Making sense of serverless

The now annual Serverlessconf NYC (2019) event was held in NYC in October. This was a great opportunity to assess the current state of ‘serverless’ – what it is, why it’s important, whether and how it should factor in enterprise cloud planning, and what challenges there are around the serverless space.

The content below is informed by presentations given during the conference, but opinions and conclusions are my own.

What exactly is serverless

‘Serverless’ architectures are often correlated with ‘functions-as-a-service’ (FaaS) but ‘serverless’ extends far beyond FaaS. At the simplest – and most pragmatic – level, ‘serverless’ architectures do not have any ‘servers’ (i.e., long-running virtual and/or physical machines) to manage – i.e., no OS patches, no upgrades, no capacity to manage, etc. Operational overhead related to infrastructure management is zero. This means development teams are 100% responsible for their serverless application architectures.

Of course, the servers do not disappear: they are merely 100% managed by ‘someone else’, with little or no engagement required between development teams and those who manage the serverless infrastructure.

Architecturally, serverless is more than simply no infrastructure to manage. ‘Pure’ serverless architectures exhibit the following characteristics:

  • Are event-driven
  • Are function-oriented
  • Use managed services
  • Are scalable to zero and up

It’s worth noting that managed services in a a serverless architecture may not themselves be using ‘serverless’ architectures. However, the technology used to deliver each managed service is not exposed in a specific serverless solution architecture.

Why is serverless important?

From a purely economic perspective, serverless enables the introduction of true value-driven businesses (see Simon Wardley’s perspective here), enabling a whole new economics around automation and infrastructure. Essentially, if no value is being delivered, then no cost is being incurred on a serverless architecture (assuming usage of services constitutes ‘value’ delivered).

Because of the traditionally high cost of supporting and maintaining application infrastructure, businesses have historically put a lot of effort into planning new software; new features tend to increase both change and operating costs, usually more than linearly. This makes businesses more change-averse over time, leading to over-planning, a lack of agility and significantly reduced pace of innovation delivery.

Currently (as of 2019), the fashion is to invest in cloud computing to dampen change and operating costs – although this investment is still predominantly in the IaaS space (i.e., using the cloud’s superior economics for compute, storage and networking). Enterprises moving beyond IaaS are faced with either committing to a specific cloud provider’s infrastructure PaaS solutions (such as AWS ECS, Fargate, etc), or investing a lot of effort in building and operating their own orchestration and runtime solutions (usually using Kubernetes as the enabler).

But the war to deliver customer value is being fought above the runtime or infrastructure PaaS, as suggested by Simon Wardley. The risk to many enterprises is that they win the battle for infrastructure PaaS but lose the war to deliver customer value. Indeed, one conference presentation (entitled ‘Killing Kubernetes’) gave a real-world example of how running down the Kubernetes path prematurely can cause a team to lose sight of the customer value to be delivered. (The team in the end decided to go full serverless, and ditch Kubernetes.)

Using a tool likely Wardley Maps enables clarity of thought with respect to critical platform-level components, and which battles it makes sense to fight, vs leveraging what industry innovation will provide.

As Ben Kehoe describes, the point of serverless is to provide focus on business value – it’s not about functions, technology, cost or operations.

Changing Build vs Buy Mindsets

Many enterprises have a policy of ‘buy over build’ – i.e., buy (and customize) a solution rather than build a solution. Customized off-the-shelf solutions have their advantages, but often ultimately lead to businesses being constrained by vendor roadmaps, or by the cost of upgrading/keeping pace with vendor advances. In particular, vendor software is optimized for configurability, whereas what enterprises need is extensible software rather than configurable software.

Serverless provides organizations which do not have a depth in engineering expertise with a path towards ‘build over buy’. Functions and workflows proprietary to a business can be done with serverless functions, while use of managed services can minimize the need for infrastructure expertise. Integrating 3rd party software-as-a-service solution also becomes second nature in a serverless environment, particularly with the advent of integration tools such as AWS EventBridge. Such an architecture is readily extensible, and better suited to meet enterprise needs.

Serverless and the Enterprise

Serverless architectures are being actively used by all kinds of businesses with great success (aCloudGuru, a sponsor of ServerlessConf NYC, is just one example of a successful serverless user). Architectures which rely exclusively on managed services (such as AWS S3, Aurora, Lambda, StepFunctions, DynamoDB, etc) can be considered serverless.

Many enterprises are partially serverless on the cloud, choosing to leverage a cloud provider’s managed service offering as part of a ‘serverful’ solution architecture (e.g., using S3 with EC2). But these architectures do not provide the full benefits of a true ‘serverless’ architecture, as considerable effort is still required to manage the non-serverless elements.

It should be noted as well that just because traditional enterprise software is made available as a ‘managed service’, doesn’t mean that the enterprise overhead of managing that service is reduced: if the cloud provider still exposes all the configurable aspects of the software, there will not be a significant benefit in moving to the managed service. (Microsoft enterprise applications offered on Azure seem to suffer from this affliction.)

Fundamentally, serverless is not yet ready to take on all enterprise workloads – there are many constraints and conflicts between standardizing the serverless runtimes (necessary to allow them to be managed efficiently at scale) and the customization needs of enterprises. In particular, how systems manage state with sufficient performance is likely to remain a challenge – although this is certainly solvable, as emerging architectural best practices for a serverless world establish themselves (the traditional model-view-controller model being a poor model for serverless applications).

For most enterprises, therefore, business solutions should be planned as if performant platform solutions exists, being clear on what the functions are, what the managed services are, and what platform capabilities are assumed. These can then drive further investment decisions to build out (or buy/rent) these capabilities. A ‘serverless first’ mindset is key to this.

Underpinning all these is organizational design – in particular, the concepts and ideas espoused in Team Topologies map very well to this approach.

Decomposing the Monolith

A key use case for serverless is enabling the decomposition of legacy monolithic architectures. Most enterprises do not have the skills or expertise to successfully migrate complex monolithic architectures to microservices, as this requires some skills in developing on and managing highly distributed systems. While technologies like CloudFoundry and SpringBoot go a long way towards minimizing the cognitive load for application developers, organizations require considerable investment to make these technologies available as true managed services across an enterprise.

Serverless offers a route to decompose monolithic architectures without first building out the full capabilities needed to deploy serverful microservice architectures. It allows enterprises to experiment with service-based business solutions without incurring significant or hard-to-reverse costs in infrastructure and/or skills. Once a decomposed architecture begins to prove its worth, it may be unavoidable (for now) to move to serverful microservices at the back-end to scale out, but the business value proposition should be clear by then.

Serverless Challenges

Serverless architectures have their own challenges that organizations need to be prepared to handle, which are different from the challenges that building serverful architectures have.

Key challenges exist around:

  • Security
  • Local development and testing
  • Debugging, tracing, monitoring and alerting
  • Limit Management
  • Resilience
  • Lock-in
  • Integration testing
  • Serverless infrastructure-as-code

The above are the challenges specifically raised during the conference. Other challenges may yet reveal themselves.

Security

Serverless requires a different security model than traditional infrastructure. Specifically, security for serverless centers around security of functions, security of data, and security of configuration.

Key attack surfaces for serverless are event data injection, unauthorized deployments, and dependency poisoning. In particular, over-privileged permissions present a significant surface attack area. A good list of attack surfaces is published by Palo Alto/Puresec, a sponsor of the conference.

Serverless components therefore need their own security solutions, as part of an over-arching defense-in-depth security strategy.

Local development and testing

By definition, managed services cannot be available on local (laptop/desktop) development environments, as neither are serverless runtimes such as lambda. Instead, development is expected to happen directly on the cloud, which can cause issues for developers who are periodically disconnected from the internet.

For some development teams, the ability to code and test away from the cloud is important, and in this regard, cloud providers are beginning to standardize more on OCI containers in their serverless runtimes, and allow developers to run these containers locally on their laptop, as well as on standard orchestrated environments such as Kubernetes. Azure and GCP seems to be leading the way in this space, but AWS is always improving its lambda run-times and offering developers more ways to customize them, so this may eventually lead to AWS offering the same features.

