- Internal discovery and connection: Pods/containers (running in a Kubernetes cluster) must be able to easily discover and connect to endpoints for Kubernetes services on which they depend in a consistent way, irrespective of whether those services exist in a different kubernetes cluster within the same cluster federation. Hence-forth referred to as "cluster-internal clients", or simply "internal clients".
- External discovery and connection: External clients (running
outside a Kubernetes cluster) must be able to discover and connect
to endpoints for Kubernetes services on which they depend.
- External clients predominantly speak HTTP(S): External clients are most often, but not always, web browsers, or at least speak HTTP(S) - notable exceptions include Enterprise Message Busses (Java, TLS), DNS servers (UDP), SIP servers and databases)
- Find the "best" endpoint: Upon initial discovery and
connection, both internal and external clients should ideally find
"the best" endpoint if multiple eligible endpoints exist. "Best"
in this context implies the closest (by network topology) endpoint
that is both operational (as defined by some positive health check)
and not overloaded (by some published load metric). For example:
- An internal client should find an endpoint which is local to its own cluster if one exists, in preference to one in a remote cluster (if both are operational and non-overloaded). Similarly, one in a nearby cluster (e.g. in the same zone or region) is preferable to one further afield.
- An external client (e.g. in New York City) should find an endpoint in a nearby cluster (e.g. U.S. East Coast) in preference to one further away (e.g. Japan).
- Easy fail-over: If the endpoint to which a client is connected
becomes unavailable (no network response/disconnected) or
overloaded, the client should reconnect to a better endpoint,
somehow.
- In the case where there exist one or more connection-terminating load balancers between the client and the serving Pod, failover might be completely automatic (i.e. the client's end of the connection remains intact, and the client is completely oblivious of the fail-over). This approach incurs network speed and cost penalties (by traversing possibly multiple load balancers), but requires zero smarts in clients, DNS libraries, recursing DNS servers etc, as the IP address of the endpoint remains constant over time.
- In a scenario where clients need to choose between multiple load balancer endpoints (e.g. one per cluster), multiple DNS A records associated with a single DNS name enable even relatively dumb clients to try the next IP address in the list of returned A records (without even necessarily re-issuing a DNS resolution request). For example, all major web browsers will try all A records in sequence until a working one is found (TBD: justify this claim with details for Chrome, IE, Safari, Firefox).
- In a slightly more sophisticated scenario, upon disconnection, a smarter client might re-issue a DNS resolution query, and (modulo DNS record TTL's which can typically be set as low as 3 minutes, and buggy DNS resolvers, caches and libraries which have been known to completely ignore TTL's), receive updated A records specifying a new set of IP addresses to which to connect.
A Kubernetes application configuration (e.g. for a Pod, Replication Controller, Service etc) should be able to be successfully deployed into any Kubernetes Cluster or Ubernetes Federation of Clusters, without modification. More specifically, a typical configuration should work correctly (although possibly not optimally) across any of the following environments:
- A single Kubernetes Cluster on one cloud provider (e.g. Google Compute Engine, GCE).
- A single Kubernetes Cluster on a different cloud provider (e.g. Amazon Web Services, AWS).
- A single Kubernetes Cluster on a non-cloud, on-premise data center
- A Federation of Kubernetes Clusters all on the same cloud provider (e.g. GCE).
- A Federation of Kubernetes Clusters across multiple different cloud providers and/or on-premise data centers (e.g. one cluster on GCE/GKE, one on AWS, and one on-premise).
It should be possible to explicitly opt out of portability across some subset of the above environments in order to take advantage of non-portable load balancing and DNS features of one or more environments. More specifically, for example:
- For HTTP(S) applications running on GCE-only Federations, GCE Global L7 Load Balancers should be usable. These provide single, static global IP addresses which load balance and fail over globally (i.e. across both regions and zones). These allow for really dumb clients, but they only work on GCE, and only for HTTP(S) traffic.
