MLOps and LLMOps: A Systematic Survey of Deploying, Monitoring, and Operating Large Language Models in Production

Machine learning operations (MLOps) emerged to close the gap between a model that works in a notebook and a system that works in production, and the arrival of large language models (LLMs) has forced the field to evolve again. Language models introduce new deployment problems — enormous model sizes, autoregressive inference, prompt-dependent behavior, and outputs that cannot be checked against a fixed ground truth — that classical MLOps tooling was never designed to handle. This paper presents a systematic survey of MLOps and its language-model successor, LLMOps. It traces the principles of MLOps from continuous integration and delivery for machine learning through model registries, data validation, and monitoring; it presents a taxonomy of production machine-learning systems organized by deployment pattern, serving architecture, and operational risk; it examines how LLM inference is actually served, from memory-efficient attention and paged key-value caches to quantization and speculative decoding; it analyzes monitoring and evaluation for systems whose outputs are free text, including drift detection, LLM-as-a-judge evaluation, retrieval-augmented generation evaluation, and tracing; it surveys the operational layer of LLM applications — prompt versioning, retrieval pipelines, and autonomous agents; and it discusses cost, energy, safety, governance, and compliance as first-class operational concerns. The survey concludes that LLMOps is not a replacement for MLOps but an extension of it: the classical disciplines of versioning, testing, and monitoring remain necessary, and the new work is to extend them to artifacts — prompts, retrieved context, and agent trajectories — that have no counterpart in classical machine learning.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24268760
Primary Topic
Software System Performance and Reliability
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

MLOps and LLMOps: A Systematic Survey of Deploying, Monitoring, and Operating Large Language Models in Production

Aref Mahmoudpour
2 links
Zenodo (CERN European Organization for Nuclear Research)
Software System Performance and Reliability
preprint

MLOps and LLMOps: A Systematic Survey of Deploying, Monitoring, and Operating Large Language Models in Production

Aref Mahmoudpour
preprint en

Abstract

Machine learning operations (MLOps) emerged to close the gap between a model that works in a notebook and a system that works in production, and the arrival of large language models (LLMs) has forced the field to evolve again. Language models introduce new deployment problems — enormous model sizes, autoregressive inference, prompt-dependent behavior, and outputs that cannot be checked against a fixed ground truth — that classical MLOps tooling was never designed to handle. This paper presents a systematic survey of MLOps and its language-model successor, LLMOps. It traces the principles of MLOps from continuous integration and delivery for machine learning through model registries, data validation, and monitoring; it presents a taxonomy of production machine-learning systems organized by deployment pattern, serving architecture, and operational risk; it examines how LLM inference is actually served, from memory-efficient attention and paged key-value caches to quantization and speculative decoding; it analyzes monitoring and evaluation for systems whose outputs are free text, including drift detection, LLM-as-a-judge evaluation, retrieval-augmented generation evaluation, and tracing; it surveys the operational layer of LLM applications — prompt versioning, retrieval pipelines, and autonomous agents; and it discusses cost, energy, safety, governance, and compliance as first-class operational concerns. The survey concludes that LLMOps is not a replacement for MLOps but an extension of it: the classical disciplines of versioning, testing, and monitoring remain necessary, and the new work is to extend them to artifacts — prompts, retrieved context, and agent trajectories — that have no counterpart in classical machine learning.

Zenodo (CERN European Organization for Nuclear Research)
Islamic Azad University, Tehran (IR)
Software System Performance and Reliability
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon