by Quinton Hoole (quinton@google.com)
Initial revision: 2015-03-05 Last updated: 2015-08-20 This doc: tinyurl.com/ubernetesv2 Original slides: tinyurl.com/ubernetes-slides Updated slides: tinyurl.com/ubernetes-whereto
Today, each Kubernetes cluster is a relatively self-contained unit, which typically runs in a single "on-premise" data centre or single availability zone of a cloud provider (Google's GCE, Amazon's AWS, etc).
Several current and potential Kubernetes users and customers have expressed a keen interest in tying together ("federating") multiple clusters in some sensible way in order to enable the following kinds of use cases (intentionally vague):
- "Preferentially run my workloads in my on-premise cluster(s), but automatically overflow to my cloud-hosted cluster(s) if I run out of on-premise capacity".
- "Most of my workloads should run in my preferred cloud-hosted cluster(s), but some are privacy-sensitive, and should be automatically diverted to run in my secure, on-premise cluster(s)".
- "I want to avoid vendor lock-in, so I want my workloads to run across multiple cloud providers all the time. I change my set of such cloud providers, and my pricing contracts with them, periodically".
- "I want to be immune to any single data centre or cloud availability zone outage, so I want to spread my service across multiple such zones (and ideally even across multiple cloud providers)."
The above use cases are by necessity left imprecisely defined. The rest of this document explores these use cases and their implications in further detail, and compares a few alternative high level approaches to addressing them. The idea of cluster federation has informally become known as "Ubernetes".
Four primary customer-driven use cases are explored in more detail. The two highest priority ones relate to High Availability and Application Portability (between cloud providers, and between on-premise and cloud providers).
Four primary federation primitives are identified (location affinity, cross-cluster scheduling, service discovery and application migration). Fortunately not all four of these primitives are required for each primary use case, so incremental development is feasible.
A central design concept in Kubernetes is that of a cluster. While loosely speaking, a cluster can be thought of as running in a single data center, or cloud provider availability zone, a more precise definition is that each cluster provides:
- a single Kubernetes API entry point,
- a consistent, cluster-wide resource naming scheme
- a scheduling/container placement domain
- a service network routing domain
- an authentication and authorization model.
The above in turn imply the need for a relatively performant, reliable and cheap network within each cluster.
There is also assumed to be some degree of failure correlation across a cluster, i.e. whole clusters are expected to fail, at least occasionally (due to cluster-wide power and network failures, natural disasters etc). Clusters are often relatively homogeneous in that all compute nodes are typically provided by a single cloud provider or hardware vendor, and connected by a common, unified network fabric. But these are not hard requirements of Kubernetes.
Other classes of Kubernetes deployments than the one sketched above are technically feasible, but come with some challenges of their own, and are not yet common or explicitly supported.
More specifically, having a Kubernetes cluster span multiple well-connected availability zones within a single geographical region (e.g. US North East, UK, Japan etc) is worthy of further consideration, in particular because it potentially addresses some of these requirements.
Let's name a few concrete use cases to aid the discussion:
"I want to preferentially run my workloads in my on-premise cluster(s), but automatically "overflow" to my cloud-hosted cluster(s) when I run out of on-premise capacity."
This idea is known in some circles as "cloudbursting".
Clarifying questions: What is the unit of overflow? Individual pods? Probably not always. Replication controllers and their associated sets of pods? Groups of replication controllers (a.k.a. distributed applications)? How are persistent disks overflowed? Can the "overflowed" pods communicate with their brethren and sistren pods and services in the other cluster(s)? Presumably yes, at higher cost and latency, provided that they use external service discovery. Is "overflow" enabled only when creating new workloads/replication controllers, or are existing workloads dynamically migrated between clusters based on fluctuating available capacity? If so, what is the desired behaviour, and how is it achieved? How, if at all, does this relate to quota enforcement (e.g. if we run out of on-premise capacity, can all or only some quotas transfer to other, potentially more expensive off-premise capacity?)