The challenge, however, will be to maintain the benefits of serverless while avoiding requiring teams end up managing containers as the new ‘servers’…a trap many teams are likely to fall into.

Debugging, tracing, monitoring and alerting

The challenges here are not unique to serverless – microservices architectures have these challenges in spades. While cloud providers typically provide managed services to assist with these (e.g, AWS X-Ray, AWS CloudWatch, etc), a rich eco-system of 3rd parties also help to address these needs.

In general, while it is possible to get by with provider-native solutions, it may be best to augment team capabilities with a vendor solution, such as Lumigo, Serverless framework, Datadog, Epsagon, etc.

Limit Management

All serverless services have limits, usually defined per account. This protects rogue applications from over-loading lambdas or managed services (such as AWS DynamoDB).

Usually, limits can be increased, but may need a service request to the cloud provider. Limits can also be imposed per account at an enterprise level (for example, via AWS Organizational Units).

It is important that the service limits are known and understood, as incorrectly assuming no limits may have a material impact on a solution architecture. While serverless solutions can scale, they cannot scale infinitely.

Resilience

Resilience for managed services is different from resilience of functions-as-a-service. Managed services need to be available at all times – but the manner and means by which such services maintain availability is generally opaque to the user. Some services may be truly global, but cloud providers tend to make managed service resilient within a specific region (through multiple availability-zones in a region), which requires solution architectures to allow for redundancy across multiple regions in the event a single region fails in its entirety. Recovery in these scenarios may not need to be 100% automated, dependending on recovery time objectives.

For functions-as-a-service (lambdas), if an invocation fails, it should be safe for the runtime to try again (i.e., idempotent processing of events). So the runtime provides most of the resilience.

However, if a lambda depends on a ‘traditional’ service (i.e., not in itself dynamically scalable), there may be resilience issues. For example, a lambda connecting to a traditional relational database via SQL may run out of available server-side connections.

Resource constraints applies to any API which is not fronting a serverless architecture. So lambdas need to ensure sufficient resilience (e.g., circuit breaker pattern) is built-in so that constraints in other APIs do not cause the lambda to fail.

Lock-in

Many enterprises are reluctant to use a particular cloud providers serverless model as they tend to be very proprietary and cloud-provider specific, and therefore moving to another cloud, or enabling a solution to run on any cloud-provider, could involve considerable re-engineering expense.

Firms which are constrained by regulatory or other drivers to avoid provider lock-in have options available. Firms can use multi-cloud serverless frameworks such as Serverless. In addition, there are vendors appearing in the multi-cloud messaging space, with vendors like TriggerMesh offering a serverless multi-cloud event bus.

Some cloud providers are making source code for their lambda services available publically – for example Google Cloud Functions and Azure Functions. Open-source serverless solutions such as Google’s Knative and OpenFaaS are also available. In addition, some vendors, such as Platform9 provide a completely independent solution for lambdas, for organizations which want to deploy lambdas internally – for example, on Kubernetes.

Other mechanisms to minimize the effect of lock-in include the use of standard OCI or docker containers to host serverless functions, which may allow containers to run in other orchestration environments without requiring significant rework. (This doesn’t really help if the container relies on external provider-specific managed services, however.)

Regardless of steps taken to avoid lock-in, some cloud providers may include managed services that may be proprietary to them: once software is built to leverage such a managed service, you have a form of lock-in (in much the same way, for example, you may be locked-in to Oracle or Microsoft databases once you commit to using proprietary features of them).

As such, focusing on avoiding lock-in is, for many firms, going to result in unnecessary complexity. It may be better to exploit a given cloud provider, and manage the business risk associated with a complete provider outage. For regulated services, however, regulators may want to ensure regulated firms are not overly concentrated in one provider.

Integration testing

Integration testing is never easy to fully automate – it is partially reason why there is so much focus on microservices, as each microservice is an independently testable and deployable component. The same applies for lambdas. But each lambda may itself depend on multiple managed services, so how to test those? An excellent piece on serverless testing by Paul Johnston describes the challenge well:

The test boundaries for unit testing a FaaS Function appears to be very close to an integration test versus a component test within a microservice approach.

In essence, because all serverless features are available through APIs, it *should* be easier to build and maintain integration tests, but for now it is still harder than it ought to be.

Serverless Infrastructure as Code

There is a growing sense of dissatisfaction with the limitations of traditional YAML-based configuration languages with respect to serverless – in particular, that the lifecycle and dependencies of serverless resources are not properly represented in existing infrastructure configuration languages. Ben Kehoe gives a flavor of the issues, but this is a complex topic likely to get more attention in the future.

Summary

The key value proposition of serverless is that it permits application developers to focus more on delivering customer value, and to spend less time dealing with infrastructure concerns such as managing servers.

The time is right for organizations to start entering the serverless mindset, and to assess business solutions in the strategic context offered by serverless – whether that means ultimately using external services or designing internal services in a serverless way.

ServerlessConf 2019 was informative, and the presentations were generally accessible to a wide audience. For many presentations, it was not necessary to be a cloud engineer to understand the content and to appreciate the potential transformational opportunities of serverless in the coming years.

I hope that in future events, a broader coalition of business strategic planners and do-ers will be in attendance. It is definitely not a Kubecon, but engineering advances made at events like Kubecon will make the serverless vision possible, while freeing serverless practitioners from the complexities of managing containers, orchestrators and servers.

Making sense of serverless

Achieving Service Resilience on Cloud-Reliant Infrastructure

[tl;dr Rather than require banks to have multi-cloud or on-premise redundant strategies for critical digital services delivered via cloud, regulators should require firms to instead adhere to ‘well-architected’ guidelines to ensure technology resilience for a given provider or PaaS solution, as part of an overall industry-wide framework of service resilience that can accommodate very rare but impactful events such as cloud provider failure.]

Resilience is the ability of a ‘system’ of processes to recover quickly from a significant shock, ideally without any noticeable impact on those depending on that system. Systems are comprised of people, technologies and environments (e.g., laws, physical locations, etc), any of which can ‘fail’.

Financial regulators are particularly concerned with resilience, as failures in banking systems can cause business disruption, serious personal inconvenience, and even the potential for economic instability.

For many years, banks have had strict requirements to ensure critical systems (“applications”) could be recovered in the event of a disaster (specifically, a data center failure incurring loss or unavailability of all software, hardware and data). Enterprise-wide people and environment ‘failure’ are often handled within an overall operational risk and business continuity management framework, but technology organizations are accountable for managing technology failure risk – which is to say, an application will be guaranteed to be available within a fixed amount of time in the event of a major infrastructure failure or breach.

In recent times, regulators have been focusing more on business service resilience (see, for example, the UK’s Financial Conduct Authority statement on business service resilience). In the UK, a succession of high-profile loss of digital services (through which most UK customers interact with their bank) in several banks has put increased focus on the area of operational resilience. To date, most such incidents can be attributed to poor technical debt management rather than anything specifically cloud-related – but on cloud, a business-as-usual approach to technical debt will greatly amplify the likelihood and impact of failures.

So, as banks move their infrastructure to leverage public cloud capabilities, the question arises as to how to meet operational resilience demands in various cloud failure scenarios – from failure of individual hosted servers, to loss of a particular provider data center, to loss of a particular infrastructure service in a region, or a complete loss of all service globally from a particular provider.

Related to provision of service is stewardship of related data. Assuming a service cannot be provided without access to and availability of underlying data services, cloud resilience solutions must also address data storage and protection.

Cloud Resilience Approaches

There are multiple ways to address cloud provider availability risk, but all have negative implications for agility and innovation – i.e., the key reasons firms want to use the cloud in the first place.

Mitigating solutions could include the following:

  • Deploy to cloud but maintain on-premise redundant storage, processing and networking facilities to fall-back on (an approach used by Barclays)
  • Implement multi-cloud solutions so that new compute infrastructure can be spun up when needed on an alternate cloud provider, with full cross-cloud data replication always enabled
  • Maintain on-premise deployment for all production and resilience workloads, and only use cloud for non-production environments.

The disadvantage with the above is that all of these strategies require treating cloud as a lowest-common-denominator – i.e., infrastructure-as-a-service. This implies firms need to build or buy capabilities to enable on-premise or multi-cloud ‘platform-as-a-service’ – a significant investment, and likely only justifiable for larger firms. It also effectively excludes a cloud provider’s proprietary services from consideration in solution architectures.