- For non-HTTP(S) applications running on GCE-only Federations within a single region, GCE L4 Network Load Balancers should be usable. These provide TCP (i.e. both HTTP/S and non-HTTP/S) load balancing and failover, but only on GCE, and only within a single region. Google Cloud DNS can be used to route traffic between regions (and between different cloud providers and on-premise clusters, as it's plain DNS, IP only).
- For applications running on AWS-only Federations, AWS Elastic Load Balancers (ELB's) should be usable. These provide both L7 (HTTP(S)) and L4 load balancing, but only within a single region, and only on AWS (AWS Route 53 DNS service can be used to load balance and fail over across multiple regions, and is also capable of resolving to non-AWS endpoints).
Ubernetes cross-cluster load balancing is built on top of the following:
- GCE Global L7 Load Balancers provide single, static global IP addresses which load balance and fail over globally (i.e. across both regions and zones). These allow for really dumb clients, but they only work on GCE, and only for HTTP(S) traffic.
- GCE L4 Network Load Balancers provide both HTTP(S) and non-HTTP(S) load balancing and failover, but only on GCE, and only within a single region.
- AWS Elastic Load Balancers (ELB's) provide both L7 (HTTP(S)) and L4 load balancing, but only within a single region, and only on AWS.
- Google Cloud DNS (or any other
programmable DNS service, like
CloudFlare can be used to route
traffic between regions (and between different cloud providers and
on-premise clusters, as it's plain DNS, IP only). Google Cloud DNS
doesn't provide any built-in geo-DNS, latency-based routing, health
checking, weighted round robin or other advanced capabilities.
It's plain old DNS. We would need to build all the aforementioned
on top of it. It can provide internal DNS services (i.e. serve RFC
1918 addresses).
- AWS Route 53 DNS service can be used to load balance and fail over across regions, and is also capable of routing to non-AWS endpoints). It provides built-in geo-DNS, latency-based routing, health checking, weighted round robin and optional tight integration with some other AWS services (e.g. Elastic Load Balancers).
- Kubernetes L4 Service Load Balancing: This provides both a virtual cluster-local and a real externally routable service IP which is load-balanced (currently simple round-robin) across the healthy pods comprising a service within a single Kubernetes cluster.
- Kubernetes Ingress: A generic wrapper around cloud-provided L4 and L7 load balancing services, and roll-your-own load balancers run in pods, e.g. HA Proxy.
The Ubernetes API for load balancing should be compatible with the equivalent Kubernetes API, to ease porting of clients between Ubernetes and Kubernetes. Further details below.
To be useful, our load balancing solution needs to work properly with real client applications. There are a few different classes of those...
These are the most common external clients. These are all well-written. See below.
- Do a DNS resolution every time they connect.
- Don't cache beyond TTL (although a small percentage of the DNS servers on which they rely might).
- Do try multiple A records (in order) to connect.
- (in an ideal world) Do use SRV records rather than hard-coded port numbers.
Examples:
- all common browsers (except for SRV records)
- ...
- Don't do a DNS resolution every time they connect (or do cache beyond the TTL).
- Do try multiple A records
Examples:
- ...
- Only do a DNS lookup once on startup.
- Only try the first returned DNS A record.
Examples:
- ...
- Never do a DNS lookup - are pre-configured with a single (or possibly multiple) fixed server IP(s). Nothing else matters.
Each cluster hosts one or more Ubernetes master components (Ubernetes API servers, controller managers with leader election, and etcd quorum members. This is documented in more detail in a separate design doc: Kubernetes/Ubernetes Control Plane Resilience.
In the description below, assume that 'n' clusters, named 'cluster-1'... 'cluster-n' have been registered against an Ubernetes Federation "federation-1", each with their own set of Kubernetes API endpoints,so, "http://endpoint-1.cluster-1, http://endpoint-2.cluster-1 ... http://endpoint-m.cluster-n .
Ubernetes Services are pretty straight-forward. They're comprised of multiple equivalent underlying Kubernetes Services, each with their own external endpoint, and a load balancing mechanism across them. Let's work through how exactly that works in practice.
