Cargo Cult Agile

It is instructive to consider human tendency towards magical thinking that finds a wonderfully didactic expression in Cargo Cults.

Magical thinking fundamentally expresses faith that performance of certain rituals alters the status of the actor and can confer benefits that are not correlated with the substance of the ritual itself.
It is not difficult to find many examples where application of Agile is an exercise in ritualistically following the form rather than focusing on substance.

The original intent of Agile was to redress the balance of power between managers/stakeholders and engineers, affording more power to the latter, and achieving better results & productivity through their empowerment.

But it is not uncommon today to find instances where Agile became co-opted as just another mechanism of control, without altering the power equation, and rests on what amounts to a magical belief that by performing formal ceremonies and rituals of Agile, the benefits will accrue even if such performance is entirely divorced from its underlying rationale.

The very same powers that demanded of engineers compliance with rigors of waterfall-style development, simply switched the rulebook and now demand compliance with the rigors of Agile. Such hollowed-out Agile practice remarkably resembles the cargo cults I referred to at the start.

On Well-Designed Event Models

I propose that events composing a well-designed event model should be of a form:

Where:

• E_Ai^x→y is the event communicating transition of aggregate/entity A_i  from business state S^x to business state S^y

• A_i is the i-th instance of aggregate or entity A consisting of attributes {α, β... ω} 

• State transition (S^x → S^y) is accompanied by changes in a subset of attributes of A_i {α, β... μ} to new values {α', β'... μ'}, these new values (if required, with previous values {α, β... μ}) are included in the data payload of the event E_Ai^x→y{α', β'... μ'}

This approach is in contrast with a trivial event model where events are simply attribute-level changes, e.g. instead of emitting E_Ai^x→y{α', β'... μ'}, emit the following series:

Ergo, developing a business state-centric event model requires starting with business state transition diagram for entity/aggregate A:

The benefit of business state-centric event model is to make explicit the state transitions (which typically involve changes to more than one attribute), instead of those having to be inferred/reconstructed by the consumer of attribute level events E^α_Ai {α'} = A_i {α'}.

The Why of Explainable AI

AI is all the rage today. And ARPA is investing into explainable AI - why?

I believe the key for us, humans, is to be able to have recourse to another human who, we viscerally feel, may understand our situation and have empathy, based on shared experience and understanding of the world and how we - mortal, fallible humans - interact with it.

So no matter how sophisticated the AI, will we ever trust it to understand us enough that we will not wish to have recourse to the judgement of another human with power to overrule the "Computer says NO"?

But, how would a human "appellate judge" be able to evaluate AI's decisions if they're a mystery concealed within the deep neural network, acting on more data than a human can readily deal with?

Hence the need for the explainable AI.

I will leave it as an exercise to the reader to infer what types of use cases ARPA might be interested in, where robotic judgement might need to be subject to human appeal.

Fast Data > Big Data

What is the next step after Big Data? Creating a fast lane for generation and utilization of insights from business data, increasing customer relevance and reducing irrelevant noise. Some call it Fast Data. What are the ingredients?

• Ability to source data near-real-time from transactional/OLTP systems either by tailing logs or consuming events 
• Augmenting traditional ETL data pipes with stream processing 
• Enabling greater model development agility and optimization via machine learning 
• Ability to feed resulting insights (e.g. segmentation changes) back to OLTP systems supporting customer interactions

Rationality of Rules vs. Rationality of Ends

"The problem of bureaucracy is that it values the rationality of rules over the rationality of ends."

– Matthew Taylor (with echoes of Max Weber) 

This can serve as a nice litmus test for when technology governance is in danger of devolving into a bureaucracy. This is why it’s so important to focus on tangible value in areas such as governance and technology architecture.

CQRS – The Sweet Spot

Experience teaches all of us that in technology there’s no panacea. Every tool has its uses and, conversely, can be misused. To really understand a piece of technology one needs to understand its design intent, what problem its creators were trying to solve, and what trade-offs they selected.