It seems that most of this boils down to:
- location affinity (pods relative to each other, and to other stateful services like persistent storage - how is this expressed and enforced?)
- cross-cluster scheduling (given location affinity constraints and other scheduling policy, which resources are assigned to which clusters, and by what?)
- cross-cluster service discovery (how do pods in one cluster discover and communicate with pods in another cluster?)
- cross-cluster migration (how do compute and storage resources, and the distributed applications to which they belong, move from one cluster to another)
- cross-cluster load-balancing (how does is user traffic directed to an appropriate cluster?)
- cross-cluster monitoring and auditing (a.k.a. Unified Visibility)
"I want most of my workloads to run in my preferred cloud-hosted cluster(s), but some are privacy-sensitive, and should be automatically diverted to run in my secure, on-premise cluster(s). The list of privacy-sensitive workloads changes over time, and they're subject to external auditing."
Clarifying questions:
- What kinds of rules determine which workloads go where?
- Is there in fact a requirement to have these rules be declaratively expressed and automatically enforced, or is it acceptable/better to have users manually select where to run their workloads when starting them?
- Is a static mapping from container (or more typically, replication controller) to cluster maintained and enforced?
- If so, is it only enforced on startup, or are things migrated between clusters when the mappings change?
This starts to look quite similar to "1. Capacity Overflow", and again seems to boil down to:
- location affinity
- cross-cluster scheduling
- cross-cluster service discovery
- cross-cluster migration
- cross-cluster monitoring and auditing
- cross-cluster load balancing
"My CTO wants us to avoid vendor lock-in, so she wants our workloads to run across multiple cloud providers at all times. She changes our set of preferred cloud providers and pricing contracts with them periodically, and doesn't want to have to communicate and manually enforce these policy changes across the organization every time this happens. She wants it centrally and automatically enforced, monitored and audited."
Clarifying questions:
- How does this relate to other use cases (high availability, capacity overflow etc), as they may all be across multiple vendors. It's probably not strictly speaking a separate use case, but it's brought up so often as a requirement, that it's worth calling out explicitly.
- Is a useful intermediate step to make it as simple as possible to migrate an application from one vendor to another in a one-off fashion?
Again, I think that this can probably be reformulated as a Capacity Overflow problem - the fundamental principles seem to be the same or substantially similar to those above.
"I want to be immune to any single data centre or cloud availability zone outage, so I want to spread my service across multiple such zones (and ideally even across multiple cloud providers), and have my service remain available even if one of the availability zones or cloud providers "goes down".
It seems useful to split this into multiple sets of sub use cases:
- Multiple availability zones within a single cloud provider (across
which feature sets like private networks, load balancing,
persistent disks, data snapshots etc are typically consistent and
explicitly designed to inter-operate).
- within the same geographical region (e.g. metro) within which network is fast and cheap enough to be almost analogous to a single data center.
- across multiple geographical regions, where high network cost and poor network performance may be prohibitive.
- Multiple cloud providers (typically with inconsistent feature sets, more limited interoperability, and typically no cheap inter-cluster networking described above).
The single cloud provider case might be easier to implement (although the multi-cloud provider implementation should just work for a single cloud provider). Propose high-level design catering for both, with initial implementation targeting single cloud provider only.
Clarifying questions: How does global external service discovery work? In the steady state, which external clients connect to which clusters? GeoDNS or similar? What is the tolerable failover latency if a cluster goes down? Maybe something like (make up some numbers, notwithstanding some buggy DNS resolvers, TTL's, caches etc) ~3 minutes for ~90% of clients to re-issue DNS lookups and reconnect to a new cluster when their home cluster fails is good enough for most Kubernetes users (or at least way better than the status quo), given that these sorts of failure only happen a small number of times a year?