Our user creates the following Ubernetes Service (against an Ubernetes API endpoint):
$ kubectl create -f my-service.yaml --context="federation-1"
where service.yaml contains the following:
kind: Service
metadata:
labels:
run: my-service
name: my-service
namespace: my-namespace
spec:
ports:
- port: 2379
protocol: TCP
targetPort: 2379
name: client
- port: 2380
protocol: TCP
targetPort: 2380
name: peer
selector:
run: my-service
type: LoadBalancer
Ubernetes in turn creates one equivalent service (identical config to the above) in each of the underlying Kubernetes clusters, each of which results in something like this:
$ kubectl get -o yaml --context="cluster-1" service my-service
apiVersion: v1
kind: Service
metadata:
creationTimestamp: 2015-11-25T23:35:25Z
labels:
run: my-service
name: my-service
namespace: my-namespace
resourceVersion: "147365"
selfLink: /api/v1/namespaces/my-namespace/services/my-service
uid: 33bfc927-93cd-11e5-a38c-42010af00002
spec:
clusterIP: 10.0.153.185
ports:
- name: client
nodePort: 31333
port: 2379
protocol: TCP
targetPort: 2379
- name: peer
nodePort: 31086
port: 2380
protocol: TCP
targetPort: 2380
selector:
run: my-service
sessionAffinity: None
type: LoadBalancer
status:
loadBalancer:
ingress:
- ip: 104.197.117.10
Similar services are created in cluster-2 and cluster-3, each of which are
allocated their own spec.clusterIP, and status.loadBalancer.ingress.ip.
In Ubernetes federation-1, the resulting federated service looks as follows:
$ kubectl get -o yaml --context="federation-1" service my-service
apiVersion: v1
kind: Service
metadata:
creationTimestamp: 2015-11-25T23:35:23Z
labels:
run: my-service
name: my-service
namespace: my-namespace
resourceVersion: "157333"
selfLink: /api/v1/namespaces/my-namespace/services/my-service
uid: 33bfc927-93cd-11e5-a38c-42010af00007
spec:
clusterIP:
ports:
- name: client
nodePort: 31333
port: 2379
protocol: TCP
targetPort: 2379
- name: peer
nodePort: 31086
port: 2380
protocol: TCP
targetPort: 2380
selector:
run: my-service
sessionAffinity: None
type: LoadBalancer
status:
loadBalancer:
ingress:
- hostname: my-service.my-namespace.my-federation.my-domain.com
Note that the federated service:
- Is API-compatible with a vanilla Kubernetes service.
- has no clusterIP (as it is cluster-independent)
- has a federation-wide load balancer hostname
In addition to the set of underlying Kubernetes services (one per cluster) described above, Ubernetes has also created a DNS name (e.g. on Google Cloud DNS or AWS Route 53, depending on configuration) which provides load balancing across all of those services. For example, in a very basic configuration:
$ dig +noall +answer my-service.my-namespace.my-federation.my-domain.com
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.117.10
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.74.77
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.38.157
Each of the above IP addresses (which are just the external load balancer ingress IP's of each cluster service) is of course load balanced across the pods comprising the service in each cluster.
In a more sophisticated configuration (e.g. on GCE or GKE), Ubernetes automatically creates a GCE Global L7 Load Balancer which exposes a single, globally load-balanced IP:
$ dig +noall +answer my-service.my-namespace.my-federation.my-domain.com
my-service.my-namespace.my-federation.my-domain.com 180 IN A 107.194.17.44
Optionally, Ubernetes also configures the local DNS servers (SkyDNS) in each Kubernetes cluster to preferentially return the local clusterIP for the service in that cluster, with other clusters' external service IP's (or a global load-balanced IP) also configured for failover purposes:
$ dig +noall +answer my-service.my-namespace.my-federation.my-domain.com
my-service.my-namespace.my-federation.my-domain.com 180 IN A 10.0.153.185
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.74.77
my-service.my-namespace.my-federation.my-domain.com 180 IN A 104.197.38.157
If Ubernetes Global Service Health Checking is enabled, multiple service health checkers running across the federated clusters collaborate to monitor the health of the service endpoints, and automatically remove unhealthy endpoints from the DNS record (e.g. a majority quorum is required to vote a service endpoint unhealthy, to avoid false positives due to individual health checker network isolation).