It should be noted as well that multi-cloud data replication strategies can be expensive, as cloud providers charge for data egress, and maintaining a complete view of data in another cloud could consume considerable data egress resources on an on-going basis.

The challenge gets more nuanced when software-as-a-service providers or business-process-as-a-service providers form part of the critical path for a regulated business service. Should a firm therefore have multiple SaaS or BPaaS providers to fall back on in the event of a complete failure of one such provider? In many cases, this is not feasible, but in other cases, the cost of keeping a backup service on ‘standby’ may be acceptably low.

Clearly, there is a point at which there are diminishing returns to actively mitigating these risks – but ignoring these ‘fat tail risks‘ completely runs the risk that a business could disappear overnight with one failure in its ‘digital supply chain’.

Regulators are taking note of this in the UK and in the US, with increased focus on business resilience for key processes, and the role that cloud providers are likely to play in the future in this.

Architecting for Resiliency

Building resilient systems is hard. A key factor in any resilient system is decentralization. The most resilient systems are highly decentralized – there is no ‘brain’ or ‘heart’ that can fail and bring down the whole system. Examples of modern resilient systems are the Internet and BitCoin. Both can survive significant failure, where the impact of failure is localised rather than systemic. (In the not-so-distant future, a truly resilient digitally-based financial infrastructure may have no choice but to be built on decentralized blockchain-like technologies…in other words, there may be diminishing cost/benefit returns for ‘too big to fail’ banks to implement digital resilience using traditional, centralised financial infrastructure architectures, even if cloud-based.)

Cloud providers, and services built on top of them, must by necessity be fundamentally resilient to failure. AWS, for example, ensures that all its regions are technically isolated from its other regions. A regional S3 outage in 2017 highlighted the dependency many businesses and services have on the S3 service. Amazon’s response to the issue implemented change of practices to mitigate the risk of a future failure, but a future regional outage of any cloud service is not only possible, it’s relatively likely (although generally less likely than the failure of an enterprise’s traditional data center, due to multiple redundant availability zones – aka data centers – in each region). So cloud providers typically provide regional isolation, and encourage (through programs like the AWS well-architected review or the Azure Architecture center) customers to design resilience of their systems around this, as well as providing hooks and interfaces to enable cross-region redundancy and fail-over of region-specific services.

Along with architecting for resiliency, organizations need to validate the resiliency of their solutions. Techniques like chaos-engineering ensure teams do not fear change or failure, and are confident in their system’s response to production failures.

Multi-cloud Options

Cloud providers, no matter how much their customers ask for it, are not likely to make it simpler for multi-cloud services to be implemented by their customers. It is in their interest (and is indeed their strategy) to assume that their customers are wholly dependent on them, and as such to take resilience extremely seriously. If cloud providers were to view their competitors as the “disaster recovery” solution for customers, then this would be a significant step backwards for the industry. While it is not likely to happen imminently, there is no room for complacency; perhaps this is where some regulatory oversight of providers may, in the future, be beneficial..

There is, however, room in the market for multi-cloud specialists (and hybrid on-premise/cloud specialists). The biggest player in this space is VMware/Dell, with their recent (re)acquisition of Pivotal, which includes cloud-neutral technologies like (open-sourced) Cloud Foundry and Pivotal Container Services. VMware is going all-in on Kubernetes, as are IBM/Red Hat and others. Kubernetes offers enterprises or 3rd party vendors the opportunity to build their own on-premise ‘cloud’ infrastructure via a standard orchestration API, which in principle should also be able to run on any cloud provider’s IaaS offerings. These steps are all pointing to a future where, irrespective of whether the ‘cloud’ infrastructure provider is public or private, container based software-provisioned infrastructure management is the future, and with it a fundamental shift in thinking about failure.

Cloud is not a Silver Bullet for Resilience

For enterprises, the choices on offer for implementing cloud are many and varied: all ultimately benefit organizations and their stakeholders by providing more flexibility at a lower cost than ever before – as long as systems on the cloud are architected for failure.

System availability (aka resilience) has traditionally been measured as mean-time-between-failure (MTBF), where MTBF is is defined as mean-time-to-failure (MTTF) + mean-time-to-recover (MTTR).

Many on-premise systems are architected to reduce MTTF – i.e., aiming for a large mean-time-to-failures rather than a reduced mean-time-to-recover (MTTR). Distributed systems favor a small MTTR over a large MTTF, and cloud is no exception. Systems explicitly architected for a large MTTF and relatively large MTTR (hours, not seconds) will, in general, find it difficult to migrate to cloud without significant re-engineering – to the extent it is unlikely to be financially feasible. (MTTR for legacy systems is roughly equivalent to ‘recovery time objective (RTO)’ in business continuity planning, and system recovery in such systems is generally assumed to be a highly manual process.)

Cloud-native applications have the advantage that they can be architected for low MTTR from the outset (the AWS well-architected reliability guidelines highlight this), taking into account a relatively small MTTF of foundational cloud components – in particular, virtual machines and docker images.

Going Serverless

It is easy to get lost in all of the infrastructure innovation happening to enable flexible, low cost and resilient systems. But ultimately enterprises don’t care about servers – they care about (business) services. Resilience planning should be built around service availability, and issues like orchestration, redundancy, fail-over, observability, load-balancing, data replication, etc should all be provided via underlying platforms, so applications can focus to the fullest extent possible on business functionality, and not on infrastructure.

Organizations seem to be taking a primarily infrastructure-first approach to cloud to date. This is helpful to build essential cloud engineering and operations competencies, but is unlikely to yield major 10x benefits sought by digital transformation agendas, as the focus is still on infrastructure (specifically, servers, networks and storage). Instead, a more useful longer term approach would be to take a serverless-first stance in solution design – i.e., to highlight where resilient stateful services are needed, and which APIs, events and services need to be created or instantiated that use them. Identify 3rd party or internal infrastructure services which are ‘serverless’ (for example, AWS Cognito, AWS Lambda/Azure Functions/GCP Serverless, Office 365, ServiceNow, …) and plan architectures which use these across all phases of the SDLC. Where required services to achieve the ‘serverless’ target state do not exist, these capabilities should be built or bought. This is essentially the approach advocated by AWS, and should, in my view, form the core of enterprise digital technology roadmaps.

Resilient architectures will consist of multiple ‘serverless’ services, connected via many structured asynchronous events streams or synchronous APIs. Behind each event/API will be a highly automated, resilient platform capable of adjusting to demand and responding to failure through graceful degradation or services. The choice of orchestration platform technology, and its physical location, should largely be neutral to service consumers – in much the same way that users of AWS Lambda don’t know or (for the most part) care about the details of how it works behind the scenes. Standards like the open-sourced AWS ‘serverless application model’ (SAM) enables this for Lambda. Other emerging cloud-neutral serverless standards like Knative and OpenFaaS also exist.

While these technologies are still maturing and are not yet robust enough to handle every demanding scenario, framing architecture in a ‘serverless’ way can be helpful to identify which technologies and providers to use where to provide overall resilience. In particular, using planning tools like Wardley Maps can help map out where and when it makes most business sense to transition from serverful on-premise current-state (custom infrastructure) to serverful (rental infrastructure/ IaaS – private or public) cloud to the serverless target state (commodity PaaS/SaaS) or ultimately to BPaaS.

Summary

Fundamentally, the near term goal is to avoid restricting firms adoption of cloud by unnecessarily limiting their technology choices to a ‘least-common-denominator’ of provider services for infrastructure services that can be delivered cloud natively. For hybrid or multi-cloud solutions, a least-common-denominator approach is currently unavoidable but a healthy, open market in Kubernetes Operators (Custom Resource Descriptors) should, over time, close the on-premise vs cloud-provider-native infrastructure service gap and raise the level of service abstractions available at platform level.

Maximizing a given provider’s managed solutions – and adhering to ‘well-architected’ practices – is likely to provide sufficient cloud-native resilience, certainly well in excess of many existing on-premise practices. For private clouds, bespoke ‘well-architected’ guidance – supported by automated tooling – should be mandatory.

