When configuring a production deployment of Istio, you need to answer a number of questions. Will the mesh be confined to a single cluster or distributed across multiple clusters? Will all the services be located in a single fully connected network, or will gateways be required to connect services across multiple networks? Is there a single control plane, potentially shared across clusters, or are there multiple control planes deployed to ensure high availability (HA)? Are all clusters going to be connected into a single multicluster service mesh or will they be federated into a multi-mesh deployment?
All of these questions, among others, represent independent dimensions of configuration for an Istio deployment.
- single or multiple cluster
- single or multiple network
- single or multiple control plane
- single or multiple mesh
In a production environment involving multiple clusters, you can use a mix of deployment models. For example, having more than one control plane is recommended for HA, but you could achieve this for a 3 cluster deployment by deploying 2 clusters with a single shared control plane and then adding the third cluster with a second control plane in a different network. All three clusters could then be configured to share both control planes so that all the clusters have 2 sources of control to ensure HA.
Choosing the right deployment model depends on the isolation, performance, and HA requirements for your use case. This guide describes the various options and considerations when configuring your Istio deployment.
The workload instances of your application run in one or more clusters. For isolation, performance, and high availability, you can confine clusters to availability zones and regions.
Production systems, depending on their requirements, can run across multiple clusters spanning a number of zones or regions, leveraging cloud load balancers to handle things like locality and zonal or regional fail over.
In most cases, clusters represent boundaries for configuration and endpoint discovery. For example, each Kubernetes cluster has an API Server which manages the configuration for the cluster as well as serving service endpoint information as pods are brought up or down. Since Kubernetes configures this behavior on a per-cluster basis, this approach helps limit the potential problems caused by incorrect configurations.
In Istio, you can configure a single service mesh to span any number of clusters.
In the simplest case, you can confine an Istio mesh to a single cluster. A cluster usually operates over a single network, but it varies between infrastructure providers. A single cluster and single network model includes a control plane, which results in the simplest Istio deployment.
Single cluster deployments offer simplicity, but lack other features, for example, fault isolation and fail over. If you need higher availability, you should use multiple clusters.
You can configure a single mesh to include multiple clusters. Using a multicluster deployment within a single mesh affords the following capabilities beyond that of a single cluster deployment:
- Fault isolation and fail over:
cluster-1goes down, fail over to
- Location-aware routing and fail over: Send requests to the nearest service.
- Various control plane models: Support different levels of availability.
- Team or project isolation: Each team runs its own set of clusters.
Multicluster deployments give you a greater degree of isolation and availability but increase complexity. If your systems have high availability requirements, you likely need clusters across multiple zones and regions. You can canary configuration changes or new binary releases in a single cluster, where the configuration changes only affect a small amount of user traffic. Additionally, if a cluster has a problem, you can temporarily route traffic to nearby clusters until you address the issue.
You can configure inter-cluster communication based on the network and the options supported by your cloud provider. For example, if two clusters reside on the same underlying network, you can enable cross-cluster communication by simply configuring firewall rules.
Within a multicluster mesh, all services are shared by default, according to the concept of namespace sameness. Traffic management rules provide fine-grained control over the behavior of multicluster traffic.
DNS with multiple clusters
When a client application makes a request to some host, it must first perform a
DNS lookup for the hostname to obtain an IP address before it can proceed with
In Kubernetes, the DNS server residing within the cluster typically handles
this DNS lookup, based on the configured
Istio uses the virtual IP returned by the DNS lookup to load balance
across the list of active endpoints for the requested service, taking into account any
Istio configured routing rules.
Istio uses either Kubernetes
Endpoint or Istio
configure its internal mapping of hostname to workload IP addresses.
This two-tiered naming system becomes more complicated when you have multiple clusters. Istio is inherently multicluster-aware, but Kubernetes is not (today). Because of this, the client cluster must have a DNS entry for the service in order for the DNS lookup to succeed, and a request to be successfully sent. This is true even if there are no instances of that service’s pods running in the client cluster.
To ensure that DNS lookup succeeds, you must deploy a Kubernetes
each cluster that consumes that service. This ensures that regardless of
where the request originates, it will pass DNS lookup and be handed to Istio
for proper routing.
This can also be achieved with Istio
ServiceEntry, rather than Kubernetes
Service. However, a
ServiceEntry does not configure the Kubernetes DNS server.
This means that DNS will need to be configured either manually or
with automated tooling such as the
Address auto allocation
feature of Istio DNS Proxying.