It is a truism that every engineering decision involves trade-offs – selecting certain strengths at the price of accepting certain limitations (not just in software/systems engineering, but in all engineering disciplines) at a given level of investment. Thus, every instrument has a sweet spot for its application.

CQRS is an interesting example to consider. It builds on the strengths of a more vanilla version of EDA (event-driven architecture) – uncoupling domains from each other – and adds asymmetric read/write models and event sourcing.

By enabling updates at highly granular, attribute level, coupled with arbitrarily denormalized aggregate reads, CQRS creates opportunities for independently scaling the read and write/update stores – much more so than adding read replicas to traditional, symmetric-model domain implementations.

These strengths come at a certain price of additional complexity – managing highly granular even models, ensuring interoperability with non-CQRS-based domains, dealing with compacting of logs in event sourcing, etc.

It would seem that domains that would benefit most from this approach are those where multiple actors are attempting simultaneous updates to varying attributes of the same aggregate – i.e. domains demanding high-concurrency writes.

Big Data and Magical Thinking

Every business is drowning in data. Every business believes they should make better use of that data. Seemingly a long time ago, Big Data came onto the scene as The Answer. It became a buzzword and a cottage industry. In some places, it simply became a synonym for Hadoop.

The challenge is that simply having more data, or combining all of the business’ data into a common pool or ‘lake’ isn’t by itself going to unlock insights, as if by magic.

Rigor is required in managing the data sources and the meaning of various data elements, and equally rigor is required in applying proper mathematical techniques in analysis of the data and avoidance of misleading conclusions.

Things like curse of dimensionality (applicable to sampling and anomaly detection among other things), misuse of p-values, and implicit assumptions about the shape of probability distribution come to mind as some of the most common omissions.

Clarity As Managerial Value Added

It is normally implicitly understood that people at more senior managerial or executive levels have to operate at (and be comfortable with) progressively higher levels of uncertainty. Often, one has to make choices being cognizant of a superposition of multiple future states that is yet to be resolved into one or the other (not unlike a quantum one).

It is also true that key role of management is to enable people to be as productive as possible. Ergo, the manager’s job has to include a meaningful reduction in uncertainty that’s being passed through to the subordinates.

Barclaycard, London

After 5 years at PayPal, most recently serving as Chief Architect for PayPal Credit, I decided to move to London, having accepted the role of CTO with Barclaycard.

You might wonder why a techie with startup background would leave Silicon Valley to work at Britain's oldest bank.

The reason is simple: financial industry as a whole and consumer financial services in particular are undergoing a transformation, some might say a seismic shift, which is significantly technology-driven.

A wave of change disrupting the status quo presents a great opportunity for innovation.

I'm tremendously excited to join Barclaycard technology team.


Photo By Diliff (Own work) [CC BY-SA 3.0 or GFDL], via Wikimedia Commons

Investing in platforms

'Platform' is a word that one hears a lot nowadays in our industry, and it sometimes suffers from over-use. Here are a few thoughts on building platforms:

1. The art of being a platform is letting someone else be the product. I.e. platforms should underlie products, enabling diversity, not attempting themselves to be all things to all people.

2. Platforms are collections of capabilities, not baskets of features. Capabilities imply higher level of abstraction, enabling development of diverse feature sets without defining them exhaustively.

3. Platform business is hard. End users interface with products, not platforms. Most businesses compete on stickiness and profitability of their customer relationships. Catering to end users gives one greater scale, and so many businesses that begin by declaring themselves platforms eventually become very keen to own end-user relationships under their own brand, rather than letting third party product developers own them. Evolution of the Apple platform is a good case in point.

4a. For a product-oriented company of sufficient scale, factoring out underlying platform/capability set is key to scaling across global markets with their diverse requirements. Without a good platform foundation, feature flow demanded by differentiated global markets becomes more expensive, there's duplication of effort (and code), gradual accumulation of tangled dependencies and increase in brittleness.

4b. Exactly the same concern arises for SaaS companies that develop tailored solutions for anchor tenants of the platform and need to migrate to true multi-tenant platform.

