| Term | Definition |
|---|---|
| ASR | Architecturally Significant Requirement — a requirement with a measurable impact on architecture, whether functional or non-functional. Ultimately measured by high cost of change |
| Quality attribute | A measurable or testable property of a system indicating how well it satisfies stakeholder needs |
| NFR | Non-functional requirement — criteria used to evaluate the whole system rather than a specific behaviour |
| Constraint | A requirement that removes design freedom; almost always architecturally significant |
| Tactic | A design decision to achieve a quality attribute response, with no trade-offs considered internally. "Atoms" |
| Pattern | A design solution for a concrete context and problem, with trade-offs built in. "Molecules" |
| Style | The highest level of granularity — layers, high-level modules, their interaction and relations |
| Structure | A set of elements and their organisation |
| View | A representation of a structure, documented per a template in a chosen notation, for some stakeholders |
| Viewpoint | The conventions for constructing, interpreting, using and analysing one type of view — where you look from |
| Perspective | A collection of activities, tactics and guidelines ensuring the system exhibits a set of related quality properties. Applied to views; never produces new views |
| AD | Architecture Description — consists of one or more views, plus possibly principles, standards and glossaries |
| SAD | Software Architecture Document |
| arc42 | A freely licensed SAD template of twelve optional sections, including architecture decisions (9), quality requirements (10), and risks and technical debt (11) |
| C4 model | Simon Brown's four-level way of drawing software structure — system context, container, component, code — plus landscape, dynamic and deployment diagrams. Notation-independent, but every diagram needs a key, typed elements and labelled relationships |
| Container (C4) | Something that must be running for a system to work — web application, API, worker, database, queue. Not a Docker container |
| Diagrams as code | Diagrams written as text (Mermaid, PlantUML, Structurizr DSL) and versioned with the code. A model (Structurizr, an ArchiMate repository) goes further: one set of elements behind many views |
| d/s/o matrix | Views and Beyond's stakeholder × view matrix: detail, some information, or overview only |
| Gap analysis | Baseline set against target, element by element: kept, changed, new or retired. Each gap becomes a transition work package (TOGAF) |
| mTLS | Mutual TLS — both ends present certificates, so each authenticates the other |
| ADR | Architecture Decision Record — a short record of one decision: title, status, context, decision, consequences (Nygard); MADR adds the options considered |
| View packet | In Views and Beyond, the smallest bundle of documentation a stakeholder is given: one primary presentation plus its element catalogue, context, rationale and links to related packets |
| Business driver | Something the business depends on to keep succeeding and growing — a resource, process or condition it must protect or improve. Named with a noun |
| Business goal / objective | A goal is a broad, qualitative statement of desired achievement; an objective is a specific, measurable step towards it. Objectives are SMART (specific, measurable, achievable, relevant, time-bound) |
| Business capability | What a business does now and must do to meet future challenges — the what, not the how |
| Value stream | An end-to-end collection of value-adding activities creating a result for a customer or stakeholder. It is drawn as stages, each delivering a value item; business capabilities are mapped to the stages they enable |
| Capability map | A non-redundant, noun-named hierarchy of business capabilities (level 1, level 2…); rated today and at target it becomes a heat map |
| Functional decomposition | A breakdown of a system or process into its functions; describes the solution, not the business (contrast business capability) |
| Power/interest grid | Mendelow's 2×2 of stakeholder power against interest; quadrants: manage closely, keep satisfied, keep informed, monitor |
| RACI | Responsible, Accountable, Consulted, Informed. Exactly one Accountable per activity; accountability precedes responsibility |
| RAID log | One register of Risks, Assumptions, Issues and Dependencies, each with an owner and a next step |
| Clarification question | A question to the customer about a gap in an RFP; until it is answered, the planned answer is logged as an assumption |
| Transition requirement | A capability needed only to move from current to future state, not needed once the change completes |
| Utility tree | QA → attribute refinement → prioritised ASR scenarios, each rated for business importance and difficulty to achieve (H/M/L). A common variant — used in this course's worked example — rates architectural impact instead of difficulty |
| QAW | Quality Attribute Workshop — SEI's eight-step scenario elicitation. Lighter variants exist: an abbreviated workshop, and architect-led elicitation |
| Six-part scenario | Source, Stimulus, Artifact, Environment, Response, Response measure |