If any regulation is to happen regarding cloud utilization for regulated services, it should be to require firms to follow a provider’s well-architected practices to a very high degree – and that provider’s need to define such practices. In other words, new regulation is not needed to change the core engineering posture of the biggest cloud providers, but rather it should elevate the existing posture as the benchmark others must follow.

For private, hybrid or multi-cloud solutions an industry-wide accepted set of ‘well-architected’ guidance to take the guesswork out of resilience compliance would be welcome, coupled with an overarching framework for maintaining minimum digital service availability in the event of ‘fat-tail failures’ occurring (i.e., extremely rare but high impact failures).

For some regulated services, therefore, regulators will need to know which cloud providers and/or vendor infrastructure orchestration solutions are being used, to better manage concentration risk and ensure the failure of a cloud provider or vendor does not present a systemic risk. But for a given financial institution, the best option is likely to go ‘all-in’ either on a particular cloud provider, or on a cloud-neutral, open-sourced PaaS based on Kubernetes.

Achieving Service Resilience on Cloud-Reliant Infrastructure

AWS CDK – why it’s worth looking at

[tl;dr AWS CDK provides a means for developers to consume compliant, reusable cloud infrastructure components in a way that matches their SDLC, improving developer experience and reducing the risk of silo’ing cloud infrastructure development and operations .]

Some weeks ago, Amazon launched the AWS Cloud Development Kit, or CDK. This article provides my initial (neutral) thoughts on the potential impact and relevance of the CDK for organizations building and deploying solutions on the cloud.

Developer Experience

First, does it work? As the CDK has been in beta for quite a while, when Amazon makes a product generally available, it means it meets a very high bar in terms of quality and stability, and that certainly is the case with the CDK. The examples all worked as expected, although the lambda code pipeline example took some mental gyrations to understand the full implications of what was actually being done – specifically, that the build/deploy pipeline created by the CDK can include running the CDK to generate templates that can be used as an input into that code-deployment pipeline – even if the resulting template is not deployed by the CDK.

All in all, the out-of-the-box experience for the CDK was excellent.

A polyglot framework

Secondly, the CDK is a software development framework. This means it uses ‘traditional’ programming languages (imperative, not declarative), it uses SDLC processes all application developers are familiar with (i.e., build/test/deploy cycles), and it provides many software abstractions that serve to hide (unnecessary) complexity from developers, while enabling developers to build ‘safe’ solutions.

The framework itself has been developed in Typescript, with an interesting technology called ‘jsii‘ used to generate native libraries in other programming languages (specifically, Java, Python, and C#/.NET as well as Javascript).

The polyglot nature of the framework is critical, as cloud infrastructure (as exposed to consumers) must be neutral to any specific programming language. But the most used languages have at least one popular framework for abstracting infrastructure necessary for building distributed applications (such as data stores, message queues, service discovery & routing, configuration, logging/tracing, caching, etc). For example, Java has the Spring framework, C# has .NET Core, Python has Django, Javascript has multiple frameworks based on NodeJS).

So, the question is, should developers now learn to use the CDK or focus instead on language-specific frameworks?

What the CDK is – and what it is not?

To answer this question, we need to be clear on what the CDK is, and what it is not. CDK is a compiler to generate CloudFormation code. If one considers CloudFormation templates to be the ‘assembly language’ for the AWS cloud ‘processor’, then what is important is that CDK generates high-quality CloudFormation templates – period.

To that extent, the CDK only needs to be as efficient as it takes to generate valid CloudFormation templates. It does not execute those templates, so CDK code will never have run-time performance sensitivities (except perhaps, as with traditional compilers, in build toolchains).

For this reason, Typescript seems to have been a pragmatic choice to implement the CDK in. Run-time performance is not the key factor here – rather, it is the creation of flexible, adaptable constructs that avoids the need for developers to write CloudFormation YAML/JSON. Languages that use the CDK libraries should expect the same performance criteria to apply.

Why CDK is necessary

After working through several of the examples, and observing the complexity of the CloudFormation code the CDK generates – and the simplicity of the example code – it is clear that having developers write CloudFormation template is no more sustainable than having developers write assembly language. CloudFormation (as with Azure Resource Manager and GCP Cloud Deployment Manager) is excellent for small, well-defined projects, but rapidly gets complex when expanded to many applications with complex infrastructure inter-dependencies.

In particular, ensuring the security and compliance of templates becomes very complex when templates are hand-crafted. While services like CloudSploit offer to statically scan CloudFormation templates for security breaches, it would be much better to ensure secure CloudFormation code was written in the first place.

Through Constructs, Stacks, and Apps, the CDK allows enterprise engineering teams to provide libraries of secure, compliant infrastructure components to developers that can be safely deployed.

For this reason, as well as the familiarity of CDK constructs to developers, CDK is likely to end up being more popular than having developers hand-craft CloudFormation templates. The complexity of CloudFormation risks enterprises splitting teams into specialists and non-specialists, reverting organizations back to silo’d infrastructure anti-patterns.

However, having specialist engineering teams focused on building, publishing and maintaining high-quality reusable constructs is a good thing, and this is likely what most organizations enterprise engineering teams should focus on (as well as the larger global community of CDK developers).

With respect to language-specific frameworks, perhaps it is only a matter of time before these frameworks generate the cloud-native templates that the software level abstractions can map directly onto. This may mean that the application footprint for such applications could get much smaller in future, as the framework abstractions are increasingly implemented by cloud constructs. Indeed, as observed by Adrian Cockroft, many of the open-source microservice components developed by Netflix ended up being absorbed by AWS, greatly simplifying the Netflix-specific code-base.

If this outlook proves correct, the correct approach for organizations already committed to a microservices framework would be to stick with it, rather than have their business-facing application developers learn CDK.

With respect to Terraform, the most popular cross-cloud provisioning, deployment and configuration tool, its principle benefit is consistent SDLC workflow across cloud providers. Organizations need to decide if a single end-to-end SDLC for application and infrastructure developers on a single cloud (using CDK) provides more benefits than a single infrastructure SDLC across multiple cloud providers, but with a different SDLC for application developers.

CDK & Serverless

For architectures which are fundamentally not ‘serverless’ in nature, the CDK presents a conflict: by allowing infrastructure to be specified and built as part of the developer lifecycle, where does responsibility for managing infrastructure lie?

The reality is, most organizations still exist in a ‘serverful’ world – where infrastructure environments, even if it’s cloud-based, is a ‘pet’ and not ‘cattle’. Environments tend to be created and managed over the long-term, especially where datastores are involved. Stacks tend to be stable across environments, changing only with new releases of software. Separate teams (from developers) are responsible for the ongoing health, security and cost of environments. These teams are likely to be much more comfortable with configuration and scripting than outright coding, using tools like Chef, Puppet, Ansible or AWS OpsWorks. They may prefer developers or architects to request infrastructure components via tools like AWS Service Catalog or ServiceNow, so that infrastructure code is firmly managed away from developers, and the benefits of SDLC-friendly CDK may be less obvious to them.

Generating and maintaining safe, secure and compliant cloud stacks is a vibrant area of growth, and CDK is unlikely to monopolise this – rather, it may spur the growth of 3rd party solutions. 3rd parties that aim to simplify and standardize cloud infrastructure management (such as Pulumi) will have a role to play, particularly for polyglot language and multi-cloud environments, but ‘serverful’ platform and infrastructure teams need to decide what infrastructure building blocks to expose to developers, and how.

With serverless, this dynamic changes significantly, and the CDK can safely become part of the development team’s SDLC. Indeed, a potential goal for enterprise’s moving towards a serverless target state (i.e., applications consisting of composable services with no fixed/bespoke infrastructure) is to use CDK constructs to define those business-level services as infrastructure components. The concept of a platform-as-a-service to integrate software-as-a-service is a concept worth exploring as this space matures, particularly with the advent of services like AWS EventBridge.

In the meantime, behind every ‘serverless’ service lie many servers..teams have many options as to how best to automate this underlying infrastructure, and CDK is another tool in the toolbox to enable this.

