Drift
Drift is a change or divergence over time in a system’s data, state, behavior, or performance relative to a reference or intended condition.
Machine Learning Platform Engineering
Data drift is a change over time in the statistical distribution or characteristics of input data relative to the data used to develop a model.
- How it is used
- Flags changes in production data that may threaten prediction quality and require investigation or maintenance.
- How deep it goes
- Develops report generation, dashboarding, and batch and API integration, while leaving microbatch size dependent on the application.
- Examples
- Uses income-classifier reports, altered image brightness, and changed movie ratings and item factors.
- What you can do after
- Apply it via Evidently or deepchecks to compare reference and current data, and integrate checks into batch or API monitoring.
Read how Machine Learning Platform Engineering explains drift →
Machine Learning System Design
Model drift is a change over time in a machine learning model’s behavior or performance caused by changes in its inputs, relationships, or operating environment.
- How it is used
- Guides monitoring and maintenance of production models as their data and operating conditions change.
- How deep it goes
- Distinguishes drift types and outlines monitoring, diagnosis, retraining, and fallback choices without exhaustively covering every response scenario.
- Examples
- Illustrates drift through a new advertising audience, popularity-driven recommendations, and retail demand prediction fallbacks.
- What you can do after
- Diagnose/debug drift by checking quality, comparing performance and distributions, and selecting retraining, rebuilding, or fallback responses.
Statistics Every Programmer Needs
Nonstationary behavior is temporal behavior whose statistical properties, such as level or variability, change over time.
- How it is used
- Flags changing time-series behavior that must be assessed before fitting a stationary-data forecasting model.
- How deep it goes
- Works through chronological splitting, visual inspection, an ADF test, first-order differencing, and post-transformation checks without establishing a universal remedy.
- Examples
- Uses Apple closing prices, hand-calculated daily differences, Python transformations, lag plots, and ADF p-values.
- What you can do after
- Diagnose/debug with it by checking a training series for nonstationarity and reassessing it after differencing.
Effective Platform Engineering
Configuration drift is the divergence between an infrastructure system's actual state and its declared or desired configuration.
- How it is used
- Uses drift detection to keep deployed infrastructure aligned with intended configuration after releases and between pipeline runs.
- How deep it goes
- Explains manual-change causes and recurring configuration tests, but stops at detection rather than specifying automated repair.
- Examples
- Uses a console change to illustrate drift and an Awspec IAM role check to illustrate configuration testing.
- What you can do after
- Diagnose/debug with it by running recurring checks that compare deployed settings with required settings and flag unexpected changes.
Automating API Delivery
Configuration drift is a divergence between a system's desired configuration and its actual deployed configuration.
- How it is used
- Identifies mismatches that API gateway deployment agents must detect and reconcile against repository-defined configuration.
- How deep it goes
- Explains causes, comparison, alerts, and correction of configuration drift, without specifying an agent implementation or command sequence.
- Examples
- Uses a direct API gateway UI change and a live API response missing a mandatory documented field.
- What you can do after
- Operate/configure API gateway changes through Git and use reconciliation to detect and correct deployed configuration drift.
Terraform in Depth: Infrastructure as Code with Terraform and OpenTofu
Drift is a discrepancy between the infrastructure or files Terraform expects and the resources or files actually present.
- How it is used
- Identifies mismatches so Terraform plans and deployment workflows can decide whether to restore configuration, accept a change, or ignore an attribute.
- How deep it goes
- Explains refresh-based detection, cause investigation, and responses to expected changes; does not detail the four referenced error categories.
- Examples
- Uses externally added tags, autoscaling counts, database upgrades, regenerated values, and local files that disappear between runs.
- What you can do after
- Diagnose/debug with Terraform plans and refresh-only mode, then investigate the cause before choosing reconciliation, acceptance, or ignore_changes.
Read how Terraform in Depth: Infrastructure as Code with Terraform and OpenTofu explains drift →
LLM Customization and Fine-Tuning
A change over time in inputs, their relationship to correct outputs, or model behavior that can degrade performance.
- How it is used
- Monitors production assistants for erosion in domain accuracy, response format, and safety as their environment or underlying model changes.
- How deep it goes
- Distinguishes four drift forms and outlines detection, threshold calibration, retraining gates, and rollback procedures.
- Examples
- Uses an IT support assistant facing Kubernetes questions, an SSO migration, and informal user shorthand.
- What you can do after
- Evaluate drift by distinguishing changes in inputs, correct answers, phrasing, and model behavior.
