Machine Learning and Software Development Lifecycle

2019-12-27

The new Software Development Lifecycle (SDLC) means finding ways to adapt to your machine learning workflow. As data scientists currently spend a significant amount of time on infrastructure and processes rather than building models, finding methods that effectively integrate the software lifecycle with machine learning is crucial for their productivity and job satisfaction.


However, this presents challenges. Machine learning introduces patterns and technical issues that SDLC cannot solve. To address this problem, we need to adapt to SDLC while also bridging the cultural gap between data scientists and other developers.It is important to remember that the field of machine learning is still evolving, and therefore, it is not standardized. Data science is more of an art than a standard software development and involves a significant amount of research tasks. In contrast, standard software developers tend to adapt their technology to the work at hand and fit into their environment. For example, they learn another language when they find a new job because it is the language most commonly used within their architecture. On the other hand, machine learning tasks are often specific to a language or a set of frameworks, so they use what is most suitable for the job, resulting in a more heterogeneous environment.


In one model, machine learning developers may use Python NLTK at a specific version level, while for other tasks, they combine R with TensorFlow and have GPU acceleration, and so on. This means that when most standard service software does not run R at all, the model must be put into production. This is a language that DevOps people have never encountered before, so they need a way to adjust their service workflow to accommodate these more heterogeneous environments.


Another area where data scientists and DevOps teams struggle to coordinate is monitoring and optimization. In the world of machine learning, testing often only occurs during the model development process and not in a production environment on a server. However, for standard developers, they not only consider whether the initial components have been set up correctly but also whether they can continuously validate it and still produce the expected results when used worldwide.The second issue is that DevOps personnel do not know how to monitor models. They are not accustomed to considering model drift or probabilistic results, so they may test a machine learning model and find that the results vary slightly each time, which may lead them to believe that the model has failed. However, data scientists would know that there must be a margin of 10% in the results.


Predictability is also a challenge. SDLC has predictable scheduled release times, while the data science cycle is unstable and unpredictable. In this situation, software developers tend to set up on a two-week schedule and plan their work in advance, while researchers tend to work on a very abstract schedule, requiring a day or two months.


The cloud environment is another area to consider. For developers who primarily write code, there are many additional operations - you need to be able to set up servers, set up and connect to databases, and often manage them in a cloud infrastructure. However, data scientists are not accustomed to this workflow. They tend to keep everything separate on their laptops or through hosted services. They are also used to training and testing in a self-managed environment and previously unlikely to work with DevOps. For them, this is a significant learning curve and often a confusing process involving unfamiliar jargon that they must decipher in order to communicate with IT staff about their work.

RECENT AI NEWS

RECENT AI TOOLS