| ADD | Attribute-Driven Design — iterative design method organised into rounds (steps 1–7) and iterations (steps 2–7) |
| ATAM | Architecture Tradeoff Analysis Method — SEI's architecture evaluation method |
| Risk / non-risk (ATAM) | A risk is a potentially problematic architectural decision; a non-risk is a sound decision, recorded with the assumptions it relies on |
| Sensitivity point | A property of one or more components that is critical for achieving a particular quality-attribute response |
| Trade-off point | A property that is a sensitivity point for more than one quality attribute, so improving one response worsens another |
| Risk theme | A group of ATAM risks sharing a systemic cause, traced to the business driver it threatens |
| ARID / CBAM / DCAR | Lighter evaluation methods: ARID has a design's future users work scenarios against an intermediate design; CBAM ranks strategies by benefit per unit of cost; DCAR reviews recorded decisions one at a time |
| Coupling / Cohesion | Coupling is interdependence between modules; cohesion is relatedness within a module. Low coupling, high cohesion |
| SOA vs microservices | Scope (SOA is enterprise-wide, microservices are application-level) and, above all, sharing: SOA shares as much as possible, usually through an ESB; microservices share as little as possible, each service owning its data |
| Mediator vs broker | Event-driven topologies: mediator orchestrates multi-step events centrally; broker chains events with no central orchestration — orchestration vs choreography |
| Modular monolith | One deployable unit split into domain modules with enforced boundaries |
| Service-based architecture | A few coarse-grained, separately deployed domain services, usually sharing a database (Richards & Ford) |
| Saga | A sequence of local transactions, each with a compensating action, replacing one distributed transaction; orchestrated or choreographed (Garcia-Molina & Salem, 1987) |
| Transactional outbox | Write the event to an outbox table in the same transaction as the state change; a relay publishes it at least once |
| Idempotent consumer | A consumer for which processing the same message twice has the same effect as processing it once |
| Anti-corruption layer | A translation boundary that keeps an external system's model out of your domain model (Evans) |
| Canonical data model | One internal representation that every channel translates to and from (Hohpe & Woolf) |
| Strangler fig | Replace a system gradually behind a routing façade, one capability at a time, retiring old paths as they empty (Fowler) |
| CQRS | Command Query Responsibility Segregation — separate models, often separate stores, for changing data and for reading it; asynchronously updated read models are eventually consistent |
| Event sourcing | Storing each change to an entity as an immutable, appended event and deriving current state by replaying the events |
| Shard key | The attribute whose value decides which shard a row belongs to; the strategy (lookup, range or hash) maps keys to shards |
| Bulkhead | Separate pools of threads, connections or instances per dependency or consumer, so one failing dependency cannot exhaust shared resources |
| Circuit breaker | A wrapper on a remote call that stops calling after repeated failures (open), fails fast, then probes with a trial call (half-open) before resuming (closed) |
| Federated identity | Delegating authentication to a trusted external identity provider whose signed tokens carry claims about the user (OpenID Connect, SAML 2.0) |
| Gatekeeper | An internet-facing layer, holding no back-end credentials, that validates every request before forwarding it to trusted services |
| Valet key | A short-lived, narrowly scoped token or signed URL that lets a client read or write one object in a store directly |
| Crypto-shredding | Making personal data in immutable storage unreadable by encrypting it per person and destroying that person's key |
| Fault → Error → Failure | The fault is the defect; the error is the incorrect state liable to lead to failure; the failure is observable non-compliance with the specification, detected by users |
| Redundancy types | Spatial (copies in different places), temporal (over time, e.g. recovery blocks), informational (multiple versions of data) |
| Little's law | L = λ × W: in a stable system, the average number inside equals the arrival rate times the average time inside. Used in Module 3 to derive concurrent sessions from volumes |
| SLI / SLO / SLA | Indicator (measured ratio of good to valid events); objective (its target over a stated window); agreement (the contractual promise, usually set looser than the SLO) |
| Error budget | 1 − SLO: the unreliability a team may spend in the window before risky changes stop |
| Burn rate | How fast an error budget is being spent: observed error ratio ÷ (1 − SLO); a rate of 1 spends it exactly over the window |
| Observability | How well a system's internal state can be inferred from the telemetry it emits — metrics, logs, traces, events, profiles — including for failures nobody predicted |