How does dynamic load balancing across clusters work, if at all? One simple starting point might be "it doesn't". i.e. if a service in a cluster is deemed to be "up", it receives as much traffic as is generated "nearby" (even if it overloads). If the service is deemed to "be down" in a given cluster, "all" nearby traffic is redirected to some other cluster within some number of seconds (failover could be automatic or manual). Failover is essentially binary. An improvement would be to detect when a service in a cluster reaches maximum serving capacity, and dynamically divert additional traffic to other clusters. But how exactly does all of this work, and how much of it is provided by Kubernetes, as opposed to something else bolted on top (e.g. external monitoring and manipulation of GeoDNS)?
How does this tie in with auto-scaling of services? More specifically, if I run my service across n clusters globally, and one (or more) of them fail, how do I ensure that the remaining n-1 clusters have enough capacity to serve the additional, failed-over traffic? Either:
- I constantly over-provision all clusters by 1/n (potentially expensive), or
- I "manually" (or automatically) update my replica count configurations in the remaining clusters by 1/n when the failure occurs, and Kubernetes takes care of the rest for me, or
- Auto-scaling in the remaining clusters takes care of it for me automagically as the additional failed-over traffic arrives (with some latency). Note that this implies that the cloud provider keeps the necessary resources on hand to accommodate such auto-scaling (e.g. via something similar to AWS reserved and spot instances)
Up to this point, this use case ("Unavailability Zones") seems materially different from all the others above. It does not require dynamic cross-cluster service migration (we assume that the service is already running in more than one cluster when the failure occurs). Nor does it necessarily involve cross-cluster service discovery or location affinity. As a result, I propose that we address this use case somewhat independently of the others (although I strongly suspect that it will become substantially easier once we've solved the others).
All of the above (regarding "Unavailibility Zones") refers primarily to already-running user-facing services, and minimizing the impact on end users of those services becoming unavailable in a given cluster. What about the people and systems that deploy Kubernetes services (devops etc)? Should they be automatically shielded from the impact of the cluster outage? i.e. have their new resource creation requests automatically diverted to another cluster during the outage? While this specific requirement seems non-critical (manual fail-over seems relatively non-arduous, ignoring the user-facing issues above), it smells a lot like the first three use cases listed above ("Capacity Overflow, Sensitive Services, Vendor lock-in..."), so if we address those, we probably get this one free of charge.
As we saw above, a few common challenges fall out of most of the use cases considered above, namely:
Can the pods comprising a single distributed application be partitioned across more than one cluster? More generally, how far apart, in network terms, can a given client and server within a distributed application reasonably be? A server need not necessarily be a pod, but could instead be a persistent disk housing data, or some other stateful network service. What is tolerable is typically application-dependent, primarily influenced by network bandwidth consumption, latency requirements and cost sensitivity.
For simplicity, lets assume that all Kubernetes distributed applications fall into one of three categories with respect to relative location affinity:
- "Strictly Coupled": Those applications that strictly cannot be partitioned between clusters. They simply fail if they are partitioned. When scheduled, all pods must be scheduled to the same cluster. To move them, we need to shut the whole distributed application down (all pods) in one cluster, possibly move some data, and then bring the up all of the pods in another cluster. To avoid downtime, we might bring up the replacement cluster and divert traffic there before turning down the original, but the principle is much the same. In some cases moving the data might be prohibitively expensive or time-consuming, in which case these applications may be effectively immovable.
- "Strictly Decoupled": Those applications that can be indefinitely partitioned across more than one cluster, to no disadvantage. An embarrassingly parallel YouTube porn detector, where each pod repeatedly dequeues a video URL from a remote work queue, downloads and chews on the video for a few hours, and arrives at a binary verdict, might be one such example. The pods derive no benefit from being close to each other, or anything else (other than the source of YouTube videos, which is assumed to be equally remote from all clusters in this example). Each pod can be scheduled independently, in any cluster, and moved at any time.
