As data-driven and AI-first applications are on the advance, we extend our best practices for DevOps and agile development with new concepts and tools. The corresponding buzz words would be continuous intelligence and continuous delivery for machine learning (CD4ML).

For our current project, we researched, tried different approaches and build a proof of concept for a continuously improved machine learning model. That’s why I got interested in this topic and went to a meet up at Thoughtwork’s office. Christoph Windheuser (Global Head of Artificial Intelligence) shared their experience in this field and gave a lot of insights. The following post summarizes these thoughts [1] with some notes from our learning process.

CD4ML continuous intelligence cycle

The continuous intelligence cycle

1- Acquire data

Get your hands on data sets. There are multiple ways, most likely the data is bought, collected or generated.

2- Store, clean, curate, featurize information

Use statistical and explorative data analysis. Clean and connect your data. At the end, it needs to be consumable information.

3- Explore models and gain insights

You are going to create mathematical models. Explore them, try to understand them and gain insights in your domain. These models will forecast events, predict values and discover patterns.

4- Productionize your decision-making

Bring your models and machine learning services into production. Apply your insights and test your hypothesizes.

5- Derive real life actions and execute upon

Take actions on your gained knowledge. Follow up with your business and gain value. This generates new (feedback) data. With this data and knowledge, you follow up with step one of the intelligence cycle.

Productionizing machine learning is hard

There are multiple experts collaborating in this process circle. We have data hunters, data scientists, data engineers, software engineers, (Dev)Ops specialists, QA engineers, business domain experts, data analysts, software and enterprise architects… For software components, we mastered these challenges with CI/CD pipelines, iterative and incremental development approaches and tools like GIT and Docker (orchestrators). However, in continuous delivery for machine learning we need to overcome additional issues:

  • When we have changing components in software development, we talk about source code and configuration. In machine learning and AI products, we have huge data sets and multiple types and permutations of parameters and hyperparameters. GitHub for example denies git pushes with files bigger than 100mb. Additionally, copying data sets around to build/training agents is more consuming than copying some .json or .yml files.
  • A very long and distributed value chain may result in a "throw over the fence" attitude.
  • Depending on your current and past history, you might need to think more about parallelism in building, testing and deploying. You might need to train different models (e.g. a random forest and an ANN) in parallel, wait for both to finish, compare their test results and only select the better performing.
  • Like software components, models must be monitored and improved.

The software engineer’s approach

In software development, the answer to this are pipelines with build-steps and automated tests, deployments, continuous monitoring and feedback control. For CD4ML the cycle looks like this [1]:

CD4ML Pipelines

There is a profusely growing demand on the market for tools to implement this process. While there are plenty of tools, here are examples of well-fitting tool chains.

stack discoverable and accessible data version control artifact repositories cd orchestration (to combine pipelines)
Microsoft Azure Azure Blobstorage / Azure Data lake Storage (ADLS) Azure DevOps Repos & ADLS Azure DevOps Pipelines
open source with google cloud platform [1] Google cloud storage Git & DVC GoCD
stack infrastructure (for multiple environments and experiments) model performance assessment monitoring and observability
Microsoft Azure Azure Kubernetes Service (AKS) Azure machine learning services / ml flow Azure Monitor / EPG *
open source with google cloud platform [1] GCP / Docker ml flow EFK *

* Aside from general infrastructure (cluster) and application monitoring, you want to:

  • Keep track of experiments and hypothesises.
  • Remember what algorithms and code version was used.
  • Measure duration of experiments and learning speed of your models.
  • Store parameters and hyperparameters.

The solutions used for this are the same as for other systems:

search engine log collector visual layer
EFK stack elasticsearch fluentd kibana
EPG stack elasticsearch prometheus grafana
ELK stack elasticsearch logstash kibana

[1]: C.Windheuser, Thoughtworks, Slideshare: https://www.slideshare.net/ChristophWindheuser/cd4ml-thoughtworks-meetup-munich-christoph-windheuser-may-8th-2019

I went to a great session about CQRS, Event Sourcing and domain-driven Design (DDD) on the Software Architecture Summit. The speaker Golo Roden (@goloroden) did a fantastic job selling these concepts to his audience with a great storytelling approach. He explained why CQRS, Event Sourcing and DDD fit together perfectly while replicating the www.nevercompletedgame.com for his daughter. This is what he shared with us.

Domain-driven Design

The more enterprise-y your customer the weirder the neologisms get.