Platform Engineering on Kubernetes
Configuration drift is the divergence between the declared configuration of a system and the configuration actually running in that system.
- How it is used
- Controls Kubernetes environment changes by making versioned Git configuration the desired state.
- How deep it goes
- Explains detection, reconciliation, validation, and rollback, but does not detail the internal implementation of reconciliation.
- Examples
- Uses Git commits, Kubernetes clusters, Helm configuration, kubectl application, and Argo CD reconciliation to illustrate drift.
- What you can do after
- Diagnose drift by comparing live workloads with Git, reconcile differences, validate deployments, and roll back through earlier commits.
Read how Platform Engineering on Kubernetes explains drift →
Delivering Machine Learning Projects: From design to deployment
Model drift is a change in the data, relationships, or operating conditions underlying a machine learning model that can reduce its suitability or performance over time.
- How it is used
- Treats drift as a post-deployment signal to review model suitability and consider repair or retraining.
- How deep it goes
- Distinguishes weak original evaluation from changing relationships and noise, then outlines retraining inputs and coordination.
- Examples
- Uses changing demand for smart trousers after a shift to working from home, and changing data noise.
- What you can do after
- Diagnose drift by reviewing prior testing, then plan retraining with updated data, a test regime, and stakeholder buy-in.
Read how Delivering Machine Learning Projects: From design to deployment explains drift →
AI Model Evaluation
A change over time in a model's data distribution, behavior, or performance relative to its earlier state.
- How it is used
- Uses evaluation to detect when model or judge behavior changes over time and threatens confidence in production scores.
- How deep it goes
- Distinguishes judge drift from product changes and points to canary or offline evaluation, without detailing a full monitoring procedure.
- Examples
- Uses historical support tickets, agent labels, and replayed outputs, alongside hosted-judge behavior changes behind stable API identifiers.
- What you can do after
- Diagnose whether evaluation score changes reflect product behavior or a drifting hosted judge, and assess internal models offline over time.
LLMs in Production
Data drift is a change over time in the statistical properties or distribution of data presented to a machine-learning system.
- How it is used
- Uses monitoring to catch data changes associated with poor ML or LLM performance and guide adjustment or retraining.
- How deep it goes
- Lists missing values, uniqueness, and standard-deviation shifts as signals, without detailing a drift-detection algorithm.
- Examples
- Uses fake reviews manipulating recommendations and highlights the difficulty of assessing drift in text data.
- What you can do after
- Diagnose silent model degradation by monitoring data and respond by adjusting the model or retraining it.
Digital Twins in Action
Data drift is a change in the distribution of input data relative to the reference data used to train a model.
- How it is used
- Monitors changing input distributions in deployed digital twins to support model lifecycle management.
- How deep it goes
- Explains distribution comparison, operator review, and retraining, without detailing detection methods or thresholds.
- Examples
- Uses solar production changes and a power-consumption comparison between reference and current readings.
- What you can do after
- Apply it via Evidently to compare a single input feature and generate a drift report.
LLM Evaluation and Alignment, The Foundational Ideas
Distribution shift is a change between the data or behavior distribution on which a model was trained and the distribution encountered during use or optimization.
- How it is used
- Monitors production translation quality and identifies calibration risks when best-of-N shifts sampled policy outputs.
- How deep it goes
- Distinguishes input-data changes from best-of-N policy shifts, but gives no general detection test or threshold.
- Examples
- Uses production translation metric drops and best-of-N candidate responses as drift-related examples.
- What you can do after
- Diagnose possible production drift by sampling expensive metrics or tracking low-cost proxy scores, then investigate sudden drops.
Read how LLM Evaluation and Alignment, The Foundational Ideas explains drift →
Reinforcement Learning from Human Feedback
Policy drift is the difference between a policy that generated data and the policy being updated after one or more optimization steps.
- How it is used
- Tracks policy mismatch during reinforcement-learning updates and movement away from assistant-like activations during conversations.
- How deep it goes
- Defines both forms, explains importance-weight correction and activation capping, and discusses bias-variance trade-offs and threshold calibration.
- Examples
- Uses multi-step policy updates and a simulated distress conversation in which persona drift precedes harmful advice.
- What you can do after
- Diagnose policy and implementation mismatches in training, or identify and mitigate activation drift during conversations.
Read how Reinforcement Learning from Human Feedback explains drift →
Kafka for Architects
Configuration drift is the divergence between a system's actual configuration and its declared desired configuration.