- "Preferentially Coupled": Somewhere between Coupled and Decoupled. These applications prefer to have all of their pods located in the same cluster (e.g. for failure correlation, network latency or bandwidth cost reasons), but can tolerate being partitioned for "short" periods of time (for example while migrating the application from one cluster to another). Most small to medium sized LAMP stacks with not-very-strict latency goals probably fall into this category (provided that they use sane service discovery and reconnect-on-fail, which they need to do anyway to run effectively, even in a single Kubernetes cluster).
From a fault isolation point of view, there are also opposites of the above. For example a master database and it's slave replica might need to be in different availability zones. We'll refer to this a anti-affinity, although it is largely outside the scope of this document.
Note that there is somewhat of a continuum with respect to network cost and quality between any two nodes, ranging from two nodes on the same L2 network segment (lowest latency and cost, highest bandwidth) to two nodes on different continents (highest latency and cost, lowest bandwidth). One interesting point on that continuum relates to multiple availability zones within a well-connected metro or region and single cloud provider. Despite being in different data centers, or areas within a mega data center, network in this case is often very fast and effectively free or very cheap. For the purposes of this network location affinity discussion, this case is considered analogous to a single availability zone. Furthermore, if a given application doesn't fit cleanly into one of the above, shoe-horn it into the best fit, defaulting to the "Strictly Coupled and Immovable" bucket if you're not sure.
And then there's what I'll call absolute location affinity. Some applications are required to run in bounded geographical or network topology locations. The reasons for this are typically political/legislative (data privacy laws etc), or driven by network proximity to consumers (or data providers) of the application ("most of our users are in Western Europe, U.S. West Coast" etc).
Proposal: First tackle Strictly Decoupled applications (which can be trivially scheduled, partitioned or moved, one pod at a time). Then tackle Preferentially Coupled applications (which must be scheduled in totality in a single cluster, and can be moved, but ultimately in total, and necessarily within some bounded time). Leave strictly coupled applications to be manually moved between clusters as required for the foreseeable future.
I propose having pods use standard discovery methods used by external clients of Kubernetes applications (i.e. DNS). DNS might resolve to a public endpoint in the local or a remote cluster. Other than Strictly Coupled applications, software should be largely oblivious of which of the two occurs.
Aside: How do we avoid "tromboning" through an external VIP when DNS resolves to a public IP on the local cluster? Strictly speaking this would be an optimization for some cases, and probably only matters to high-bandwidth, low-latency communications. We could potentially eliminate the trombone with some kube-proxy magic if necessary. More detail to be added here, but feel free to shoot down the basic DNS idea in the mean time. In addition, some applications rely on private networking between clusters for security (e.g. AWS VPC or more generally VPN). It should not be necessary to forsake this in order to use Cluster Federation, for example by being forced to use public connectivity between clusters.
This is closely related to location affinity above, and also discussed there. The basic idea is that some controller, logically outside of the basic Kubernetes control plane of the clusters in question, needs to be able to:
- Receive "global" resource creation requests.
- Make policy-based decisions as to which cluster(s) should be used to fulfill each given resource request. In a simple case, the request is just redirected to one cluster. In a more complex case, the request is "demultiplexed" into multiple sub-requests, each to a different cluster. Knowledge of the (albeit approximate) available capacity in each cluster will be required by the controller to sanely split the request. Similarly, knowledge of the properties of the application (Location Affinity class -- Strictly Coupled, Strictly Decoupled etc, privacy class etc) will be required. It is also conceivable that knowledge of service SLAs and monitoring thereof might provide an input into scheduling/placement algorithms.
- Multiplex the responses from the individual clusters into an aggregate response.
There is of course a lot of detail still missing from this section, including discussion of:
- admission control
- initial placement of instances of a new service vs scheduling new instances of an existing service in response to auto-scaling
- rescheduling pods due to failure (response might be different depending on if it's failure of a node, rack, or whole AZ)
- data placement relative to compute capacity, etc.
Again this is closely related to location affinity discussed above, and is in some sense an extension of Cross-cluster Scheduling. When certain events occur, it becomes necessary or desirable for the cluster federation system to proactively move distributed applications (either in part or in whole) from one cluster to another. Examples of such events include:
- A low capacity event in a cluster (or a cluster failure).
