Certification exam glossary

Definitions of terms that recur across the exams covered on this site, written the way the exams actually use them rather than the way a textbook would.

AI and generative AI

Foundation model. A large model trained on broad data and adapted to many tasks, rather than built for one. Its generality is exactly what makes its behaviour hard to enumerate and test.

Embedding. Text represented as a vector of numbers, positioned so that similar meanings sit close together. This is what makes semantic similarity search possible. Not compression, not encryption.

Context window. How much text a model can consider in one request, measured in tokens, covering both what you send and what it generates. It is not memory between conversations.

Token. The unit models read and bill in, roughly a word fragment. Both input and output are billed, which is why stuffing a whole knowledge base into every prompt gets expensive fast.

Hallucination. Fluent output that is unsupported or fabricated. Distinct from drift, which is degradation over time.

RAG (retrieval-augmented generation). Retrieving relevant documents at query time and passing them to the model, so answers reflect current content without retraining. The standard answer whenever source material changes regularly.

Chunking. Splitting documents before indexing. Chunks that are too large dilute the embedding, so retrieval matches the broad topic instead of the specific answer — the most common cause of “related but useless” results.

Groundedness. Whether an answer is actually supported by the retrieved sources. Different from relevance, which is whether it addresses the question. High groundedness with low relevance means faithful answers to the wrong question.

Zero-shot / few-shot prompting. No examples in the prompt, versus a handful of worked examples in the prompt. Neither changes the model; fine-tuning does.

Prompt injection. Instructions hidden in content the model processes, causing it to ignore its own rules. Indirect injection arrives through retrieved documents or web pages. The real mitigation is isolating retrieved content from the instruction channel and treating it as data, not filtering for suspicious words.

Agent. A system that chooses among tools and takes several dependent steps, rather than answering in one shot. Its blast radius is whatever its tool permissions allow, which is why least-privilege scoping matters more than prompt wording.

Managed identity / workload identity. Platform-issued identity so a workload authenticates without a stored secret. The answer to almost any “credentials must not be stored in code” question.

Machine learning

Regression / classification / clustering. Predicting a number, predicting a category from labelled data, and finding groups in unlabelled data.

Overfitting. Excellent training performance, poor performance on unseen data. The model memorised rather than generalised.

Target leakage. A feature that encodes the answer, or that is unavailable at inference time. Symptom: superb validation scores that collapse in production.

Training-serving skew. Features computed differently in training and serving, so predictions diverge from what validation promised. Fixed by one shared feature definition, typically a feature store.

Data drift / concept drift. The input distribution shifting, versus the relationship between inputs and the target shifting. The second is only visible once labels arrive.

Precision / recall. Of what you flagged, how much was right, versus of what was there, how much you caught. When missing a positive case is costly, recall is the metric that matters.

Class imbalance. A rare positive class, where accuracy becomes meaningless — a model predicting “no” every time can score 99%.

Data engineering

Delta Lake. ACID transactions, schema enforcement and time travel layered over files in object storage.

Medallion architecture. Bronze for raw as-received data, silver for cleansed, gold for business-level aggregates.

Idempotent write. A write that produces the same result if it runs twice, normally via a merge keyed on an identifier. What makes retries safe.

Schema evolution. Allowing new columns to appear without breaking the write.

Small file problem. Many tiny files making listing and opening dominate query time. Scales far worse than data volume, and is a frequent cause of jobs that suddenly take hours.

Data skew. Disproportionate data in a few partitions, so a handful of tasks dominate a stage while the rest finish quickly.

Shuffle. Redistributing data across the cluster, which creates a stage boundary and is usually the most expensive part of a job.

Broadcast join. Sending a small table to every executor to avoid the shuffle entirely.

Lineage. The recorded relationship between tables and jobs, which answers “what breaks if I change this” without reading code.

Networking

OSI layers. Layer 2 forwards on MAC addresses, layer 3 on IP addresses. Most exam questions hinge on which of those two is in play.

VLAN and trunk. A logical network segment, and a link carrying tagged traffic for several of them.

Duplex mismatch. Two ends disagreeing on duplex, producing late collisions and poor throughput on a link that is otherwise up.

RTO / RPO. Maximum acceptable time to restore service, versus maximum acceptable data loss.

Security and governance

Least privilege. Granting only the access required. Recurs across every security exam here, and is the correct answer more often than any other single principle.

Private endpoint. A service reachable on a private address with public access disabled. Not the same as a firewall rule allowing your office IP, which still exposes a public endpoint.

Confidential computing. Hardware-backed protection of data while it is being processed, not merely at rest or in transit.

Model extraction. Reconstructing a model’s behaviour through many crafted queries. Membership inference instead determines whether a specific record was in the training data.

Data poisoning. Corrupting training data to influence model behaviour, often through a supply chain you do not control.

Segregation of duties. The person who builds something does not approve and deploy it unchecked.

Operating effectiveness. Evidence that a control actually ran over a period, as opposed to design effectiveness, which is evidence that it exists. Auditors care about the first; policy documents only demonstrate the second.

Terms here are used as the exams use them. For how each exam weights these areas, see the exam sections, and for choosing between exams the certification paths.