Conclusion

AWS CDK is ground-breaking technology that is a big step towards improving the developer experience and capabilities on (AWS) cloud. Other polyglot cloud providers will likely follow suit or risk widening the gap between cloud infrastructure teams and application development teams. Organizations should consider investing in building and publishing CDK construct libraries to be used by application teams – constructs which can be verified to be secure and with sufficient guardrails to allow less-experienced engineers to safely experiment with solutions.

In the meantime, as cloud platforms extend their capabilities, expect language-specific microservices frameworks to get simpler and smaller (or at least more modular in implementation), enabling application developers to fully exploit a given cloud provider’s platform services. Teams relying on these frameworks should understand and drive the roadmap for how these frameworks leverage cloud-native services, and ensure they align with their wider platform cloud/infrastructure automation strategy.

AWS CDK – why it’s worth looking at

The changing role of data lakes

[tl;dr A single data lake, data warehouse or data pipeline to “rule them all” is less useful in hybrid cloud environments, where it can be feasible to query ‘serverless’ cloud-native data sources directly rather than rely on traditional orchestrated batch extracts. Pipeline complexity can be reduced by open extensions to SQL such as the recently announced AWS PartiQL language. Opportunities exist to integrate enterprise human-oriented data governance and meta-data platforms with data pipelines using serverless technologies.]

The need for Data Lakes

A data lake is a centralized repository that allows you to store all your structured and unstructured data at any scale. The data lake concept was created to address a number of issues with traditional data analytics and reporting solutions, specifically:

  • the growing number of applications across an enterprise depending on a given dataset;
  • business and regulatory drivers for governing dataset discovery, quality, creation and/or consumption;
  • the increasing difficulty of IT teams to respond in a timely manner to growing business demand for access to high quality datasets.

The data lake allows data to be made available from its source without making any assumptions about its use. This is particularly critical when the data originates from batch extracts of load-sensitive OLTP databases, most of which are still operating on-premise. Streaming data pipelines, while growing in popularity, are not as common as batch-driven pipelines – although this should change over time as more digital platform architectures become more event-driven in nature.

Data lakes are a key component in data pipelines, a construct (or set of constructs) that┬áprovides consolidation of data from multiple sources and makes it available for use. A data pipeline can be orchestrated (via a scheduler) or choreographed (responding to events) – the more jobs a pipeline has to do, the more complex the orchestration or choreography, which has implications for supportability. So reducing the number of jobs a pipeline has to support is key to managing data pipeline complexity.

The Components of a Data Lake

A data lake consists of a few key components:

FeatureDescriptionVirtualTraditional
A storage repositoryDurable, resilient storage of data objects.NoYes
An ingestion mechanismA means to upload content to the repository (no transformation)NoYes
A tagging & metadata mechanismA means to associate metadata with data objects, including user-defined tags.YesYes
A metadata search mechanismA means to search objects in the data lake based on metadata and tags (not content)YesYes
A query engineA means to search the content of objects in the data lakeYesPartially
An access control mechanismA means to ensure that users can only access datasets and parts of data sets that they are entitled to see, and to audit all activity.YesYes

In effect, data lakes have become a kind of data warehouse – the main significant difference being that input sources into data lakes tend to be familiar files – CSVs, Avro, JSON, etc. from multiple sources rather than highly optimized domain-specific schemas – i.e., no assumptions are made about how (or why) the data in the data lake will be consumed. Data lakes also do not concern themselves with scheduling or orchestration.

Datawarehouses, datawarehouses everywhere…

For mature data use cases (i.e., situations where relatively stable, well-known data requirements exist), and where consistent high performance is material to meeting customer needs, data warehouses are still the best solution. A data warehouse stores and manages all of its data locally, and only relies on the data lake as an initial ingestion point.

A data warehouse will transform datasets to the form needed for the specific use cases it supports, and will optimize performance for the consumption of those datasets. Modern data warehouses will use ML/AI techniques to optimize performance rather than relying on human database specialists. But, as this approach is compute intensive, such solutions are more amenable to cloud environments than on-premise environments. Snowflake is an example of this model. As more traditional data warehouses (e.g., Oracle Exadata) move to the cloud, we can expect these to also get ‘smarter’ – however, data gravity will mean such solutions will need to be fundamentally multi-cloud compatible.

For on-premise data warehouses, the tendency is for business lines or functions to create ‘one data warehouse to rule them all’ – mainly because of the traditionally significant storage and compute infrastructure and resources necessary to support data warehouses. Consequently considerable effort is spent on defining and maintaining high performance, appropriately normalized, enterprise data models that can be used in as many enterprise use cases as possible.

In a hybrid/cloud world, multiple data warehouses become more feasible – and in fact, will be inevitable in larger organizations. As more enterprise data becomes available in these dynamically scalable, cloud-based (or HDFS/Hadoop based) data warehouses (such as AWS EMR, AWS Redshift, Snowflake, Google Big Query, Azure SQL Data Warehouse), ‘virtual data warehouses’ avoid the need to move data from its source for query handling, allowing data storage and egress costs to be kept to a minimum, especially if assisted by machine-learning techniques.

Virtual Data Warehouses

Virtual Data warehouse technologies have been around for a while, allowing users to manage and query multiple data sources through a common logical access point. For on-premise solutions, virtual data warehouses have limited use cases, as the cost/effort of scaling out in-house solutions can be prohibitive and not particularly agile in nature, precluding experimental use cases.

On hybrid or cloud environments, virtual data warehouses can leverage the scalability of cloud-native data warehouses, driving queries to the relevant engine for execution, and then leveraging its own scalable infrastructure for executing join queries.

Technologies like Dremio reflect the state of the art in cloud-based data warehouses, which push down queries to the source system where possible, but can process them in-memory directly from a data lake or other source if not.

However, there is one thing that all data warehouses have in common: they leverage SQL and (implicitly) a relational view of the data. Standard ANSI SQL queries are generally supported by all data warehouses, but may mean that some data cannot be queried if it is not in tabular form amenable to SQL processing.

Extending SQL with PartiQL

Enter PartiQL, an open-source project sponsored by Amazon to drive extensions to standard SQL that can cope with non-relational data types, including structured, unstructured, nested, and schemaless (NoSQL, Document).

Historically, all data ingested into a data lake had to be transformed into a format that could be queried by SQL-like commands or processed by typical data warehouse bulk-upload tools. This adds complexity to data pipelines (i.e., more jobs), and may also force premature schema design (i.e., forcing the design of an optimal schema before all critical use cases are fully understood).

PartiQL potentially allows tools such as Snowflake, Dremio (as well as the tools AWS uses internally) to query data using SQL-like syntax, but to also include non-relational data in those queries so they can avoid those separate transformation steps, aiding pipeline complexity reduction.

PartiQL claims to be fully ANSI-compliant, but extended in specific ways to support alternate data formats. While not an official ISO/ANSI standard, it may have the ability to become a de-facto standard – especially as the language has already been used in anger with success within AWS. This will provide a skill path for relational data warehouse experts to become proficient in leveraging modern data pipelines without committing to one specific vendor’s technology.

Technologies like PartiQL will make it much easier to include event-sourced streams into a data pipeline, as events are defined as nested or other non-relational structures. As more data pipelines become event driven rather than batch-driven, having a standard like PartiQL will be key. (It will be interesting to see if Confluent’s KSQL and PartiQL will converge to a single event-stream query standard.)

As PartiQL has only just been released, it’s too soon to tell how the big data ecosystem or ISO/ANSI will respond. Expect more on this topic in the future. For now, virtual data warehouses must rely on their proprietary SQL extensions.

Non-SQL Data Processing

Considerable investment is being made by third party vendors on advanced technology focused on making distributed, scalable processing of SQL (or SQL-like) queries fast and reliable with little or no human tuning required. As such, it is wise to pick a vendor demonstrating a clear strategy in this space, and continuing to invest in SQL as the lingua-franca of transformation logic.