| OpenTelemetry (OTel) | Vendor-neutral CNCF standard of APIs, SDKs, semantic conventions, the OTLP protocol and the Collector for producing and moving telemetry |
| Four golden signals | Latency, traffic, errors, saturation — the minimum to monitor for a user-facing system (Beyer et al., 2016) |
| RED / USE | Rate, Errors, Duration for request-driven services; Utilisation, Saturation, Errors for resources |
| Correlation ID | An identifier carried through every request, message and log record of one business transaction, so its whole history can be retrieved |
| Unit cost | Run cost ÷ a business unit such as a referral; the cost-efficiency measure, since total spend rises with success |
| RTO / RPO | Recovery time objective (longest acceptable time to restore service) and recovery point objective (most data loss acceptable, measured as time) |
| Hot / warm / cold standby | Determines switchover speed; checkpoints help a warm standby catch up faster |
| Blue-green / canary | Zero-downtime deployment techniques. Blue-green switches traffic between two full environments and rolls back by switching again; canary releases to a small cohort first to limit blast radius. Both complement automated tests — neither replaces them |
| CAP | Consistency, Availability, Partition tolerance — under partition you choose availability or consistency; "CA" is not a real option for a distributed system |
| PACELC | Abadi's extension of CAP: if Partitioned, choose Availability or Consistency; Else, choose Latency or Consistency |
| BASE | Basically available, Soft state, Eventual consistency |
| Scale cube | X = cloning, Y = functional decomposition, Z = data sharding |
| Cone of uncertainty | The best-case estimate accuracy at each project point — from 0.25×–4× at initial concept to 0.9×–1.1× once detailed design is complete (McConnell). Narrows only through project control — otherwise it becomes a cloud |
| Focus factor | The share of a working day that goes to planned work. Calendar effort = ideal person-days ÷ focus factor (e.g. 30 ÷ 0.75 = 40) |
| Parkinson's Law / Student Syndrome | Why overestimation's penalty is linear and bounded, while underestimation's is nonlinear and unbounded |
| Accuracy vs precision | Independent. The precision you present should match the accuracy you actually have |
| WBS | Work Breakdown Structure — deliverable-oriented (scope) or task-oriented (work) |
| Law of Large Numbers | Why decomposed bottom-up estimates beat one big estimate: errors partly cancel |
| Wideband Delphi | Anonymous iterative group estimation; any "no" vote returns the group to discussion |
| 100% rule | A WBS contains all the work the scope defines — project management included — and nothing outside it; at every level the children sum to their parent (PMI) |
| Three-point (PERT) estimate | Best (a), most likely (m), worst (b): mean = (a + 4m + b) ÷ 6, σ = (b − a) ÷ 6. Totals add means and variances, never standard deviations or worst cases |
| P50 / P80 | The effort or date with a 50% / 80% chance of not being exceeded; for a sum of many independent tasks, P80 ≈ mean + 0.84σ |
| Contingency | The gap between the percentile you commit to and the sum of most-likely estimates — derived, not a flat percentage |
| Fixed price / T&M / capped T&M | Pricing models that put estimation risk on the supplier, on the client, or share it up to a cap |
| Hypercare / warranty | Hypercare: a short period of heightened support after go-live with the build team on hand. Warranty: a contractual period in which defects in delivered scope are fixed at the supplier's cost |
| Go/no-go and rollback criteria | Measurable conditions agreed before cutover that decide whether it goes ahead, and the triggers and point of no return for reverting it |
| TCO | Total Cost of Ownership — initial plus all continuing costs through final decommissioning |
| FinOps | An operational framework and cultural practice for managing technology spend through shared accountability between engineering, finance and the business; phases Inform, Optimize, Operate |
| Commitment discount | Reserved capacity, a savings plan or a committed-use discount: a lower rate for committing to steady usage, paid whether used or not |
| Showback / chargeback | Reporting allocated cost to its owner without moving money / billing it to their budget |
| NIST essential characteristics | The five characteristics in NIST SP 800-145: on-demand self-service, broad network access, resource pooling, rapid elasticity, measured service |
| hpaPaaS | High-productivity aPaaS — declarative, model-driven, low-code/no-code, opaque infrastructure. Gartner's earlier term; the category is now usually called enterprise low-code application platforms (LCAP) |
| Container orchestration | Keeping declared container state true: scheduling, restarts, service discovery, rolling updates (e.g. Kubernetes) |
| Pets vs cattle | Whether losing one server takes everything down, or the herd carries on unaffected |
| Immutable infrastructure | Never modify instances in place; replace to update; plan for failure; don't let instances get stale |
| DIKW | Data → Information → Knowledge → Insight (Ackoff's original pyramid names the top level wisdom). Each level adds context and meaning — and data quality still matters, since data can carry capture and measurement errors |
