Glossary

This glossary centralizes key terms used across full stack engineering, ML, DevOps, and collaboration topics. Use it as a quick reference while reading other pages.

How to Use - Skim alphabetically or search in-page. - Terms intentionally concise; for deeper context, follow cross-links. - Bold indicates especially foundational terms.

---

A

A/B Test – Controlled experiment comparing two (or more) variants to measure impact on a metric.
Access Token – Credential granting a client permission to call protected APIs (limited scope & lifetime).
API Contract – Versioned specification of endpoints, inputs, outputs, and errors (e.g., OpenAPI).
Artifact – Versioned build output (container image, model weights, compiled bundle).
Authentication (AuthN) – Verifying identity of a user or system.
Authorization (AuthZ) – Determining what an authenticated entity is allowed to do.

B

Baseline (Model) – Simple heuristic or naive predictor used for initial comparison.
Batch Processing – Running computations on large groups of items offline (non-interactive).
Blue/Green Deployment – Two production environments: one live (blue), one prepared (green); switch traffic after validation. Burn Rate (Error Budget) – Speed at which allowable failures are consumed relative to SLO.

C

Caching – Storing computed data for faster subsequent access (memory, CDN, database).
Canary Deployment – Gradually shifting a small portion of traffic to a new version to reduce release risk.
CI/CD – Continuous Integration / Continuous Delivery: automated build, test, and deploy pipelines.
Circuit Breaker – Pattern that halts calls to a failing dependency to prevent cascading failure.
Cloud Region – Geographic data center grouping in a cloud provider.
Confusion Matrix – Table summarizing classification predictions vs actual classes.
Container – Isolated runtime packaging code + dependencies on a shared kernel.
Continuous Monitoring – Ongoing collection of metrics & logs to assess health & performance in production.

D

Data Drift – Change in input feature distribution over time relative to training data.
Data Leakage – When information from outside the training scope (future, test set) contaminates training and inflates metrics.
Deployment – Releasing code/model to a production environment where users or systems rely on it.
DevOps – Culture & practices combining development + operations for faster, safer delivery.
Drift (Concept) – Evolution of the relationship between inputs and target variable.
Dry Run – Executing logic in a mode that logs actions instead of performing them (safety check).

E

Embedding – Dense vector representation capturing semantic or relational meaning (text, images, users).
Environment (Runtime) – Isolated configuration grouping environment variables, dependencies, and infrastructure settings.
Error Budget – Allowed amount of unreliability before halting feature releases.
Experiment Tracking – Logging parameters, code versions, and metrics for ML iterations.

F

Feature (ML) – Input variable used for model prediction.
Feature Engineering – Creating transformations/aggregations of raw data to improve predictive performance.
Feature Flag – Toggle controlling exposure of code paths or features in production.
Feature Store – System centralizing feature definitions for training & online inference consistency.
Fork (Git) – Personal copy of a repository, often used for open-source contributions.

G

GitHub Flow – Lightweight branching model using main plus short-lived feature branches merged via pull requests.
GPU – Highly parallel processor for matrix-heavy workloads (deep learning).
Guardrail Metric – Secondary metric ensuring primary optimization doesn’t harm critical constraints (e.g., precision while optimizing recall).

H

Health Check – Endpoint or probe verifying application readiness or liveness.
Horizontal Scaling – Adding more instances to handle load (vs vertical scaling resources per instance).
HTTP Status Code – Standardized numeric response indicator (200 OK, 404 Not Found, etc.).

I

Idempotency – Repeating an operation yields the same result (critical for safe retries).
Infrastructure as Code (IaC) – Managing infrastructure using version-controlled declarative configuration.
Input Validation – Ensuring incoming data matches expected shape, type, range before processing.
Integration Test – Validates multiple components working together (API ↔ DB).
Isolation (Container) – Separation of filesystem, process space, and network context per container.

J

JWT (JSON Web Token) – Signed token encoding claims (subject, expiry, scopes) for stateless auth.

K

K-Fold Cross Validation – Evaluation technique rotating validation subsets to estimate generalization.
Key Rotation – Replacing cryptographic keys/secrets periodically to reduce exposure risk.

L

Latency (P95/P99) – Time for a request; percentiles highlight tail performance.
Least Privilege – Grant only needed access rights (principle of minimal access).
Load Shedding – Proactively rejecting work to preserve service for existing traffic under overload.
Logging (Structured) – Emitting machine-parseable key/value event records.

M

MLOps – Engineering discipline unifying ML development & production operations.
Model Card – Standardized documentation summarizing a model’s purpose, performance, data, limitations.
Monitoring (Model) – Tracking performance + drift metrics in production.
Multicollinearity – Correlation among predictors inflating variance of coefficient estimates.

N

Namespace (Kubernetes) – Logical cluster partition scoping resources & access.
Normalization (Feature) – Scaling data (e.g., z-score) to stabilize training.
N+1 Query – Inefficient pattern of issuing one query per item instead of a set-based query.

O

Observability – Ability to understand internal state from external outputs (logs, metrics, traces).
Overfitting – Model memorizes noise; performance drops on unseen data.
OpenAPI – Specification format for describing RESTful APIs.

P

Pagination (API) – Mechanism to segment large result sets (offset or cursor).
PCA (Principal Component Analysis) – Dimensionality reduction technique.
Pipeline (ML) – Sequence of reproducible steps from raw data → deployed model.
Precision / Recall – Metrics balancing false positives vs false negatives in classification.
Provisioning – Allocating infrastructure resources (compute, storage, networking).

Q

Queue (Message) – Buffer decoupling producers and consumers for asynchronous processing.

R

RBAC – Role-Based Access Control; permissions assigned to roles, roles to users.
Regression (Statistical) – Predicting a continuous variable.
Reliability – Probability a system performs correctly over time.
Rollback – Reverting to a prior stable deployment state.

S

Scaling (Vertical) – Increasing resources (CPU/RAM) of a single instance.
Schema (Data) – Structural definition (columns, types, constraints).
Session (Auth) – Server-maintained state linking a user to an authenticated identity.
Shadow Deployment – New version receives mirrored traffic (responses ignored) for evaluation.
Skew (Train/Serve) – Mismatch in feature computation between training and production.
SLO (Service Level Objective) – Target reliability/latency measure.
Span (Tracing) – Timed operation with metadata in a distributed trace.
Streaming – Processing data continuously as it arrives.

T

Tail Latency – High percentile latency (P95, P99) that impacts user perception disproportionately.
Token (Auth) – Opaque or signed credential used for authorization decisions.
Trace ID – Unique identifier correlating logs, metrics, spans for one request.
Transformation (Feature) – Operation converting raw attribute into model-ready value.

U

Underfitting – Model too simple to capture data patterns.
Unit Test – Verifies correctness of a small piece of code in isolation.
UTC – Coordinated Universal Time; store timestamps in UTC to avoid timezone confusion.

V

Version Pinning – Freezing dependency versions to guarantee reproducibility.
Vertical Slice – Thin end-to-end increment demonstrating a feature across layers.
Virtual Environment – Isolated Python environment with its own dependencies.

W

Warm Start (Model Serving) – Preloading model weights and performing initial inference to prime caches.
WebSocket – Bi-directional persistent communication protocol over a single TCP connection.
Workload – Type and volume of work a system processes (read-heavy, write-heavy, CPU-bound).

X

XSS (Cross-Site Scripting) – Injection attack executing malicious scripts in users' browsers.

Y

YAML – Human-readable configuration format used for pipelines, configs, and manifests.

Z

Zero Downtime Deploy – Deployment pattern avoiding user-visible interruptions (rolling, blue/green, canary).