---
title: What is MLOps? Lifecycle, principles, and MLOps vs DevOps
canonical: https://telnyx.com/learn-ai/mlops
---
# What is MLOps? Lifecycle, principles, and MLOps vs DevOps

[Start building](https://telnyx.com/sign-up)

### Ask AI

* [![GPT](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/chatgpt_icon-ea55e09d.jpeg?width=92&format=webp)](https://chat.openai.com/?q=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.)
* [![Claude](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/claude_icon-c7c8eef3.webp?width=92&format=webp)](https://claude.ai/new?q=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.)
* [![Perplexity](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/perplexity_icon-cd5b74f1.webp?width=92&format=webp)](https://www.perplexity.ai/search?q=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.)
* [![Gemini](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/gemini_icon-4ff1cbf4.webp?width=92&format=webp)](https://www.google.com/search?udm=50&aep=11&q=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.)
* [![Grok](https://d2efw2akbx8gaf.cloudfront.net/https://dotcom-payload-assets-prod.s3.us-east-1.amazonaws.com/cms-media/grok_icon-e588e5bd.webp?width=92&format=webp)](https://x.com/i/grok?text=What makes Telnyx's global network different, breakdown call quality for normal phone calls and voice AI calls.)

What is MLOps? Lifecycle, principles, and MLOps vs DevOps

[Back to Glossary](https://telnyx.com/learn-ai)

# What is MLOps? Lifecycle, principles, and MLOps vs DevOps

MLOps is how teams take machine learning models to production and keep them working. The lifecycle, core principles, maturity levels, and MLOps vs DevOps.

Editor: [Andy Muns](https://www.linkedin.com/in/andy-muns/)

**Updated October 2026**

MLOps, short for machine learning operations, is the set of practices that takes machine learning models from development into production and keeps them working there. It combines machine learning, software engineering and data engineering: versioning data, code and models, automating training and deployment, and monitoring models after release so they can be retrained when the world changes.

In their paper on hidden technical debt, [Sculley and colleagues](https://papers.nips.cc/paper%5Ffiles/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html) show that the machine learning code is only a small part of a real-world machine learning system. Most of it is the infrastructure around the model.

## What is MLOps?

This content was generated with the assistance of AI. Our AI [prompt chain](https://telnyx.com/learn-ai/promptchain-ai) workflow is carefully [grounded](https://telnyx.com/learn-ai/grounding-ai) and preferences .gov and .edu citations when available. **All content is reviewed by a Telnyx employee to ensure accuracy, relevance, and a high standard of quality.**

## Sign up and start building.

[Sign Up](https://telnyx.com/sign-up)[Contact Us](https://telnyx.com/contact-us)

MLOps is what turns a model that works in a notebook into a system that keeps working in production.

[Kreuzberger, Kühl and Hirschl](https://doi.org/10.1109/ACCESS.2023.3262138) define it as covering the whole life of a machine learning product, from design to deployment, monitoring and scaling.

An intent model, the model that tells a [voice AI agent](https://telnyx.com/products/voice-ai-agents) why someone is calling, shows the gap. A data scientist can train a first version on last year's calls in a notebook.

Running it in production is a different job from training it. The model has to be retrained as callers' reasons change, deployed without breaking live calls, and watched so a drop in accuracy is caught before customers notice.

Kreuzberger and colleagues also list the roles that share this work, including data scientists, data engineers, software engineers, DevOps engineers and ML engineers, who combine skills from the other roles.

## Why MLOps matters

After launch, the data a model sees keeps changing, and the model has to change with it.

Sculley and colleagues describe the result as technical debt: building a model is fast, but maintaining it in production carries "massive ongoing maintenance costs".

One reason is entanglement, which they summarize as "changing anything changes everything". A new input feature, a different training sample or a changed threshold can shift how every other part of the model behaves.

Another is change in the outside world. When live data drifts away from the training data, accuracy falls even though no code changed. [Data drift and concept drift](https://telnyx.com/learn-ai/concept-drift-vs-data-drift) are two of the main ways this happens.

Data drift changes the inputs: callers start describing the same problems in new words. Concept drift changes what the inputs mean: the same words now point to a different intent.

For the intent model, a new product launch can add a reason for calling that the model has never seen. Nothing breaks, but more calls are routed to the wrong place.