However, for use cases for which SQL is not appropriate, distributed computing platforms like Spark are still needed. The expectation here is that such platforms will ingest data from a data lake, and output results into a data lake. In some cases, the distributed computing platform offers its own storage (e.g., HDFS), but increasingly it is more appropriate to question whether data needs to reside permanently in a HDFS cluster rather than in a data lake. For example, Amazon’s EMR service allows Hadoop clusters to be created ephemerally, and to consume their initial dataset from AWS S3 repositories or other data sources,

Enforcing Enterprise Data Collaboration and Governance

Note that all data warehouse solutions (virtual or not) must support some form of meta-data tagging and management used by their SQL query engines – otherwise they cannot act as a virtual database source (generally an ODBC end-point that applications can connect directly to). This tagging can be automated if sources included meta-data (e.g., field headers, Avro schema definitions, etc) , but can be enhanced by human tagging, which is increasingly augmented by machine-learning to help identify, for example, where data may be sensitive, etc.

But data governance needs extend beyond the needs of the virtual data warehouse query engines, and this is where there are still gaps to be filled in the current enterprise data management tools.

Tools from vendors like Alation, Waterline, Informatica, Collibra etc were created to augment people’s ability to properly tag content in the data-lake with meaningful information to make it discoverable and governable. Consistent tagging in principle allows tag-based governance rules to be defined to automatically enforce data governance policies in data consumers. This data, coupled with schema information which can be derived directly from data-sources, is all the information needed to allow users (or developers) to source the data they need in a secure, compliant way.

But meta-data for data governance has humans as the primary user (e.g. CDOs, business/data analysts, process owners, etc) – or, as Alation describes it – meta-data for human collaboration.

Currently, there is no accepted standards for ensuring the consistency of ‘meta-data for human collaboration’ with ‘meta-data for query execution’.

Ideally, the human-oriented tools would generate standard events that tools in the data pipeline could pick up and act on (via, for example, something like AWS EventBridge), thereby avoiding the need for data governance personnel to oversee multiple data pipelines directly…

Summary

With the advent of cloud-based managed compute and data storage services, a multi-data warehouse and pipeline strategy is viable and may even be desirable, potentially involving multiple data lakes.

Solutions like PartiQL have the potential to eliminate many transformation job phases and greatly simplify data pipeline complexity in a standardized way, leveraging existing SQL skills rather than requiring new skills.

To ensure consistent governance across multiple data pipelines, a serverless event-based approach to connecting human data governance solutions with cloud-native data pipeline solutions may be the way forward – for example, using AWS EventBridge to action events originating from SaaS-based data governance services with data pipelines.

The changing role of data lakes

The cloudy future of data management & governance

[tl;dr The cloud enables novel ways of handling an expected explosion in data store types and instances, allowing stakeholders to know exactly what data is where at all times without human process dependencies.]

Data management & governance is a big and growing concerns for more and more organizations of all sizes. Effective data management is critical for compliance, resilience, and innovation.

Data governance is necessary to know what data you have, when you got it, where it came from, where it is being used, and whether it is of good quality or not.

While the field is relatively mature, the rise of cloud-based services and service-enabled infrastructure will, I believe, fundamentally change the nature of how data is managed in the future and enable greater agility if leveraged effectively.

Data Management Meta-Data

Data and application architects are concerned about ensuring that applications use the most appropriate data storage solution for the problem being solved. To better manage cost and complexity, firms tend to converge on a handful of data management standards (such as Oracle or SQL Server for databases; NFS or NTFS for filesystems; Netezza, Terradata for data warehousing, Hadoop/HDFS for data processing, etc). Expertise is concentrated around central teams that manage provisioning, deployments, and operations for each platform. This introduces dependencies that project teams must plan around. This also requires forward planning and long-term commitment – so not particularly agile.

Keeping up with data storage technology is a challenge – technologies like key/value stores, graph databases, columnar databases, object stores, and document databases exist as these represent varying datasets in a more natural way for applications to consume, reducing or eliminating the ‘impedance mismatch‘ between how applications view state and how that state is stored.

In particular, may datastore technologies are used to scaling up rather than out; i.e., the only way to make them perform faster is to add more CPU/memory, or faster IO hardware. While this keeps applications simpler, it require significant forward planning and longer-term commitments to scale up, and is out of the control of application development teams. Cloud-based services can typically handle scale-out transparently, although applications may need to be aware of the data dimensions across which scale out happens (e.g., sharding by primary key, etc).

Fulfilling provisioning requests for a new datastore on-premise is mostly ticket driven, but fulfillment is still mostly by humans and not by software within enterprises – which means an “infrastructure-as-code” approach is not feasible.

Data Store Manageability vs Application Complexity

Most firms decide that it is better to simplify the data landscape such that fewer datastore solutions are available, but to resource those solutions so that they are properly supported to handle business critical production workloads with maximum efficiency.

The trade-off is in the applications themselves, where the data storage solutions available end up driving the application architecture, rather than the application architecture (i.e., requirements) dictating the most appropriate data store solution, which would result in the lowest impedance mismatch.

A typical example of an impedance mismatch are object-oriented applications (written in, say C++ or Java) which use relational databases. Here, object/relational mapping technologies such as Hibernate or Gigaspaces are used to map the application view of the data (which likes to view data as in-memory objects) to the relational view. These middle layers, while useful for naturally relational data, can be overly expensive to maintain and operate if what your application really needs is a more appropriate type of datastore (e.g., graph).

This mismatch gets exacerbated in a microservices environment where each microservice is responsible for its own persistence, and individual microservices are written in the language most appropriate for the problem domain. Typical imperative, object-oriented languages implementing transactional systems will lean heavily towards relational databases and ORMs, whereas applications dealing with multi-media, graphs, very-large objects, or simple key/value pairs will not benefit from this architecture.

The rise of event-driven architectures (in particular, transactional ‘sagas‘, and ‘aggregates‘ from DDD) will also tend to move architectures away from ‘kitchen-sink’ business object definitions maintained in a single code-base into multiple discrete but overlapping schemas maintained by different code-bases, and triggered by common or related events. This will ultimately lead to an increase in the number of independently managed datastores in an organisation, all of which need management and governance across multiple environments.

For on-premise solutions, the pressure to keep the number of datastore options down, while dealing with an explosion in instances, is going to limit application data architecture choices, increase application complexity (to cope with datastore impedance mismatch), and reduce the benefits from migrating to a microservices architecture (shared datastores favor a monolithic architecture).

Cloud Changes Everything

So how does cloud fundamentally change how we deal with data management and governance? The most obvious benefit cloud brings is around the variety of data storage services available, covering all the typical use cases applications need. Capacity and provisioning is no longer an operational concern, as it is handled by the cloud provider. So data store resource requirements can now be formulated in code (e.g., in CloudFormation, Terraform, etc).

This, in principle, allows applications (microservices) to choose the most appropriate storage solution for their problem domain, and to minimize the need for long-term forward planning.

Using code to specify and provision database services also has another advantage: cloud service providers typically offer the means to tag all instantiated services with your own meta-data. So you can define and implement your own data management tagging standards, and enforce these using tools provided by the cloud provider. These can be particularly useful when integrating with established data discovery tools, which depend on reliable meta-data sources. For example, tags can be defined based on a data ontology defined by the chief data office (see my previous article on CDO).

These mechanisms can be highly automated via service catalogs (such as AWS Service Catalog or ServiceNow), which allow compliant stacks to be provisioned without requiring developers to directly access the cloud providers APIs.

Let a thousand flowers bloom

The obvious downside to letting teams select their storage needs is the likely explosion of data stores – even if they are selected from a managed service catalog. But the expectation is that each distinct store would be relatively simple – at least compared to relational stores which support many application use cases and queries in a single database.

In on-premise situations, data integration is also a real challenge – usually addressed by a myriad of ad-hoc jobs and processes whose purpose is to extract data from one system and send it to another (i.e., ETL). Usually no meta-data exists around these processes, except that afforded by proprietary ETL systems.

In best case integration scenarios, ‘glue’ data flows are implemented in enterprise service buses that generally will have some form of governance attached – but which usually has the undesirable side-effect of introducing yet another dependency for development teams which needs planning and resourcing. Ideally, teams want to be able to use ‘dumb’ pipes for messaging, and be able to self-serve their message governance, such that enterprise data governance tools can still know what data is being published/consumed, and by whom.

