Intel MLOps Professional flashcards

33 cards. Read the term and the definition here. The deck at the top of this page is optional.

MLOps maturity level 1

A level where the ML pipeline is automated end to end and supports continuous training on new data, but full CI/CD of the pipeline components themselves is not yet in place.

CRISP-DM iteration

The data preparation and modeling phases loop repeatedly as understanding improves; deployment is reached only after evaluation against business success criteria, not after a single linear pass.

Latency vs throughput scoping

Real-time serving optimizes single-request tail latency, batch scoring optimizes records per second, and streaming optimizes window completion time; the choice constrains architecture before model selection.

Point-in-time correctness

Features used for training must reflect only data available at the prediction timestamp, preventing future information from leaking into the model through later-arriving values.

Train-serve skew

A mismatch between feature computation during training and during serving, usually caused by duplicated transformation code; sharing one transformation implementation across offline and online stores prevents it.

Temporal data split

Training on earlier periods and validating on later periods; random splits leak future information when rows are autocorrelated over time, inflating offline estimates.

Feature versioning rule

Any change to a feature's definition, unit, or aggregation window requires a new version name, because overwriting in place invalidates historical training datasets built on the old definition.

Data quality gate

An automated pipeline check on schema conformance, null rate, value range, cardinality, or freshness that fails the run loudly instead of silently dropping or imputing bad rows.

Reproducibility triad

A result is reproducible only when the code commit, the data snapshot version, and the full hyperparameter plus environment specification are all recorded together.

Model registry promotion

Immutable model versions move from staging to production only after passing stored evaluation gates, with lineage links back to the training run and data snapshot.

PR-AUC vs ROC-AUC

Use precision-recall AUC when positives are rare because ROC-AUC stays optimistic under heavy class imbalance; use ROC-AUC when classes are roughly balanced.

Calibration metric

A reliability curve or Brier score measures whether predicted probabilities match observed frequencies; accuracy alone does not, which matters when probabilities drive decisions.

Grouped cross-validation

A split that keeps all rows of an entity in the same fold, used when the same entity repeats across rows so plain k-fold would leak entity-specific information.

oneDNN role

Intel's deep learning library providing optimized CPU primitives used as the backend by frameworks such as TensorFlow and PyTorch on Intel hardware; oneDNN Graph fuses operator patterns.

Intel Neural Compressor

A toolkit for post-training quantization, quantization-aware training, and accuracy-aware tuning that searches configurations within a user-specified accuracy tolerance.

bfloat16 vs FP16

bfloat16 keeps the FP32 exponent range so it usually needs no loss scaling, while FP16 has narrower range and typically requires loss scaling to avoid gradient underflow.

OpenVINO Intermediate Representation

A converted model format produced by Model Optimizer from ONNX, TensorFlow, PaddlePaddle, or PyTorch graphs, enabling runtime layer fusion and layout optimization on Intel hardware.

INT8 quantization tradeoff

Reducing weights and activations to 8-bit integers typically yields large CPU speedups with modest accuracy loss; validate on a representative calibration set before deployment.

Roofline memory bound

When arithmetic intensity is low, performance is limited by memory bandwidth, so reducing bytes moved via lower precision helps more than adding compute capacity.

KServe InferenceService

A Kubernetes custom resource defining predictor, transformer, and explainer components for a model, supporting standard request protocols and scale-to-zero behavior.

Shadow deployment

A release pattern that mirrors production traffic to a new model version without returning its predictions to users, allowing comparison of outputs and latency before any traffic shift.

Readiness versus liveness

Liveness reports the process is running; readiness reports the model is loaded and dependencies are reachable, so orchestrators avoid routing traffic to warming replicas.

Idempotent orchestration task

A pipeline task that produces the same result when retried, typically by writing with merge or upsert semantics keyed on entity and event time, making retries safe.

Canary promotion criteria

A canary is promoted only when error rate, tail latency, and the business metric on the canary slice match or beat the current production baseline within predefined tolerances.

Concept drift

A change in the relationship between inputs and target, P(Y|X), which degrades model performance even when the input distribution itself remains unchanged.

Population Stability Index

PSI equals the sum over bins of (actual minus expected share) times the natural log of their ratio; values above roughly 0.25 indicate a significant distribution shift.

Delayed label monitoring

When ground-truth labels arrive late, monitor proxies such as prediction distribution, confidence, and input drift, then run scheduled backfill evaluation once labels land.

Four golden signals

Latency, traffic, errors, and saturation are the baseline service health signals; model monitoring adds prediction distribution, unknown rate, and missing-feature rate.

ML CI model behavior test

An automated test asserting expected model properties, such as invariance to irrelevant perturbations, directional response to key features, and a minimum performance threshold on a fixed holdout.

Model card

A governance document stating intended use, training data provenance, evaluation metrics across subgroups, known limitations, and out-of-scope uses for a released model.

Environment promotion parity

Promoting the identical container image across dev, staging, and production, changing only configuration and model version, eliminates environment-specific behavior differences.

Membership inference attack

An attack that determines whether a specific record was in the training set by exploiting overconfident, overfit model outputs; mitigation includes regularization, differential privacy, and limiting confidence exposure.

Differential privacy epsilon

A parameter bounding privacy loss for a query or training procedure; smaller epsilon gives stronger privacy with more utility loss, and budgets compose across repeated releases.