- How it is used
- Detects discrepancies between Kafka cluster configuration and its Git-declared desired state.
- How deep it goes
- Defines drift and describes detection, warnings, and corrective suggestions, but notes that authorized direct changes cannot be prevented.
- Examples
- Uses an authorized user creating a topic directly as an example of a change that can cause drift.
- What you can do after
- Operate Kafka configuration through Git-driven comparison, reconciliation, and periodic drift validation.
Bootstrapping Microservices, Second Edition: With Docker, Kubernetes, GitHub Actions, and Terraform
Configuration drift is the divergence between the recorded or declared configuration of infrastructure and its actual state.
- How it is used
- Uses Terraform state refresh and configuration comparison to detect and reconcile infrastructure changed outside declared code.
- How deep it goes
- Explains ownership, state refresh, and reconciliation while stopping short of a detailed remediation workflow.
- Examples
- Uses manually tweaked infrastructure and an Azure resource group recorded in terraform.tfstate as concrete examples.
- What you can do after
- Diagnose drift by inspecting Terraform state and understanding how refresh compares live resources with declared configuration.
AI Governance
A change in a model’s behavior over time caused by model updates or changes in its inputs.
- How it is used
- Tracks post-release behavior changes as governance risks that can prompt policy, tooling, rule, or configuration updates.
- How deep it goes
- Defines drift by its causes and connects operational findings to governance updates without detailing detection methods.
- What you can do after
- Operate a feedback loop that turns incidents, audit findings, and user feedback into policy, tooling, rule, or configuration updates.
Financial AI in Practice
A change over time in a model’s predictive behavior or the relationships on which its performance depends.
- How it is used
- Uses stability monitoring to detect behavioral changes after LLM provider updates.
- How deep it goes
- Mentions drift detection but does not detail a monitoring method.
- Examples
- Identifies an LLM provider updating its system as a potential trigger.
- What you can do after
- Diagnose behavioral drift by monitoring model stability after provider updates.
Also covered in
- Infrastructure as Code, Patterns and Practices: With examples in Python and Terraform
- 100 SQL Server Mistakes and How to Avoid Them
- Secure APIs
- Experimentation for Engineers
- Acing the CCNA Exam: Volume 2 Advanced Networking and Security
- Authorization in Action
- Designing Deep Learning Systems: A guide for software engineers
- Data Without Labels
- Domain-Specific Small Language Models
- How Large Language Models Work
- Software Architecture Decision-MakingIn early access (MEAP)
- Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI
What to know first
- language modelA language model is a computational model that represents patterns in linguistic sequences to assign probabilities, predict, generate, or evaluate language.
- productionProduction is the live operational setting in which a deployed software system serves its intended users or processes real workloads.
- accuracyAccuracy is the degree to which a result agrees with a specified truth, reference, or criterion, commonly measured as the proportion of evaluated predictions that are correct.
- fine-tuningFine-tuning is the continued training of a pretrained model on targeted data or objectives to adapt its parameters or behavior to a task, domain, or use.
Related concepts
- concept driftincludes · Concept drift is a change over time in the data-generating process, including the distribution of inputs or the relationship between inputs and target outcomes.
- monitoringaddressed by · Monitoring is the ongoing observation and measurement of a system or process to assess its behavior or condition and detect changes, failures, or other problems.
- infrastructure as codeaddressed by · Infrastructure as code is the practice of defining, provisioning, and managing computing infrastructure through machine-readable code or configuration rather than manual procedures.
- reconciliationaddressed by · Reconciliation is the process of comparing related states, records, or outputs to identify differences and, when needed, resolve them or bring them into agreement.
- AI agentaddressed by · An AI agent is a computational system that uses artificial intelligence to interpret information and select or perform actions toward goals while interacting with an environment.
Questions about drift
What is drift?
Drift is a change or divergence over time in a system’s data, state, behavior, or performance relative to a reference or intended condition.
Which Manning books explain drift?
Machine Learning Platform Engineering, Machine Learning System Design, Statistics Every Programmer Needs, Effective Platform Engineering, Automating API Delivery, Terraform in Depth: Infrastructure as Code with Terraform and OpenTofu, LLM Customization and Fine-Tuning, Platform Engineering on Kubernetes and 10 more.
What should I know before learning drift?
Manning books that teach drift usually introduce language model, production, accuracy and fine-tuning first.
What kinds of drift are there?
Kinds of drift include concept drift.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.

