Istio uses a simplified definition of network to refer to workload instances that have direct reachability. For example, by default all workload instances in a single cluster are on the same network.
Many production systems require multiple networks or subnets for isolation and high availability. Istio supports spanning a service mesh over a variety of network topologies. This approach allows you to select the network model that fits your existing network topology.
In the simplest case, a service mesh operates over a single fully connected network. In a single network model, all workload instances can reach each other directly without an Istio gateway.
A single network allows Istio to configure service consumers in a uniform way across the mesh with the ability to directly address workload instances. Note, however, that for a single network across multiple clusters, services and endpoints cannot have overlapping IP addresses.
You can span a single service mesh across multiple networks; such a configuration is known as multi-network.
Multiple networks afford the following capabilities beyond that of single networks:
- Overlapping IP or VIP ranges for service endpoints
- Crossing of administrative boundaries
- Fault tolerance
- Scaling of network addresses
- Compliance with standards that require network segmentation
In this model, the workload instances in different networks can only reach each other through one or more Istio gateways. Istio uses partitioned service discovery to provide consumers a different view of service endpoints. The view depends on the network of the consumers.
This solution requires exposing all services (or a subset) through the gateway. Cloud vendors may provide options that will not require exposing services on the public internet. Such an option, if it exists and meets your requirements, will likely be the best choice.
Control plane models
An Istio mesh uses the control plane to configure all communication between workload instances within the mesh. Workload instances connect to a control plane instance to get their configuration.
In the simplest case, you can run your mesh with a control plane on a single cluster.
A cluster like this one, with its own local control plane, is referred to as a primary cluster.
Multicluster deployments can also share control plane instances. In this case, the control plane instances can reside in one or more primary clusters. Clusters without their own control plane are referred to as remote clusters.
To support remote clusters in a multicluster mesh, the control plane in a primary cluster must be accessible via a stable IP (e.g., a cluster IP). For clusters spanning networks, this can be achieved by exposing the control plane through an Istio gateway. Cloud vendors may provide options, such as internal load balancers, for providing this capability without exposing the control plane on the public internet. Such an option, if it exists and meets your requirements, will likely be the best choice.
In multicluster deployments with more than one primary cluster, each primary
cluster receives its configuration (i.e.,
DestinationRule, etc.) from the Kubernetes API Server residing in the same
cluster. Each primary cluster, therefore, has an independent source of
This duplication of configuration across primary clusters does require
additional steps when rolling out changes. Large production
systems may automate this process with tooling, such as CI/CD systems, in
order to manage configuration rollout.
Instead of running control planes in primary clusters inside the mesh, a service mesh composed entirely of remote clusters can be controlled by an external control plane. This provides isolated management and complete separation of the control plane deployment from the data plane services that comprise the mesh.
A cloud vendor’s managed control plane is a typical example of an external control plane.
For high availability, you should deploy multiple control planes across clusters, zones, or regions.
This model affords the following benefits:
Improved availability: If a control plane becomes unavailable, the scope of the outage is limited to only workloads in clusters managed by that control plane.
Configuration isolation: You can make configuration changes in one cluster, zone, or region without impacting others.
Controlled rollout: You have more fine-grained control over configuration rollout (e.g., one cluster at a time). You can also canary configuration changes in a sub-section of the mesh controlled by a given primary cluster.
Selective service visibility: You can restrict service visibility to part of the mesh, helping to establish service-level isolation. For example, an administrator may choose to deploy the
HelloWorldservice to Cluster A, but not Cluster B. Any attempt to call
HelloWorldfrom Cluster B will fail the DNS lookup.
The following list ranks control plane deployment examples by availability:
- One cluster per region (lowest availability)
- Multiple clusters per region
- One cluster per zone
- Multiple clusters per zone
- Each cluster (highest availability)
Endpoint discovery with multiple control planes
An Istio control plane manages traffic within the mesh by providing each proxy with the list of service endpoints. In order to make this work in a multicluster scenario, each control plane must observe endpoints from the API Server in every cluster.
To enable endpoint discovery for a cluster, an administrator generates a
remote secret and deploys it to each primary cluster in the mesh. The
remote secret contains credentials, granting access to the API server in the
The control planes will then connect and discover the service endpoints for
the cluster, enabling cross-cluster load balancing for these services.
By default, Istio will load balance requests evenly between endpoints in each cluster. In large systems that span geographic regions, it may be desirable to use locality load balancing to prefer that traffic stay in the same zone or region.