Cloud provides two main game-changing capabilities to manage data complexity management at scale. Specifically:

  • All resources that manage data can be tagged with appropriate meta-data – without needing to, for example, examine tables or know anything about the specifics about the data service. This can also extend to messaging services.
  • Serverless functions (e.g., AWS Lambda, Azure Functions, etc) can be used to implement ‘glue’ logic, and can themselves be tagged and managed in an automated way. Serverless functions can also be used to do more intelligent updates of data management meta-data – for example, update a specific repository when a particular service is instantiated, etc. Serverless functions can be viewed as on-demand microservices which may have their own data stores – usually provided via a managed service.

Data, Data Everywhere

By adopting a cloud-enabled microservice architecture, using datastore services provisioned by code, applying event driven architecture, leveraging serverless functions, and engaging with the chief data officer for meta-data standards, it will be possible to have an unprecedented up-to-date view of what data exists in an organization and where. It may even address static views of data in motion (through tagging queue and notification topic resources). The data would be maintained via policies and rules implemented in service catalog templates and lambda functions triggered automatically by cloud configuration changes, so it would always be current and correct.

The CDO, as well as data and enterprise architects, would be the chief consumer of this metadata – either directly or as inputs into other applications, such as data governance tools, etc.

Conclusion

The ultimate goal is to avoid data management and governance processes which rely on reactive human (IT) input to maintain high-quality data management metadata. Reliable metadata can give rise to a whole new range of capabilities for stakeholders across the enterprise, and finally take IT out of the loop for business-as-usual data management queries, freeing up valuable resources for building even more data-driven applications.

The cloudy future of data management & governance

The future of modularity is..serverless

[tl;dr As platform solutions evolve and improve, the pressure for firms to reduce costs, increase agility and be resilient to failure will drive teams to adopt modern infrastructure platform solutions, and in the process decompose and simplify monoliths, adopt microservices and ultimately pave the way to building naturally modular systems on serverless platforms.]

“Modularity” – the (de)composition of complex systems into independently composable or replaceable components without sacrificing performance, security or usability – is an architectural holy grail.

Businesses may be modular (commonly expressed through capability maps), and IT systems can be modular. IT modularity can also be described as SOA (Service Oriented Architecture), although because of many aborted attempts at (commercializing) SOA in the past, the term is no longer in fashion. Ideally, the relationship between business ‘modules’ and IT application modules should be fully aligned (assuming the business itself has a coherent underlying business architecture).

Microservices are the latest manifestation of SOA, but this is born from a fundamentally different way of thinking about how applications are developed, tested, deployed and operated – without the need for proprietary vendor software.

Serverless takes takes the microservices concept one step further, by removing the need for developers (or, indeed, operators) to worry about looking after individual servers – whether virtual or physical.

A brief history of microservices

Commercial manifestations of microservices have been around for quite a while – for example Spring Boot, or OSGi for Java – but these have very commercial roots, and implement a framework tied to a particular language. Firms may successfully implement these technologies, but they will need to have already gone through much of the microservices stone soup journey. It is not possible to ‘buy’ a microservices culture from a technology vendor.

Because microservices are intended to be independently testable and deployable components, a microservices architecture inherently rejects the notion of a common framework for implementing/supporting the microservices natures of an application. This therefore puts the onus on the infrastructure platform to provide all the capabilities needed to build and run microservices.

So, capabilities like naming, discovery, orchestration, encryption, load balancing, retries, tracing, logging, monitoring, etc which used to be handled by language-specific frameworks are now increasingly the province of the ‘platform’. This greatly reduces the need for complex, hard-to-learn frameworks, but places a lot of responsibility on the platform, which must handle these requirements in a language-neutral way.

Currently, the most popular ‘platforms’ are the major cloud providers (Azure, Google, AWS, Digital Ocean, etc), IaaS vendors (e.g., VMWare, HPE), core platform building blocks such as Kubernetes, and platform solutions such as Pivotal Cloud Foundry,  Open Shift and Mesophere. (IBM’s BlueMix/Cloud is likely to be superseded by Red Hat’s Open Shift.)

The latter solutions previously had their own underlying platform solutions (e.g., OSGi for BlueMix, Bosh for PKS), but most platform vendors have now shifted to use Kubernetes under the hood. These solutions are intended to work in multiple cloud environments or on-premise, and therefore in principle allow developers to avoid caring about whether their applications are deployed on-premise or on-cloud in an IaaS-neutral way.

Decomposing Monolithic Architectures

With the capabilities these platforms offer, developers will be incentivized to decompose their applications into logical, distributed functional components, because the marginal additional cost of maintaining/monitoring each new process is relatively low (albeit definitely not zero). This approach is naturally amenable to supporting event driven architectures, as well as more conventional RESTful and RPC architectures (such as gRPC), as running processes can be mapped naturally to APIs, services and messages.

But not all processes need to be running constantly – and indeed, many processes are ‘out-of-band’ processes, which serve as ‘glue’ to connect events that happen in one system to another system: if events are relatively infrequent (e.g., less than one every few seconds), then no resources need to be used in-between events. So provisioning long-running docker containers etc may be overkill for many of these processes – especially if the ‘state’ required by those processes can be made available in a low-latency, highly available long-running infrastructure service such as a high-performance database or cache.

Functions on Demand

Enter ‘serverless’, which aims to specify the resources required to execute a single piece of code (basically a functional monolith) on-demand in a single package – roughly the equivalent of, for example, a declarative service in OSGi. The runtime in which the code runs is not the concern of the developer in a serverless architecture. There are no VMs, containers or side-cars – only functions communicating via APIs and events.

Currently, the serverless offerings by the major cloud providers are really only intended for ‘significant’ functions which justify the separate allocation of compute, storage and network resources needed to run them. A popular use case are ‘transformational’ functions which convert binary data from one form to another – e.g., create a thumbnail image from a full image – which may temporarily require a lot of CPU or memory. In contrast, an OSGi Declarative Service, for example, could be instantiated by the runtime inside the same process/memory space as the calling service – a handy technique for validating a modular architecture without worrying about the increased failure modes of a distributed system, while allowing the system to be readily scaled out in the future.

Modular Architectures vs Distributed Architectures

Serverless functions can be viewed as ‘modules’ by another name – albeit modules that happen to require separately allocated memory, compute and storage to the calling component. While this is a natural fit for browser-based applications, it is not a great fit for monolithic applications that would benefit from modular architectures, but not necessarily benefit from distributed architectures. For legacy applications, the key architectural question is whether it is necessary or appropriate to modularize the application prior to distributing the application or migrating it to an orchestration platform such as Kubernetes, AWS ECS, etc.

As things currently stand, the most appropriate (lowest risk) migration route for complex monolithic applications is likely to be a migration of some form to one of the orchestrated platforms identified above. By allowing the platform to take care of ‘non-functional’ features (such as naming, resilience, etc), perhaps the monolith can be simplified. Over time, the monolith can then be decomposed into modular ‘microservices’ aligned by APIs or events, and perhaps eventually some functionality could decompose into true serverless functions.

Serverless and Process Ownership

Concurrently with decomposing the monolith, a (significant) subset of features – mainly those not built directly using the application code-base, or which straddle two applications – may be meaningfully moved to serverless solutions without depending on the functional decomposition of the monolith.

It’s interesting to note that such an architectural move may allow process owners to own these serverless functions, rather than relying on application owners, where often, in large enterprises, it isn’t even clear which application owner should own a piece of ‘glue’ code, or be accountable when such code breaks due to a change in a dependent system.

In particular, existing ‘glue’ code which relies on centralized enterprise service buses or equivalent would benefit massively from being migrated to a serverless architecture. This not only empowers teams that look after the processes the glue code supports, but also ensures optimal infrastructure resource allocation, as ESBs can often be heavy consumers of infrastructure resources. (Note that a centralized messaging system may still be needed, but this would be a ‘dumb pipe’, and should itself be offered as a service.)

Serverless First Architecture

Ultimately, nirvana for most application developers and businesses, is a ‘serverless-first’ architecture, where delivery velocity is only limited by the capabilities of the development team, and solutions scale both in function and in usage seamlessly without significant re-engineering. It is fair to say that serverless is a long way from achieving this nirvana (technologies like ‘AIOps‘ have a long way to go), and most teams still have to shift from monolithic to modular and distributed thinking, while still knowing when a monolith is still the most appropriate solution for a given problem.