So far we have a federated service defined, with a resolvable load balancer hostname by which clients can reach it, but no pods serving traffic directed there. So now we need a Federated Replication Controller. These are also fairly straight-forward, being comprised of multiple underlying Kubernetes Replication Controllers which do the hard work of keeping the desired number of Pod replicas alive in each Kubernetes cluster.
$ kubectl create -f my-service-rc.yaml --context="federation-1"
where my-service-rc.yaml contains the following:
kind: ReplicationController
metadata:
labels:
run: my-service
name: my-service
namespace: my-namespace
spec:
replicas: 6
selector:
run: my-service
template:
metadata:
labels:
run: my-service
spec:
containers:
image: gcr.io/google_samples/my-service:v1
name: my-service
ports:
- containerPort: 2379
protocol: TCP
- containerPort: 2380
protocol: TCP
Ubernetes in turn creates one equivalent replication controller (identical config to the above, except for the replica count) in each of the underlying Kubernetes clusters, each of which results in something like this:
$ ./kubectl get -o yaml rc my-service --context="cluster-1"
kind: ReplicationController
metadata:
creationTimestamp: 2015-12-02T23:00:47Z
labels:
run: my-service
name: my-service
namespace: my-namespace
selfLink: /api/v1/namespaces/my-namespace/replicationcontrollers/my-service
uid: 86542109-9948-11e5-a38c-42010af00002
spec:
replicas: 2
selector:
run: my-service
template:
metadata:
labels:
run: my-service
spec:
containers:
image: gcr.io/google_samples/my-service:v1
name: my-service
ports:
- containerPort: 2379
protocol: TCP
- containerPort: 2380
protocol: TCP
resources: {}
dnsPolicy: ClusterFirst
restartPolicy: Always
status:
replicas: 2
The exact number of replicas created in each underlying cluster will of course depend on what scheduling policy is in force. In the above example, the scheduler created an equal number of replicas (2) in each of the three underlying clusters, to make up the total of 6 replicas required. To handle entire cluster failures, various approaches are possible, including:
- simple overprovisioing, such that sufficient replicas remain even if a cluster fails. This wastes some resources, but is simple and reliable.
- pod autoscaling, where the replication controller in each cluster automatically and autonomously increases the number of replicas in its cluster in response to the additional traffic diverted from the failed cluster. This saves resources and is relatively simple, but there is some delay in the autoscaling.
- federated replica migration, where the Ubernetes Federation Control Plane detects the cluster failure and automatically increases the replica count in the remainaing clusters to make up for the lost replicas in the failed cluster. This does not seem to offer any benefits relative to pod autoscaling above, and is arguably more complex to implement, but we note it here as a possibility.
The implementation approach and architecture is very similar to Kubernetes, so if you're familiar with how Kubernetes works, none of what follows will be surprising. One additional design driver not present in Kubernetes is that Ubernetes aims to be resilient to individual cluster and availability zone failures. So the control plane spans multiple clusters. More specifically:
- Ubernetes runs it's own distinct set of API servers (typically one or more per underlying Kubernetes cluster). These are completely distinct from the Kubernetes API servers for each of the underlying clusters.
- Ubernetes runs it's own distinct quorum-based metadata store (etcd, by default). Approximately 1 quorum member runs in each underlying cluster ("approximately" because we aim for an odd number of quorum members, and typically don't want more than 5 quorum members, even if we have a larger number of federated clusters, so 2 clusters->3 quorum members, 3->3, 4->3, 5->5, 6->5, 7->5 etc).
Cluster Controllers in Ubernetes watch against the Ubernetes API server/etcd state, and apply changes to the underlying kubernetes clusters accordingly. They also have the anti-entropy mechanism for reconciling ubernetes "desired desired" state against kubernetes "actual desired" state.