- A change of scheduling policy ("we no longer use cloud provider X").
- A change of resource pricing ("cloud provider Y dropped their prices - lets migrate there").
Strictly Decoupled applications can be trivially moved, in part or in whole, one pod at a time, to one or more clusters (within applicable policy constraints, for example "PrivateCloudOnly").
For Preferentially Decoupled applications, the federation system must first locate a single cluster with sufficient capacity to accommodate the entire application, then reserve that capacity, and incrementally move the application, one (or more) resources at a time, over to the new cluster, within some bounded time period (and possibly within a predefined "maintenance" window). Strictly Coupled applications (with the exception of those deemed completely immovable) require the federation system to:
- start up an entire replica application in the destination cluster
- copy persistent data to the new application instance (possibly before starting pods)
- switch user traffic across
- tear down the original application instance
It is proposed that support for automated migration of Strictly Coupled applications be deferred to a later date.
These are often left implicit by customers, but are worth calling out explicitly:
- Software failure isolation between Kubernetes clusters should be retained as far as is practically possible. The federation system should not materially increase the failure correlation across clusters. For this reason the federation control plane software should ideally be completely independent of the Kubernetes cluster control software, and look just like any other Kubernetes API client, with no special treatment. If the federation control plane software fails catastrophically, the underlying Kubernetes clusters should remain independently usable.
- Unified monitoring, alerting and auditing across federated Kubernetes clusters.
- Unified authentication, authorization and quota management across clusters (this is in direct conflict with failure isolation above, so there are some tough trade-offs to be made here).
Two distinct potential architectural approaches have emerged from discussions thus far:
- An explicitly decoupled and hierarchical architecture, where the Federation Control Plane sits logically above a set of independent Kubernetes clusters, each of which is (potentially) unaware of the other clusters, and of the Federation Control Plane itself (other than to the extent that it is an API client much like any other). One possible example of this general architecture is illustrated below, and will be referred to as the "Decoupled, Hierarchical" approach.
- A more monolithic architecture, where a single instance of the Kubernetes control plane itself manages a single logical cluster composed of nodes in multiple availability zones and cloud providers.
A very brief, non-exhaustive list of pro's and con's of the two approaches follows. (In the interest of full disclosure, the author prefers the Decoupled Hierarchical model for the reasons stated below).
- Failure isolation: The Decoupled Hierarchical approach provides better failure isolation than the Monolithic approach, as each underlying Kubernetes cluster, and the Federation Control Plane, can operate and fail completely independently of each other. In particular, their software and configurations can be updated independently. Such updates are, in our experience, the primary cause of control-plane failures, in general.
- Failure probability: The Decoupled Hierarchical model incorporates numerically more independent pieces of software and configuration than the Monolithic one. But the complexity of each of these decoupled pieces is arguably better contained in the Decoupled model (per standard arguments for modular rather than monolithic software design). Which of the two models presents higher aggregate complexity and consequent failure probability remains somewhat of an open question.
- Scalability: Conceptually the Decoupled Hierarchical model wins here, as each underlying Kubernetes cluster can be scaled completely independently w.r.t. scheduling, node state management, monitoring, network connectivity etc. It is even potentially feasible to stack federations of clusters (i.e. create federations of federations) should scalability of the independent Federation Control Plane become an issue (although the author does not envision this being a problem worth solving in the short term).
- Code complexity: I think that an argument can be made both ways here. It depends on whether you prefer to weave the logic for handling nodes in multiple availability zones and cloud providers within a single logical cluster into the existing Kubernetes control plane code base (which was explicitly not designed for this), or separate it into a decoupled Federation system (with possible code sharing between the two via shared libraries). The author prefers the latter because it:
- Promotes better code modularity and interface design.
- Allows the code bases of Kubernetes and the Federation system to progress largely independently (different sets of developers, different release schedules etc).