5. However, a typical product manager is not incentivized to be platform-minded. Most corporate incentive structures are targeting delivery of specific, discrete, value-added features, not investment in broader capabilities that may yield dividends over the longer arc of time.

A humble proposal: 

For every investment, there should be an explicitly factored out portion that genuinely accrues to platform building (with the rest going towards feature flow). If it's 0% than at least it's explicitly 0%, not ambiguously implying platform investment or enablement in their absence.

Sometimes, making implied things explicit is all one needs to reach agreements or effect change.

Factoring and pipelining complex projects

We're all familiar with the concept of minimum viable product.

In fintech, those minimum requirements include a lot of regulatory compliance, legal and other complex concerns that can have significant ramifications on how every product feature is supposed to work. Throw in requirements from key partners and customers, and one can easily end up with an MVP with a large scope, and a lot of uncertainty and interdependencies among requirements/stories.

It's very easy then to start spending a lot of time reconciling it all together in attempt to arrive at a coherent picture all teams can align on in detail.

Finally, large investments are typically in the opposite of a skunk works/'fly under the radar until you've got something good to show' situation. Large investments imply commitments - in time as well as revenue. As in - in which quarter can it be delivered/ramped/booked. So for very good business reasons one ends up with a defined budget & timeline on one hand and a quivering mass of stories on the other. What to do?

The common mistake is to attempt to mitigate uncertainty by figuring it all out upfront in sufficient detail. Then everyone is spending most of their effort drawing and redrawing the complete picture and overall progress is very slow.

And, of course, every requirement/story has a product owner behind it, and no PO worth their salt will say 'uh, you know - the thing I'm responsible for - don't worry about it for now'.


Factoring:

In my experience, key has been to factor the entire whole into subgroups of mutually influencing requirements but with fewer fine-grained dependencies between these subgroups. Architecture can serve in leading or analytical capacity for this factoring.

Then decide by fiat which subgroup goes first and start cranking away at it. Yes, without other subgroups the 'MVP' won't see the light of day, and attacking other subgroups later doesn't mean they're less important (as their POs will be worrying about). But one has to start somewhere.

Pipelining:

And then, guess what - while subgroup 1 is being delivered, [C]POs, architects, dev leads and others with responsibilities beyond a single agile team or sprint can free up time to focus on subgroup 2, 3, n, n+1... creating a pipeline.

Seemingly simple, but requires persuading passionate people that being in n-th place in delivery pipeline doesn't imply being in the same place in terms of importance.

Microservices Architecture - Finding The Sweet Spot

Microservices architecture has been more in the spotlight lately, and with good reason.

As I wrote before in  Loose coupling and integration - the pendulum swings, splitting execution into multiple processes requiring remote invocation isn't free however.

In my recent discussions with many practitioners, I noticed that microservices architecture is typically put forth in opposition to a more coarse-grained approach where multiple domains  are lumped together, sometimes resulting in a 'God service' or 'everything service' anti-pattern.

However, it's dangerous to posit that more services is always better and create incentives to which teams and managers might respond - e.g. how many services did your team create last quarter? Nice, but that other team created twice as many!

So how to find the sweet spot in this spectrum? My answer is not to lump conceptually unrelated domains together. Another rule of thumb would be to look at persistence domains - if data/persistence architecture segregates domains properly, they should be exposed via separate service interfaces.

Fundamentally, services are simply public interfaces for domains. Each domain should have at least one. But creating too much fragmentation within each domain, exposing separate services for interrelated concerns shifts burden onto clients in terms of discovery, orchestration and invocation latency.

Ultimately, the goal of service interface design should be to allow clients to reason about the domain the service represents and integrate with it as simply and efficiently as possible.

Martin Fowler provides an  excellent write-up on microservices here.

Emergent software design?

Software seems to be the only area where we often try to combine designing a system with building it. Or declare that one should just start building it and the design will emerge.

Perhaps this is because in the material world, limitations of physics, materials properties and manufacturing processes are intuitively understood by non-specialists.