As platform solutions improve and mature, however, and the pressure mounts on businesses whose value proposition is not in the platform engineering space to reduce costs, increase agility and be increasingly resilient to failures of all kinds, the path from monolith to orchestrated microservices to serverless (and perhaps ‘low-code’) applications seems inevitable.

The future of modularity is..serverless

Message Evolution in High Performance Messaging Environments

tl;dr

Moving to an event-driven architecture in a high-performance environment has specific needs that do not yet have widely standardized solutions, and as such require a high degree of focus on both software engineering and business architecture.

Context

Event- or message-driven applications exist in at least two contexts – an application-specific context, and a domain or enterprise context. For high-performance applications, latency is typically more sensitive within the application context, and less sensitive in the domain/enterprise context.

For the purposes of this article, the application-specific context is assumed to related to components that are typically deployed together when a new feature is released – i.e., there is high coupling and high cohesion between components. All application context components are generally tested and deployed together as a unit.

Application Context

The high-coupling and high cohesion of the application context is usually a compromise to achieve the low latency performance requirements, as normal microservice architecture best practice states that services are independently deployable and hence loosely coupled. This impacts the overall agility of the architecture, but fundamentally, with automated configuration, testing and deployment, it should be manageable by a ‘pizza-sized’ team without losing the integrity of the platform or the performance of the network communications.

A given application can reasonably require a single/common version of a highly optimized serialization library, and related message schema definitions, to be used by all components at any given time, as enforcing such deployments can be guaranteed through deployment/configuration automation processes.

In general, these deployments do not require meta-data to be included in messages, as the meta-data will be explicit in the application code. High-performance binary-encoded protocols like Protocol Buffers and Avro can handle a certain amount of schema evolution, but in general for highly coupled applications, all components should use the same version of serialized objects and be deployed simultaneously.

Application messaging contexts can get complicated, and may include both low-latency and normal-latency scenarios, with different degrees of cohesion and coupling. The key point is that messaging and deployment decisions lie within one ‘pizza-size’ team and are not subject to wider enterprise governance. Eventually, when a single pizza-size team becomes insufficient for the application complexity, technical debt relating to these decisions will be revealed when teams split, and will eventually need to be addressed, as governance, agility and latency needs may all be in conflict.

Domain/Enterprise Context

The domain context may include many pizza-sized application teams. For maximum agility, these disparate applications should have relatively low (functional) cohesion and low coupling, but each may consist of a number of high cohesion microservices – the level of coupling depending on the extent to which message codecs are optimised for network performance or agility.

To maximize decoupling and ensure maximum independence of testing, deployment, configuration and operation, the codecs used in the enterprise context should be as flexible as possible.

A consequence of this approach is that, to ensure decoupling, there will in many cases need to be some translation between events in the application context to events in the domain/enterprise context. Typically this can be done by a separate event-driven component which can do the necessary translations, consuming from one messaging channel and publishing to another. The additional overhead of this should be weighed against the agility cost of trying to maintain application schema consistency at enterprise scale, which can initially be considerable as teams begin to adopt event-driven architecture. (The legacy of Enterprise Service Buses, and the messaging bottlenecks they often cause, show how extreme this cost can be.)

Message Standards & Governance

In general, there is a trade-off between agility and data architecture compliance at the domain/enterprise level. In order to avoid the insertion of another technology team between producers and consumers, it is generally best to follow the microservice best practice of ‘dumb pipes and smart end-points’ – i.e., any compliance with standards is not enforced by the messaging infrastructure but instead at the application (or ‘project’) level.

It is feasible to develop run-time tools to assess the data architecture compliance of messages over the bus – in many circumstances this may offer the best balance between compliance and agility, especially if they run in lower environments prior to production deployment.

Enterprise Message Characteristics

Messages in the enterprise domain have some specific characteristics:

  • No assumptions should be made about the consuming applications, in terms of the languages, libraries or frameworks used, except to the extent that the serialization mechanism is supported.
  • Messages should be readily readable by authorized human readers irrespective of schema versions (in particular, operations & support staff).
  • Messages (or parts of messages) should not be readable by unauthorized readers (human or computer)

When it comes to choosing serialization technology, it is all about compromises. There is no silver bullet. There are trade-offs on performance vs flexibility vs readability vs precision etc. An excellent read on this topic is Martin Kleppman’s book ‘Design Data Intensive Applications‘.

Schema Evolution using a Message Bus

Schema evolution is one of the biggest challenges any technology team has to deal with. Databases, events/messages and RESTful APIs all require schemas to be managed.

Microservices aims to minimize the complexity of managing database schema evolution through ensuring any applications that depend on a particular dataset access that data only through a microservice. In effect, the database has only one reader/writer, and so schema evolution can be tied to deployments of that one application – much easier to manage.

However, this pushes inter-application schema evolution to the messaging or API layer. For breaking schema changes (i.e., a new version of a schema is incompatible with a prior version), two principle approaches can be considered:

  • A single ‘channel’ (topic or URI) handles all schema versions, and each consumer must be able to handle all schema versions received over that channel
  • Each schema version has its own message channel (topic or URL), and new consumers are created specifically to consume messages from that channel.

Note that the approaches above only discuss the technical aspects of decoding incompatible message versions: it does not address semantic changes, which can only be resolved at the application level.

Single Channel

In this case, the consumer must have multiple versions of the deserializer ‘built-in’, so it can interpret the version header and invoke the correct deserializer. For many languages, such as Java, this is difficult, as it requires supporting multiple versions of the same classes in the same process.

It is, in principle, doable using OSGi, but otherwise, the consuming application may be forced to delegate incoming messages to other processes for deserializing, which could be expensive. Alternatively, the IDL parser for the serializer could generate unique encoding/decoding classes for each version of a schema so they could reside in the same process. However, message meta-data indicating the correct schema version would need to be very reliable to ensure this works well.

Multiple Channel

In this case, each new (breaking) schema definition will have its own channel (topic/URI), such that new processes specifically built to consume those schema messages can be deployed that subscribe to that channel.

This avoids the need for delegating deserialization, and may be easier to debug when issues occur. However, it can add additional complexity to channel/topic namespaces, and mechanisms may need to be in place to ensure all expected consumers were running and that there are no accidental ‘orphan’ messages being published (i.e., messages for which there is no consumer active).

Architectural Implications of EDA

Fundamentally, to handle an enterprise-wide event-driven architecture, organizations must be fully committed to implementing a microservices architecture. At the simplest level, this means that the cost and overhead of deploying operating new application components is very low. This means that orchestration, configuration, logging, monitoring and control mechanisms are all standardized across all deployed components, so that there is no operational resistance to deploying separate processes for each message type and/or channel as needed, or to deploy various adaptors/gateways to cope with potentially multiple incompatible enterprise consumers.

Implementing any form of an EDA architecture without addressing the above will likely not substantially improve business agility or lead-time reduction. Instead, it could lead to increased co-dependencies across components, reducing overall system availability and stability, and requiring coordinated integration testing and deployments on a periodic basis (every 2-3 months, for example).

Conclusion

The points above are based on observing the architectural evolution of systems I have directly been involved with, the challenges teams faced by multiple teams moving to an event driven architecture, and the lessons learned from this process.

While a number of issues are squarely in the technical domain, some of the hardest decisions relate to what should be considered in the ‘application’ domain vs what belongs in the ‘enterprise’ domain. Usually, there will be strong business drivers behind merging applications previously in separate domains into a single domain – and this will have implications on team size, message standards, etc. Fundamentally, however, IT should not attempt to draw technical message or data boundaries around applications that are not directly aligned with business architecture goals.

In essence, if business architects and/or product owners are not directly involved in dictating messaging standards (including semantic definitions of fields) across applications, then Conways Law applies: messaging standards remain local to the teams that use them, with many message flows existing as bi-lateral agreements between applications.

This naturally gives rise to a ‘spaghetti’ architecture, but if this reflects how business processes are actually aligned and communicate, and the business is happy with this, then all IT can do is manage it, not eliminate it.

Message Evolution in High Performance Messaging Environments