- Administration complexity: Again, I think that this could be argued both ways. Superficially it would seem that administration of a single Monolithic multi-zone cluster might be simpler by virtue of being only "one thing to manage", however in practise each of the underlying availability zones (and possibly cloud providers) has it's own capacity, pricing, hardware platforms, and possibly bureaucratic boundaries (e.g. "our EMEA IT department manages those European clusters"). So explicitly allowing for (but not mandating) completely independent administration of each underlying Kubernetes cluster, and the Federation system itself, in the Decoupled Hierarchical model seems to have real practical benefits that outweigh the superficial simplicity of the Monolithic model.
- Application development and deployment complexity: It's not clear to me that there is any significant difference between the two models in this regard. Presumably the API exposed by the two different architectures would look very similar, as would the behavior of the deployed applications. It has even been suggested to write the code in such a way that it could be run in either configuration. It's not clear that this makes sense in practise though.
- Control plane cost overhead: There is a minimum per-cluster overhead -- two possibly virtual machines, or more for redundant HA deployments. For deployments of very small Kubernetes clusters with the Decoupled Hierarchical approach, this cost can become significant.
It is proposed that this look a lot like the existing Kubernetes API but be explicitly multi-cluster.
- Clusters become first class objects, which can be registered, listed, described, deregistered etc via the API.
- Compute resources can be explicitly requested in specific clusters, or automatically scheduled to the "best" cluster by the Cluster Federation control system (by a pluggable Policy Engine).
- There is a federated equivalent of a replication controller type (or perhaps a deployment), which is multicluster-aware, and delegates to cluster-specific replication controllers/deployments as required (e.g. a federated RC for n replicas might simply spawn multiple replication controllers in different clusters to do the hard work).
The Policy Engine decides which parts of each application go into each cluster at any point in time, and stores this desired state in the Desired Federation State store (an etcd or similar). Migration/Replication Controllers reconcile this against the desired states stored in the underlying Kubernetes clusters (by watching both, and creating or updating the underlying Replication Controllers and related Services accordingly).
This should ideally be delegated to some external auth system, shared by the underlying clusters, to avoid duplication and inconsistency. Either that, or we end up with multilevel auth. Local readonly eventually consistent auth slaves in each cluster and in the Cluster Federation control system could potentially cache auth, to mitigate an SPOF auth system.
The services comprising the Cluster Federation control plane) have to run somewhere. Several options exist here:
- For high availability Cluster Federation deployments, these
services may run in either:
- a dedicated Kubernetes cluster, not co-located in the same availability zone with any of the federated clusters (for fault isolation reasons). If that cluster/availability zone, and hence the Federation system, fails catastrophically, the underlying pods and applications continue to run correctly, albeit temporarily without the Federation system.
- across multiple Kubernetes availability zones, probably with some sort of cross-AZ quorum-based store. This provides theoretically higher availability, at the cost of some complexity related to data consistency across multiple availability zones.
- For simpler, less highly available deployments, just co-locate the Federation control plane in/on/with one of the underlying Kubernetes clusters. The downside of this approach is that if that specific cluster fails, all automated failover and scaling logic which relies on the federation system will also be unavailable at the same time (i.e. precisely when it is needed). But if one of the other federated clusters fails, everything should work just fine.
There is some further thinking to be done around the data consistency model upon which the Federation system is based, and it's impact on the detailed semantics, failure and availability characteristics of the system.
Identify concrete applications of each use case and configure a proof of concept service that exercises the use case. For example, cluster failure tolerance seems popular, so set up an apache frontend with replicas in each of three availability zones with either an Amazon Elastic Load Balancer or Google Cloud Load Balancer pointing at them? What does the zookeeper config look like for N=3 across 3 AZs -- and how does each replica find the other replicas and how do clients find their primary zookeeper replica? And now how do I do a shared, highly available redis database? Use a few common specific use cases like this to flesh out the detailed API and semantics of Cluster Federation.