This article is based on a guest feature by Debashis Saha, Vice President of Platform Engineering at Intuit, originally published on insidebigdata.com. Saha previously served as Executive VP and CTO at Jiff, Inc., and VP of Commerce Platform Infrastructure at eBay. He holds an M.S. in Electrical Engineering and Computer Science from MIT and a B.Tech in Computer Science from IIT Kharagpur.
Why Machine Learning Needs Its Own DevOps
DevOps transformed software delivery by breaking down the wall between development and operations teams, enabling faster, more reliable releases. As machine learning moves from research notebooks into production systems, it faces an analogous—and arguably more complex—challenge. The term ML Ops describes the discipline of applying DevOps principles to the full machine learning lifecycle, spanning data engineering, model development, and production deployment.
Where traditional DevOps mediates between two primary personas—developers and operations engineers—ML Ops must reconcile four distinct roles, each with its own tooling, cadence, and constraints:
- Data Engineers – build and maintain pipelines that move and transform raw data
- Data Scientists – develop, train, and validate machine learning models
- Software Engineers – integrate models into production applications
- Operations Teams – monitor, scale, and maintain live systems
Balancing the requirements, velocity, and toolchains of all four groups—while satisfying business goals and regulatory constraints—is one of the central problems ML Ops is designed to solve.
The Core Problem: Models That Never Leave the Lab
A widely recognized pain point in applied machine learning is that a significant proportion of data science projects never reach production. The reasons are largely operational rather than scientific:
- Integrating a trained model with a production environment requires manual, error-prone steps
- Data preparation, model training, and deployment pipelines often rely on different, incompatible tools
- There is little visibility into the end-to-end process, making debugging and auditing difficult
- Handoffs between data engineers, data scientists, and software engineers are informal and slow
The cumulative effect is that organizations sit on large volumes of potentially valuable data without being able to extract actionable intelligence from it at scale. This gap between experimentation and operationalization is where ML Ops intervenes.
What ML Ops Encompasses
ML Ops is not a single tool or platform—it is a set of practices and, increasingly, a category of tooling that stitches together three complex lifecycle stages:
1. Data Engineering
Raw data must be prepared, cleaned, refined, and transformed before it is suitable for machine learning. This is an iterative process. Without automation and versioning at this stage, downstream model quality is unpredictable and difficult to reproduce.
2. Data Science (Model Lifecycle)
Model development involves iterative cycles of training, testing, validation, and optimization. Tracking experiments, managing model versions, and ensuring reproducibility are foundational requirements—especially in regulated industries such as banking and insurance where auditability of model decisions is a legal obligation.
3. Deployment and Operations
Deployment covers integration testing, staged rollout, A/B experimentation, real-time monitoring, performance engineering, and ongoing operations. Models degrade over time as real-world data distributions shift, so continuous monitoring and rapid retraining are critical capabilities.
A mature ML Ops platform connects all three stages into a seamless, auditable workflow—making collaboration between disciplines practical rather than aspirational.
Key Principles for Applying ML Ops in Practice
End-to-End Automation
Manual handoffs are the enemy of velocity and reliability. Automating the pipeline from data ingestion through model deployment reduces human error, speeds iteration, and creates a repeatable process that can be version-controlled and audited.
Transparency and Observability
Every stage of the ML lifecycle—data transformations, training runs, model versions, deployment events, and live inference metrics—should be logged and queryable. Lack of visibility is consistently cited as the first-order barrier to trustworthy production ML.
Continuous Testing
Testing must occur at every phase: data quality checks before training, model evaluation during training, integration tests before deployment, and statistical monitoring after deployment. Skipping testing at any stage compounds risk and erodes stakeholder trust in ML-driven decisions.
Rapid Model Refresh
Business conditions and data distributions change continuously. An ML Ops platform must support near-real-time model retraining and redeployment, otherwise models become stale and their predictions misleading.
Reproducibility
The ability to recreate any model artifact from a fixed dataset and a fixed version of training code is both a quality requirement and, in regulated industries, a compliance requirement. Data versioning, environment management, and experiment tracking tools are essential building blocks.
The State of the Ecosystem
When Saha’s original article was published in 2018, the ML Ops tooling landscape was nascent—analogous, as he noted, to the era when the first compilers and editors were emerging for software development. Two factors constrained the ecosystem at that time:
- A rapidly shifting landscape – data platforms, ML frameworks, and cloud infrastructure were all evolving quickly, making it difficult to build durable integrations
- Skill set gaps – few practitioners understood data engineering, data science, and software engineering deeply enough to design abstractions that worked across all three
The ecosystem has matured substantially since 2018. Categories such as feature stores, model registries, pipeline orchestrators, and model monitoring platforms now have multiple commercial and open-source offerings. However, integration across the full lifecycle remains an active engineering challenge for most organizations.
The Business Case for ML Ops
The fundamental objective of ML Ops is straightforward: deliver trustworthy, actionable intelligence from data to business problems, reliably and at speed. Organizations that invest in ML Ops infrastructure gain several concrete advantages:
- Faster time from experimentation to production value
- Greater consistency and reliability of model behavior in production
- Ability to safely experiment with different models, datasets, and configurations in parallel
- Auditability and compliance readiness, particularly in regulated sectors
- Reduced dependency on heroic individual effort; institutional knowledge encoded in repeatable processes
As cloud providers continue to expand managed ML services, the infrastructure cost of building an ML Ops platform is decreasing. The organizational and cultural investment—getting data engineers, scientists, software engineers, and operations teams to work within a shared workflow—remains the harder, more enduring challenge.
Looking Ahead
A well-implemented ML Ops platform has the potential to be transformative at scale. By unlocking the latent intelligence in enterprise data and making it accessible to business processes in a reliable, repeatable way, ML Ops is not merely an engineering convenience—it is the operational foundation on which trustworthy AI systems are built.
For DevOps engineers moving into the ML space, the core instincts transfer directly: automate ruthlessly, instrument everything, version all artifacts, and design for failure. The domain is different; the discipline is the same.