Intel MLOps Professional key terms

A centralized system that stores curated feature data and serves it consistently to both training pipelines and online inference, reducing skew between how features are computed in development and in production.

The gap between how features or preprocessing are computed during model training and how they are computed at inference time, which silently degrades production accuracy even when offline metrics look strong.

Building training datasets using only data that was actually available at each historical prediction timestamp, preventing future information from leaking into features used to train the model.

A modeling error where information unavailable at real prediction time, such as target-derived columns or post-event data, contaminates training and inflates measured performance relative to what deployment will achieve.

Automated checks on incoming data for schema conformance, type correctness, allowed ranges, and distribution expectations, run before data enters training or scoring so bad inputs fail loudly rather than silently.

Declaring and verifying the expected column names, data types, and nullability of a dataset at ingestion, so upstream changes break the pipeline at a known point instead of corrupting downstream models.

The automated sequence of transforms that turns raw source data into model-ready features, typically orchestrated as scheduled or event-triggered jobs with versioned code and reproducible outputs.

Recording each training run's parameters, code version, data snapshot, metrics, and artifacts in a searchable store, enabling comparison, reproduction, and audit of how a model was produced.

A versioned catalog of trained model artifacts with metadata, lineage, and lifecycle stages such as staging or production, acting as the controlled handoff point between training and deployment.

The recorded chain linking a deployed model back to its training code commit, hyperparameters, dataset version, and parent models, used for debugging, audit, and impact analysis when something changes.

The ability to recreate a model or metric exactly or within stated tolerance by pinning code, data, environment, seeds, and hardware configuration, which is essential for debugging and regulatory defensibility.

Systematically searching configurations such as learning rate, batch size, and regularization strength, often with early-stopping schedulers, to improve a chosen metric without changing the underlying algorithm family.

Reducing the numeric precision of model weights and sometimes activations, for example from 32-bit floats to 8-bit integers, to shrink memory footprint and speed inference at some cost in accuracy.

Converting an already trained model to lower precision using calibration data to choose scaling factors, requiring no retraining and typically the fastest path to smaller, faster inference.

A small, representative sample of real input data used during quantization to estimate activation ranges so that scaling factors preserve accuracy on the workloads the model will actually see.

Training or inference that combines lower-precision arithmetic, such as 16-bit floats or bfloat16, with higher-precision accumulation or master weights to gain throughput while keeping numerical stability.

A 16-bit floating-point format with the same exponent range as 32-bit floats but reduced mantissa precision, widely used on modern accelerators to speed training and inference with minimal tuning.

The family of techniques, including pruning, quantization, distillation, and low-rank factorization, that reduce model size or compute cost so a model meets latency, memory, or cost constraints.

Training a smaller student model to imitate the outputs or intermediate representations of a larger teacher model, transferring much of the teacher's accuracy into a cheaper artifact for deployment.

Removing weights, channels, or layers judged least important, often followed by fine-tuning, to reduce model size and compute while limiting the drop in predictive quality.

Tuning a trained model for serving through graph-level fusion, operator replacement, precision reduction, batching, and runtime selection, measured against latency and throughput targets on the target hardware.

Grouping multiple inference requests into one execution pass to improve accelerator utilization and throughput, traded against added per-request latency while the batch fills.

A stated service-level objective for response time, such as a percentile target under expected load, which constrains model size, precision, batching strategy, and infrastructure sizing choices.

The volume of inference requests or training samples a system processes per unit time, a primary capacity metric that interacts with latency and hardware utilization.

Exposing a trained model behind a stable interface, such as an HTTP or gRPC endpoint, with request handling, preprocessing, and response formatting managed by a serving runtime.

Routing a small share of live traffic to a new model version while monitoring error and quality signals, then gradually expanding or rolling back based on observed behavior.

Running a candidate model on mirrored production traffic without returning its outputs to users, allowing comparison against the incumbent before any real traffic is shifted.

Restoring a previously validated model or configuration after a release shows degraded behavior, which requires versioned artifacts, retained prior deployments, and a tested reversal procedure.

Automatically adjusting serving replicas or compute resources in response to load or queue depth, balancing latency objectives against infrastructure cost and cold-start effects.

A system that schedules and sequences pipeline tasks, handles dependencies, retries, and resource placement, and provides run history for data and model workflows.

Degradation in model performance over time as real-world relationships or input populations change, detected by monitoring prediction quality or proxies when ground truth arrives late.

A shift in the distribution of model inputs relative to the training reference, which can precede accuracy loss and is often detectable earlier than outcome-based performance metrics.

A change in the true relationship between inputs and target, so the same input now warrants a different prediction; it cannot be fixed by reweighting inputs alone and usually requires retraining.

The ability to understand a deployed system's internal state from its emitted logs, metrics, and traces, enabling diagnosis of failures and performance regressions without direct access to the host.

A configured boundary on a monitored signal that triggers notification when crossed, chosen to catch real incidents while limiting noise from normal variation.

Automated stages that build, test, and release code or models on every change, providing repeatable quality gates and traceable promotion from development to production.

A packaged, immutable filesystem and configuration that includes an application and its dependencies, giving consistent runtime behavior across environments and a versionable deployment artifact.

A short structured document describing a model's intended use, training data, evaluation results, limitations, and ethical considerations, supporting informed review by downstream users and reviewers.

Evaluating a model's error rates and outcomes across protected or sensitive subgroups to identify disparate performance, then documenting and mitigating material disparities before release.

Detecting and removing or masking personally identifiable information in data used for training, logging, or monitoring, reducing privacy exposure and supporting data-minimization obligations.