## The MLOps lifecycle

The MLOps lifecycle is a loop, not a line. Each step hands its output to the next, and monitoring feeds back into retraining. For the intent model, the loop looks like this:

1. Data collection and preparation gathers labeled calls, checks their quality and saves a versioned copy of the training set.
2. Training and evaluation trains candidate models, tracks every experiment and compares them on held-out calls.
3. Validation tests the chosen model against quality checks before release, such as accuracy on each caller intent.
4. [Model deployment](https://telnyx.com/resources/model-deployment) puts the model into production, where it answers live requests.
5. Monitoring tracks accuracy, input data and latency, and flags drift.
6. Retraining starts when monitoring shows a drop: the model learns from new data and the loop begins again.

Kreuzberger and colleagues describe the architecture behind this loop as nine components. Besides a code repository, CI/CD, training infrastructure, model serving and monitoring, four may be new to a software team:

* A workflow orchestrator runs the pipeline steps in the right order.
* A feature store keeps the inputs that models share, such as call features, in one place for both training and live predictions.
* A model registry stores each trained model with its metadata, so deployment pulls a known version.
* A metadata store records each training run: when it ran, its parameters, its results and which data and code produced the model.

## MLOps principles and practices

Kreuzberger and colleagues name nine principles: CI/CD automation, workflow orchestration, reproducibility, versioning, collaboration, continuous training and evaluation, metadata tracking, continuous monitoring and feedback loops.

Reproducibility means a team can rerun an experiment and get the same result. It rests on versioning, supported by workflow orchestration and metadata tracking.

Versioning covers data, models and code, not code alone. It lets a team reproduce any past model and trace which data trained it, which compliance and audits require.

Collaboration means data scientists and engineers work on the same data, models and code instead of handing files across a wall.

CI/CD automation builds, tests and deploys changes automatically, so a bad change fails a test instead of reaching users.

Continuous training retrains the model on new data on a schedule or when monitoring triggers it. Kreuzberger and colleagues require every retraining run to include an evaluation run that checks the new model's quality.

Continuous monitoring checks data, the model, code, infrastructure and serving performance, such as prediction accuracy. Feedback loops carry what monitoring finds back into development and retraining.

Open-source tools cover each practice: [DVC](https://dvc.org/) versions data the way Git versions code. Apache Airflow and Kubeflow Pipelines orchestrate workflows, Feast is a feature store, and Prometheus collects monitoring metrics.

## MLOps maturity levels

MLOps maturity models describe how much of the lifecycle a team has automated. They disagree on the number of levels: a [2025 comparison](https://doi.org/10.1109/CAI64502.2025.00118) by Stone and colleagues lists models with three, four and five levels.

The models agree on the direction. They start with manual processes and move toward full automation of data collection, model development, versioning, delivery, monitoring and retraining.

For the intent model, the climb starts with a data scientist who retrains by hand and gives engineers a model file. Next, retraining runs as an automated pipeline on new calls.

Then CI/CD tests and deploys each new version automatically. At the top, monitoring detects drift and starts retraining without anyone stepping in.

A team on the first step can climb with three habits. It versions the training data and the model alongside the code.

It adds an evaluation test that a new model must pass before release, and it logs the model's predictions so they can be monitored.

## Challenges of MLOps

MLOps is difficult because problems show up at every stage, not only in the model. A [survey of deployment case studies](https://doi.org/10.1145/3533378) by Paleyes, Urma and Lawrence found that practitioners face them from data collection to monitoring.

Data quality, drift and reproducibility are the problems the earlier sections cover. Two more are easy to miss.

Interpretability is one concern the survey lists. In regulated settings, a team may need to explain why a model made a decision.

Accuracy alone is not enough to judge a model in production. The survey notes that performance metrics should reflect what users and the business care about, which can include bias and fairness.

## MLOps vs DevOps vs AIOps vs LLMOps

MLOps extends DevOps from code to data and models. DevOps, which emerged around 2008 and 2009, closes the gap between software development and operations with collaboration and CI/CD.

MLOps is more than DevOps because of what changes. In software, the code defines the behavior. In machine learning, the data shapes the model as much as the code does, and a model can degrade without any code change.

Unlike MLOps, AIOps is artificial intelligence for IT operations. It applies machine learning to running software systems, for example to detect failures in large distributed systems, as [a survey by Zhang and colleagues](https://arxiv.org/abs/2406.11213) describes.

LLMOps is MLOps applied to large language models. Stone and colleagues fold it into the MLOps lifecycle and start with a decision: an existing model or a custom one.

Unless a team needs a specialized model and has the data, compute and expertise to keep it maintained, the paper finds an existing model generally preferable. The work then shifts to prompts, fine-tuning, retrieval and evaluation.

## MLOps for voice AI agents

A voice AI agent built on a language model needs the same discipline as any model in production: versions, tests before release, a gradual rollout and monitoring on real calls.

For an agent on a hosted model, the unit that changes is its prompts, tools, model settings and voice, not its weights.

