas, cpu allocs, etc.
This is the first step of the install. Please do not remove or edit any CRD - config currently requires all CRDs to be present. On each upgrade it is recommended to reapply the file, to make sure you get all CRDs.
kubectl apply -f crds.yamlSecurity should be installed in istio-system, since it needs access to the root CA. For upgrades from the official installer, it is recommended to install the security component in istio-system, install the other components in different namespaces, migrate all workloads - and at the end uninstall the official installer, and lock down istio-system.
This is currently required if any MTLS is used. In future other Spifee implementations can be used, and it is possible to use other tools that create the expected certificates for Istio.
iop istio-system istio-system-security $IBASE/istio-system-securityImportant options: the 'dnsCerts' list allows associating DNS certs with specific service accounts. This should be used if you plan to use Galley or Sidecar injector in different namespaces. By default it supports "istio-control", "istio-master" namespaces used in the examples.
Access to the security namespace and istio-system should be highly restricted.
TODO. This is an optional step - CNI must run in a dedicated namespace, it is a 'singleton' and extremely security sensitive. Access to the CNI namespace must be highly restricted.
It is possible to add Istio-CNI later, and gradually migrate.
Control plane contains 3 components.
This can be run in any other cluster having the CRDs configured via CI/CD systems or other sync mechanisms. It should not be run in 'secondary' clusters, where the configs are not replicated.
Galley provides config access and validation. Only one environment should enable validation - it is not currently supported in multiple namespaces.
iop istio-control istio-config $IBASE/istio-control/istio-config --set configValidation=true
# Second Galley, using master version of istio
TAG=master-latest-daily HUB=gcr.io/istio-release iop istio-master istio-config-master $IBASE/istio-control/istio-configOther MCP providers can be used - currently the address and credentials need to match what galley is using.
Discovery, Policy and Telemetry components will need to be configured with the address of the config server - either in the local cluster or in a central cluster.
This can run in any cluster. A mesh should have at least one cluster should run Pilot or equivalent XDS server, and it is recommended to have Pilot running in each region and in multiple availability zones for multi cluster.
iop istio-control istio-discovery $IBASE/istio-control/istio-discovery
TAG=master-latest-daily HUB=gcr.io/istio-release iop istio-master istio-discovery-master $IBASE/istio-control/istio-discovery \
--set policy.enable=false \
--set global.istioNamespace=istio-master \
--set global.telemetryNamespace=istio-telemetry-master \
--set global.policyNamespace=istio-policy-maste
This is optional - kube-inject can be used instead.
If installed, namespaces can select the injector by setting 'istio-env' label on the namespace.
Only one auto-injector environment should have enableNamespacesByDefault=true, which will apply that environment to any namespace without an explicit istio-env label.
If istio-system has set 'enableNamespaceByDefault' you must set 'istio-inject: disabled' label to prevent istio-system from taking over. In this case, it is recommended to first install istio-control autoinject with the default disabled, test it, and move the default from istio-system to istio-control.
iop istio-control istio-autoinject $IBASE/istio-control/istio-autoinject --set enableNamespacesByDefault=true
# Second auto-inject using master version of istio
# Notice the different options
TAG=master-latest-daily HUB=gcr.io/istio-release iop istio-master istio-autoinject-master $IBASE/istio-control/istio-autoinject \
--set global.istioNamespace=istio-master
A cluster may use multiple Gateways, each with a different load balancer IP, domains and certificates.
Since the domain certificates are stored in the gateway namespace, it is recommended to keep each gateway in a dedicated namespace and restrict access.
For large-scale gateways it is optionally possible to use a dedicated pilot in the gateway namespace.
To support K8S ingress we currently use a separate namespace. In Istio 1.1, this requires using a dedicated Pilot instance in the ingress namespace. This will be fixed in future releases.
Note that running a dedicated Pilot for ingress/gateways is supported and recommended for very large sites, but in the case of K8S ingress it is currently required.
TODO - see example
TODO - see example
A number of helm test setups are general-purpose and should be installable in any cluster, to confirm Istio works properly and allow testing the specific install.