Building a house is a good example: a house is a composite of multiple subsystems: structural, electrical, plumbing, HVAC and so on. These subsystems won't just evolve independently to end up well-enmeshed in a coherent whole, and no future homeowner would assume such a thing and tell their contractors to just start building the house and not worry about blueprints.

Yet software has very real limitations in terms of complexity, physical resource utilization efficiency and so on, but these are not intuitively obvious to non-technologists. It is very easy to build something that appears functionally correct but is crippled in terms of scalability, maintainability, scalability, stability and so on.

So while not every detail has to be figured out and specified in advance, overall structure/architecture and how different subsystems should interact does need some forethought and definition.

Startup pivots - technology angle

This is what I've seen throughout my career as a typical [successful] startup lifecycle:

1. Build a product really fast - focus on experience, hopefully decent app code, basic persistence abstraction, monolithic database to keep things fast & simple.

2. If the product takes off, rapidly add features, variations for markets/segments, etc. Code grows more complex & tangled, resulting in greater brittleness and more bugs, database backend shows signs of strain.

3. Increase in number of users leads to business placing greater value on  availability and scalability, as well as performance, graceful degradation and similar matters.

Depending on the length of the the ramp up between stages 2 & 3, and the quality of tech design decisions at stage 1 despite pressure to get things out the door as fast as possible, techies at this point will be looking at a pile of technical debt sized somewhere between a molehill and a mountain.

What to do? Business still requires new features and capabilities to be added at the rapid clip. The more you build on the old without investing into refactoring or rewriting, the more the technical debt compounds (just like financial debt), increasing the tension between improving the internals vs the user-facing features.

Hopefully, as some money is now being made, reinvestment in improving technology platform is possible: break up the monolithic database (hopefully not too much logic is in the DB layer - stored procedures, etc.), try to break up increasingly tangled codebase, attempt to separate operations demanding synchronous, strong consistency guarantees from those that can be intermediated by async events. Words like SOA and EDA start to bounce around. If the business demands availability levels that require multiple datacenters, CAP theorem comes into play, eventual consistency strategies become critical, etc., etc.

In the meantime, some of the self-made components & modules may now have equivalent (or better) counterparts available as open-source or commercial software.

And it's at this point that investment in architecture, with focus on decomposition and encapsulation, as well as build/buy analysis unbiased by who wrote original in-house modules, becomes critical.

One often hears the term 'pivot' in startup parlance these days, usually referring to change in business model or product vision. I think the above illustrates that technology-wise, startups also have to execute one or more pivots to avoid being bound by the initial technical decisions that, just like business and product decisions, can rarely support the business forever.

Should we perhaps place greater emphasis on technology pivots? I think so.

iPhone as a platform

Following Apple's announcement of iPhones 5C and especially 5S, there's been much discussion about the merits of their various features and components.

Some have regretted that Apple didn't come out with any new 'whiz bang' features. Others questioned M7 motion co-processor in 5S or outright called it a gimmick.

I think such arguments reveal a superficial understanding of Apple's strategy. A few years back, everyone in the Internet industry couldn't stop talking about 'platforms'. Unsurprisingly, despite the hype, few companies actually figured out how to make money on their 'platform'.

Now, I think Apple has figured out how to both create a hardware+software platform AND make money on it - this is the entire iOS ecosystem.

So when Apple introduces new capability to its platform - whether it's 64-bit A7, or low-power-consumption M7 - it's not about what those things do by themselves, it's what they enable millions of developers to do. That's what happened with adding accelerometer, compass and other capabilities in the past.

Finally, from my own experience with various apps that attempt to track your activity (like the number of steps you've made) throughout the day, they consume way too much power to run all the time, reducing battery life unacceptably. M7 should be an awesome soltuion for that. Also, most avid joggers I know use their iPhones as fitness trackers and media players in one and are thrilled about the prospect of not having to recharge their iPhones more often.

Code maintenance & oil changes

Efficient organizations are typically KPI/metric-driven to a significant degree. Technology investments have to be translatable into business benefits, typically within 1 year horizon.