The risk is quiet: a changed prompt can make an agent misread a common request or call the wrong tool while the call still connects, so nothing looks broken in the logs.

In [Telnyx AI Assistants](https://developers.telnyx.com/docs/inference/ai-assistants/version-testing-traffic-distribution), each change of instructions, tools or voice can be saved as a new assistant version. AI Tests, a feature in the Telnyx Portal, check a version against success criteria before it goes live.

For testing at scale, AI Assistants integrate with Coval, a simulation and evaluation platform for voice and chat agents.

In a CI/CD pipeline, Coval runs a library of test scenarios on every assistant change and blocks deployments that introduce regressions.

Versions can then take a percentage of live calls, and the remaining calls go to the main version as a safety fallback. Teams can also route their own number or a test group to the new version first, as a [canary deployment](https://telnyx.com/resources/voice-ai-canary-deployments).

For monitoring, the [Langfuse integration](https://developers.telnyx.com/docs/inference/ai-assistants/agent-observability) traces every model call, tool execution and conversation turn, including token usage, latency and cost.

AI Insights, a Telnyx feature, can analyze each conversation and send the results to a webhook when the conversation completes.

What monitoring finds becomes the next assistant version, which goes back through AI Tests and Coval before it takes live calls. That closes the loop.

Teams running their own models on the [Voice API](https://telnyx.com/products/voice-api) get the raw call audio over WebSockets through [media streaming](https://developers.telnyx.com/docs/voice/programmable-voice/media-streaming). Versioning, testing and monitoring are then theirs to build.

## Frequently asked questions

### Is MLOps harder than DevOps?

MLOps has more moving parts than DevOps. Code, data and models all change, and a model can lose accuracy without any code change. So MLOps adds data versioning, retraining and model monitoring to the DevOps toolkit.

### Does MLOps require coding?

Yes. Pipelines, tests and deployment steps are code that ML engineers write, version and maintain like any other software.

### What are the three levels of MLOps?

There is no single standard. The three levels often quoted come from one cloud provider's model: manual process, ML pipeline automation and CI/CD pipeline automation. Other models use four or five levels.

### What is LLMOps?

LLMOps is MLOps for large language models. It usually starts from an existing model and focuses on prompts, retrieval, fine-tuning and evaluation instead of training a model from scratch.

### What is the difference between ML and MLOps?

Machine learning builds models that learn from data. MLOps runs those models in production: it versions, deploys, monitors and retrains them so they keep working as data changes.

## Sources

* Sculley, D., et al. (2015). [Hidden technical debt in machine learning systems](https://papers.nips.cc/paper%5Ffiles/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html). NeurIPS 2015.
* Kreuzberger, D., Kühl, N., and Hirschl, S. (2023). [Machine learning operations (MLOps): overview, definition, and architecture](https://doi.org/10.1109/ACCESS.2023.3262138). IEEE Access, 11, 31866-31879.
* Paleyes, A., Urma, R.-G., and Lawrence, N. D. (2022). [Challenges in deploying machine learning: a survey of case studies](https://doi.org/10.1145/3533378). ACM Computing Surveys.
* Stone, J., Patel, R., Ghiasi, F., et al. (2025). [Navigating MLOps: insights into maturity, lifecycle, tools, and careers](https://doi.org/10.1109/CAI64502.2025.00118). IEEE CAI 2025.
* Zhang, L., Jia, T., Jia, M., et al. (2024). [A survey of AIOps for failure management in the era of large language models](https://arxiv.org/abs/2406.11213).
* DVC. [Data Version Control](https://dvc.org/).
* Telnyx. [Testing and traffic distribution for AI Assistants](https://developers.telnyx.com/docs/inference/ai-assistants/version-testing-traffic-distribution), [Agent observability](https://developers.telnyx.com/docs/inference/ai-assistants/agent-observability) and . Telnyx developer documentation.

Share on Social

## Jump to:

[What is MLOps?](#what-is-mlops)[Why MLOps matters](#why-mlops-matters)[The MLOps lifecycle](#the-mlops-lifecycle)[MLOps principles and practices](#mlops-principles-and-practices)[MLOps maturity levels](#mlops-maturity-levels)[Challenges of MLOps](#challenges-of-mlops)[MLOps vs DevOps vs AIOps vs LLMOps](#mlops-vs-devops-vs-aiops-vs-llmops)[MLOps for voice AI agents](#mlops-for-voice-ai-agents)[Frequently asked questions](#frequently-asked-questions)[Sources](#sources)

## Sign up for emails of our latest articles and news

[Media streaming](https://developers.telnyx.com/docs/voice/programmable-voice/media-streaming)