Agile approaches

DevOps

DevOps is a set of practices that combines software development (Dev) and IT Operations (Ops). With respect to Agile approaches, it encourages more collaboration and process sharing. These practices are essentially based on shorter release cycles with an increase in the frequency of deployments and continuous deliveries.

• Continious Integration/Continious Deployement (CI/CD):

  • Continuous Integration: Frequently integrate and merge code changes into the main code branch by setting up automated tests that run whenever new code is validated.
  • Continuous Deployment: Respect frequent and automated deployment of new application versions in a production environment. (Tools: Jenkins, CicleCI, Gitlab…)

• Version management: Manage different versions of code by tracking revisions and change history to facilitate code review and retrieval. (Tools: Git, Github, Apache Subversion, bitbucket…)

• Continuous monitoring: Set up continuous monitoring and alerting tools to have complete real-time visibility of the applications/infrastructure performance and integrity. (Tools: ELK, Sensu, Zabbix, Nagios, Amazon CloudWatch…)

The Product Owner

The Product Owner plays an important role in an Agile organization. We can categorize his tasks and responsibilities into 5 main families:

• Client / team interface: The PO is the representative of the customer and the product users in the Agile team. He/She transforms customer needs into technical features and guarantees that the product meets their expectations. He/She sets up metrics and KPIs to assess product usage and users’ behavior in order to identify potential improvements.

• Product Development management: The PO is the master of the Backlog and the referent of product roadmap. He/She is in charge of:

  • Proposing functional specifications and ensuring the correct translation in terms of technical specifications
  • Writing User Stories and defining their acceptance criteria
  • Defining acceptance protocols and criterias
    
  • Validating achievements using functional tests
  • Guarantee of delivered features’ quality
  • Feeding the roadmap with prioritized features.

• Collaboration: By nature, the PO must collaborate with different members of the development team. As a product referent, he/she also needs to coordinate with other teams and departments: sales, marketing, infrastructure, etc.

• Communication: The PO must share the product vision inside and outside the company. He/she can organize Product Discovery workshops with the team and clients. He/She is also responsible for the quality and completeness of the documentation.

• Technology watch: The PO must ensure a continuous watch on the market in order to be able to situate the product among the competition.

Data Science produtcs

We can categorize Data Science products into two main families: Data Analytics & Machine Learning.

• Data Analytics: They essentially consist of a data processing pipelines attempting to extract knowledge and insights of the studied phenomena by using collected raw data:

  • Data Extraction: Frequency, disponibility, schemas, formats, data models…
  • Data Storage: Data warehousing, datalake, relational Databases, NoSql Databases, Cloud Storage platforms….
  • Data Exploration: Profiling, Descriptive statistics, visualization…
  • Data Transformation: Cleaning, aggregating, automatic reporting…
    

• Machine Learning: It refers to products that are based on learning models (supervised, unsupervised, self-supervised) and generate services aiming at replacing human tasks. We can distinguish 5 families of ML solutions on the market: recommendation and scoring systems (e.g. movie recommendation), prediction of actions or events (e.g. prediction of clicks on a video), classification (e.g. classification of emails as spam / not spam), generative models (e.g. translation), and clustering (e.g. clients segmentation).

In addition to the data analytics pipelines, ML products integrate specific techniques:

• Data transformation: division into test/train, feature engineering, data augmentation…
• Modeling: testing different statistical models, model design, hyperparameters tuning……
• Model training
• Model Evaluation and validation: cross validation, evaluation metrics…
• Model serving
• Model monitoring

DataOps & MLOps

DataOps and MLOps are relatively new disciplines that seek to systemize the entire Data Science workflows (data pipelines and ML lifecycles), from science to production. They refer to a set of practices and tools aiming at improving the collaboration between data science and DevOps teams in order to accelerate the time to market of data science applications.

You can find here a comprehensive documentation about MLOps principles and workflow.

MLOps: Continuous delivery and automation pipelines in machine learning (source: “https://ml-ops.org/content/mlops-principles”)

The Data Science Product Owner

Due to the differences between traditional IT solutions and data-driven solutions, new skills are needed for accurately managing Data and ML/AI products. Specificities of Data Science products (in terms of mathematical foundation, exploratory dimension and data-driven workflow) are significantly impacting and reshaping product development lifecycle. Therefore, Product Owner role and knowledge may expand to include the following aspects:

• Product Development management:

  • Feasibility assessment: Streamlining the interface between R&D processes (feasibility assessment, exploration, literature review…) and production.
  • Scientific approach: Integrating the necessary steps for implementing scientific approach in product development (observation, formulation of problems, hypothesis testing, phenomenon modelling, experimentation, evaluation…)
  • Intermediate delivery: Designing intermediate delivery adapted to the experimental nature of data science workflows which do not necessarily result in a deployed feature.
  • Complex estimation: Estimating efforts and charges for deploying a Data Science feature may be complex due to the exploratory phase and dependency to data availibility and quality. The PO needs hands-on experience to better estimate featuress value and complexity and integrate them to the development cycle.
  • Data-driven: As a referent of a data-driven product, the PO must know very well the input data (quality, availability, formats, source, etc.). Manipulating the data helps for better anticipating constraints and potential improvements of the delivered solution.
  • Modelling: The Data Science PO should master the Machine Learning pipeline (model designing, testing, training, validating,serving…) and be able to explain model weakness for guiding the product roadmap to the most effective improvements.
    
  • Validation: The validation of data-driven features requires skills in data wrangling, use of notebooks to standardize tests, and a good understanding of evaluation metrics, etc.

• Collaboration: In addition to the traditional development team, the Data Science PO must coordinate the tasks of various profiles making up the Data team: Data Analysts, Data Scientists, Data engineers, Data Architects… He/she should know their responsibilities, skills and constraints in order to better coordinate the development phases.

• Technology watch: The development of industrialized applications based on data science and machine learning remains a fairly young profession. It is important for the data product team, and particularly the PO, to be up to date in terms of tools, techniques, methods, algorithms and technologies… This makes it possible to better anticipate the changes to be integrated in order to improve some performance.

I dedicate this article to my ex-colleagues in OpenfieldLive, together we tried to apply Scrum framework for developing Location Intelligence solutions, we tried to find our own methodology and collaboration degree between Data and DevOps teams, and with them I experimented the role of Data Science Product Owner.