We – as software engineers – struggle to understand business and domain experts. Once we understand something, we try to map it to technical concepts. Understood the word "ferret"? Guess we need a database table called "ferret" somehow. We then proceed to inform our business colleagues, "deploying a new schema is easy as we use Entity Framework or Hibernate as OR-mapper". He thinks we understood, we think he understood. Actually, nobody understood anything.
As software engineers we tend to fit every trivial and every complex problem into CRUD-operations. Why? Because its "easy" and everyone does it. If it was that easy, software development would be effortless. Rather trying to fit problems in a crud pattern, we should transform business stories into software.
That’s why we should use domain-driven design and ubiquitous language.
Golo Roden proceeds to create a view on the nevercompletedgame with ubiquitous language. So nobody asks, "what does open a game mean" and there is no mental mapping.
I won’t go into detail here, but an example can show why we need this.

  • Many words have one meaning: When developing a software for a group of people, sometimes we call them users, sometimes end-users, sometimes customers etc. If we use different words in the code or documentation and developers join a project later – they might think there is a difference between these entities.
  • One word has many meanings: Every insurance software has "policies" somewhere in its system. Sometimes it describes a template for a group of coverages, sometimes it’s a contract underwritten by an insurance, sometimes a set of government rules. You don’t need to be an expert to guess this can go wrong horribly.

CQRS

Asking a question should not change the answer

Golo Roden jokes, "CQRS is CQS on application level", but actually it’s easy to understand this way, once you read a single article about CQS. Basically, it’s a pattern where you separate commands (writes) and queries (reads): CQS.

  • Writes do not return any values and change the state of an object.
    stack.push(23); // pushes value 23 onto the stack; returns nothing
  • Reads return a value and don’t change the state.
    stack.isEmpty() // does not change state; returns a isEmpty boolean
  • But don’t be fooled! Stacks are not following the CQS pattern.
    stack.pop() // returns a value and changes state

Separating them on application level means: exposing different APIs for reading (return a value; do not change state) and writing (change state; do not return value *). Or phrased differently: Segregate responsibilities for commands and queries: CQRS.

* for http: always returns 200 before doing anything

Enforcing CQRS could have this effect on your application:

For synchronizing patterns see patterns like the saga pattern or 2 phase commit . For more reference see: Starbucks Does Not Use Two-Phase Commit

Event Sourcing

When talking about databases (be it relational or NoSQL) often we save the current state of some business item persistently. When we are ambitious we save a history of these states. Event sourcing follows a different approach. There is only one initial state, change requests to this state (commands) and following manipulating operations (events). When we want to change the state of an object, we set up a command. This triggers an event (that’s published so some kind of queue) and most likely is persistent in a database.

Bank account example: we start with 0 € and do not change this initial value when we add or withdraw money. We save the events something like this:

Date EventId Amount Message
2019-01-07 e5f9e618-39ad-4979-99a7-342cb1962266 0 account created
2019-01-11 f2e98590-7795-4cf7-bdc2-1794ad39874d 1000 manual payment received
2019-01-29 cbf44bfc-7a5e-4514-a906-a313a6e0fb5e 2000 saylary received
2019-02-01 32bc638c-4783-45b8-8c1e-bebe2b4528a1 -1500 rent payed

When we want to see the current balance, we read all the events and replay what happened.

const accountEvents = [0, 1000, 2000, -1500];
const replayBalance = (total, val) =>  total + val;
const accountBalance = accountEvents.reduce(replayBalance);

Once every n (e.g. 100) values we save a snapshot to not have to replay too many events. Aside from the increased complexity this has some side effects which should not be unadressed.

  • As we append more and more events, the data usage increases endless. There are ways around removing "old" events etc. and replacing them with snapshots, but this destroys the intention of the concept.
  • Additionally, as more events are stored, the system gets slower as it has to replay more events to get the current state of an object. Though, snapshotting every n events can get deterministic maximum execution time.

While there are many contra arguments there is one the key benefit why its worth: your application is future proof, as you save "everything" for upcoming changes and new requirements. Think of the account example from the previous step. You can implement/analyze everything of the following:

  • "How long does it take people to pay their rent once they got their salary"
  • "How many of our customers have two apartments? How much is the difference between both rents?"
  • "How many of our customers with two apartments with at least 50% in price difference need longer to pay off their car credit?"

To sum it up and coming back to our initial challenge, our simple CRUD application with domain-driven Design, CQRS and event sourcing would have transformed our architecture to something like this:

While this might solve some problems in application and system development this is neither a cookie-cutter approach nor "the right way" to do things. Be aware of the rising complexity of your application, system and enterprise ecosystem and the risk of over-engineering!