However, all code rots, with cruft and technical debt gradually accumulating within the codebase, necessitating refactoring and other maintenance that is bound to consume time and scarce resources. How can one demonstrate the necessity of such investment against competing priorities?

An analogy that comes to mind is engine oil changes. We spend time and money on them even though there's no direct, near-term proof that they are necessary. We essentially believe in their necessity because the experts tell us so. If we wait until the car starts showing signs of trouble, then it is probably too late and the repair will be far more expensive than an oil change.

So while managing technical debt is important, it is difficult to attach metrics to it that reliably demonstrate near-term benefits to business. Allocation of sufficient resources for it is then essentially a matter of trust between business and technology organizations.

Web apps - back to client-server?

Are we moving back to client server architecture for web applications?

The web started as a medium for sharing and linking content, with very limited interactivity. In time, ubiquity, elimination of software deployment to and maintenance on the clients, ability to update software on the server side transparently, SaaS and other factors made the web an application delivery channel as much as a content medium.

The biggest limitation that soon emerged is the quality of experience when using server-side UI state management - due to communication latency, responsiveness of UI leaves much to be desired, even as we moved from full page refreshes to partial updates, AJAX, parallelization of requests, etc.

Now that JavaScript has become a powerful, fully featured language, are we effectively moving back to client server model, in with rich JavaScript clients are running in browser containers and maintaining UI state on the client, with server side basically providing data and business logic processing?

The answer could well be yes, but to really match native apps in terms of quality of UI/experience, we need better resource management and control of the 'active' vs 'inactive' apps in the browser, when there are many windows & tabs open, refreshing, etc.

I think web as the medium for application delivery and browser as container is a trend that will continue and in time will force improvements in browsers, comm protocols and markup to enable better UI/experience.

Hiring Solution Architects!

Like many platform companies, PayPal has many strategic partners. Important partners come with their own requirements, often expressed as desired tweaks to core logic in multiple domains.

Needless to say, directly, programmatically implementing such tweaks across multiple domains would result in brittle, poorly maintainable code. The challenge is to ensure that there's enough flexibility in the key domains to support plurality of partner requirements without creating cruft/adding technical debt to the overall system.

This is best accomplished by technical solution architects on one hand working with partners to help shape their requirements against the fundamental rules and assumptions of the platform, and on the other hand, working with core domains comprising this platform to ensure enough flexibility is built in to support these requirements.

I'm currently looking for highly qualified individuals to fill the role of solution architects. Since all comments on this blog are moderated, you can apply by commenting on this post (I will not make submissions public).

People, Process & Org structures

Perhaps I'm biased by my start-up experience, but I strongly believe that having the right talent in the critical roles is the most important enabler of consistent, successful execution.

To put it another way, having the very best process and well-thought-out org structure in place will help, but will not guarantee success if you don't have the right people in critical roles. Conversely, if you do have the right people in those roles, they will find a way to work around the deficiencies in or change the process and the org structures.

In a start-up, you live and die by the quality of your team. In my experience, even large organizations still depend on a contingent of critical talent. So the most dangerous thing a technology company can do is to start treating its technical talent as a fungible resource - a pure cost to be managed down on a per-unit basis.

Paul Graham's analysis of Yahoo is a great illustration of this.

Errors & exceptions - client vs service

Let's say you're designing a service that exposes a getByQuery() interface. The client is responsible for composing a query and interpreting the result set returned.

Interesting question to consider: if a query results in an empty record set, is it because the query is valid but no matching records are found, or that the query is invalid in that particular context?

The designer of the service can choose to be helpful and indicate error conditions to distinguish the two cases, but at a cost of having to be more aware of the calling application's context.

Alternatively, the service could be context-agnostic and just return an empty record set. It then would be up to the client to have a complete enough suite of unit tests to have an assertion fail and indicate  an error in such a case.

I personally lean towards the services being more agnostic of client context and not getting encumbered by context-specific helper methods too early in their lifecycle. YMMV.