In some advanced scenarios, load balancing across clusters may not be desired. For example, in a blue/green deployment, you may deploy different versions of the system to different clusters. In this case, each cluster is effectively operating as an independent mesh. This behavior can be achieved in a couple of ways:
Do not exchange remote secrets between the clusters. This offers the strongest isolation between the clusters.
DestinationRuleto disallow routing between two versions of the services.
In either case, cross-cluster load balancing is prevented. External traffic can be routed to one cluster or the other using an external load balancer.
Identity and trust models
When a workload instance is created within a service mesh, Istio assigns the workload an identity.
The Certificate Authority (CA) creates and signs the certificates used to verify the identities used within the mesh. You can verify the identity of the message sender with the public key of the CA that created and signed the certificate for that identity. A trust bundle is the set of all CA public keys used by an Istio mesh. With a mesh’s trust bundle, anyone can verify the sender of any message coming from that mesh.
Trust within a mesh
Within a single Istio mesh, Istio ensures each workload instance has an appropriate certificate representing its own identity, and the trust bundle necessary to recognize all identities within the mesh and any federated meshes. The CA creates and signs the certificates for those identities. This model allows workload instances in the mesh to authenticate each other when communicating.
Trust between meshes
To enable communication between two meshes with different CAs, you must exchange the trust bundles of the meshes. Istio does not provide any tooling to exchange trust bundles across meshes. You can exchange the trust bundles either manually or automatically using a protocol such as SPIFFE Trust Domain Federation. Once you import a trust bundle to a mesh, you can configure local policies for those identities.
Istio supports having all of your services in a mesh, or federating multiple meshes together, which is also known as multi-mesh.
The simplest Istio deployment is a single mesh. Within a mesh, service names are
unique. For example, only one service can have the name
mysvc in the
namespace. Additionally, workload instances share a common identity since
service account names are unique within a namespace, just like service names.
A single mesh can span one or more clusters and one or more networks. Within a mesh, namespaces are used for tenancy.
Multiple mesh deployments result from mesh federation.
Multiple meshes afford the following capabilities beyond that of a single mesh:
- Organizational boundaries: lines of business
- Service name or namespace reuse: multiple distinct uses of the
- Stronger isolation: isolating test workloads from production workloads
You can enable inter-mesh communication with mesh federation. When federating, each mesh can expose a set of services and identities, which all participating meshes can recognize.
To avoid service naming collisions, you can give each mesh a globally unique mesh ID, to ensure that the fully qualified domain name (FQDN) for each service is distinct.
When federating two meshes that do not share the same trust domain, you must federate identity and trust bundles between them. See the section on Trust between meshes for more details.
In Istio, a tenant is a group of users that share common access and privileges for a set of deployed workloads. Tenants can be used to provide a level of isolation between different teams.
You can configure tenancy models to satisfy the following organizational requirements for isolation:
Istio supports three types of tenancy models:
A cluster can be shared across multiple teams, each using a different namespace. You can grant a team permission to deploy its workloads only to a given namespace or set of namespaces.
By default, services from multiple namespaces can communicate with each other, but you can increase isolation by selectively choosing which services to expose to other namespaces. You can configure authorization policies for exposed services to restrict access to only the appropriate callers.
Namespace tenancy can extend beyond a single cluster.
When using multiple clusters, the namespaces in each
cluster sharing the same name are considered the same namespace by default.
Service B in the
Team-1 namespace of cluster
Service B in the
Team-1 namespace of cluster
East refer to the same service, and Istio merges their
endpoints for service discovery and load balancing.
Istio supports using clusters as a unit of tenancy. In this case, you can give each team a dedicated cluster or set of clusters to deploy their workloads. Permissions for a cluster are usually limited to the members of the team that owns it. You can set various roles for finer grained control, for example:
- Cluster administrator
To use cluster tenancy with Istio, you configure each team’s cluster with its own control plane, allowing each team to manage its own configuration. Alternatively, you can use Istio to implement a group of clusters as a single tenant using remote clusters or multiple synchronized primary clusters. Refer to control plane models for details.
In a multi-mesh deployment with mesh federation, each mesh can be used as the unit of isolation.
Since a different team or organization operates each mesh, service naming
is rarely distinct. For example, a
Service C in the
foo namespace of
Team-1 and the
Service C service in the
foo namespace of cluster
Team-2 will not refer to the same service. The most common example is the
scenario in Kubernetes where many teams deploy their workloads to the
When each team has its own mesh, cross-mesh communication follows the concepts described in the multiple meshes model.