| Analytics maturity | Descriptive → Diagnostic → Predictive → Prescriptive → Cognitive |
| MLOps | Engineering practice for building and running ML systems in production: data pipelines, training, model registry, serving, monitoring and retraining |
| Model registry | A versioned store of models with their metrics, model card and approval state; the release gate for serving |
| Training–serving skew | A model meeting inputs in production computed differently from its training features; prevented by sharing feature code and point-in-time correctness |
| Data drift / concept drift | A change in a model's input distribution / a change in the relationship between inputs and outcome; both degrade a model and trigger retraining |
| Model card | A short document of a model's intended use, data, evaluation (including per-group results), limitations and owner |
| Disparate impact | A group's rate of an outcome divided by the most favoured group's; 0.8 is a common screening threshold |
| RAG | Retrieval-augmented generation — answering from documents retrieved at query time, with citations, instead of from model memory |
| LLM gateway | A single point between applications and models for routing, quotas, redaction, logging, cost tracking and policy |
| Prompt injection | Input the model follows as instructions — direct from the user, or indirect via retrieved content or tool results |
| Groundedness | Whether every claim in a generated answer is supported by the sources supplied |
| Golden set | Curated evaluation cases with expected answers, sources and refusals, run as a release gate |
| LLM-as-judge | Using one model to grade another's output; scalable but biased, so calibrated against human labels |
| TTFT | Time to first token — the delay before a streamed answer starts to appear |
| CDC (change data capture) | Reading a database's change log and applying every insert, update and delete to another store, so copies trail the source by seconds or minutes |
| Open table format | Apache Iceberg, Delta Lake or Apache Hudi: table metadata over files in object storage, adding atomic commits, schema evolution and time travel |
| Lakehouse | Analytics that runs warehouse-style SQL directly on open-table-format data in object storage |
| Polling consumer | A consumer that fetches messages or extracts on its own schedule; its latency floor is the polling interval (Hohpe & Woolf) |
| Embedding | A dense vector produced by a model so that items with similar meaning lie close together |
| Approximate nearest neighbour (ANN) | An index (HNSW, IVF, PQ) that finds near vectors fast by trading a little recall |
| Hybrid search | Lexical (BM25) plus vector retrieval, merged — typically by reciprocal rank fusion |
| Reciprocal rank fusion (RRF) | Merges ranked lists by summing 1 ÷ (k + rank) per item, with k = 60 by convention |
| nDCG | Normalised discounted cumulative gain: a ranking metric that rewards relevant results more the higher they appear |
| Splittable compression | Compressed files must remain splittable or parallel frameworks lose their main advantage |
| Avro vs Parquet | Row-oriented, self-describing, schema-evolving (raw data) vs columnar with predicate pushdown (processed data) |
| Split-brain / fencing | Two active masters corrupting data; fencing ensures only one remains active |
| Consumer group | Kafka's abstraction generalising queueing (same group = load balanced) and pub-sub (different groups = broadcast) |
| AMQP vs JMS | AMQP specifies the wire format but no standard API (1.0 is ISO/IEC 19464; 0-9-1, with exchanges and bindings, is RabbitMQ's native model); JMS — now Jakarta Messaging — specifies the API but not the message format |
| Visibility timeout | SQS hides a received message for this long; if it is not deleted in time it reappears. Too small causes duplicate processing that cascades and chokes threads; too large delays failover. Remedies: extend visibility while processing, make consumers idempotent, use a dead-letter queue |
| Governance characteristics | Discipline, Transparency, Independence, Accountability, Responsibility, Fairness |
| Architecture Contract | In TOGAF, the joint agreement between those developing an architecture and its sponsors on deliverables, quality and fitness for purpose |
| Design authority | The standing forum that owns the technical design decisions of one programme, usually chaired by the lead solution architect |
| Fitness function | Any mechanism giving an objective assessment of an architectural characteristic, ideally automated in the delivery pipeline (Ford, Parsons & Kua) |
| Guardrail / paved road | A guardrail prevents or flags a breach in the platform itself; a paved road (golden path) is the supported default way to build, compliant by construction |
| Dispensation | A recorded, time-limited exception to an architecture principle, with an owner and an expiry date |
| PHI | Protected health information — identifiable health information; used in this course for any regime's equivalent |
| Trust boundary | A line on a data-flow diagram where data passes between zones that trust each other differently; threat analysis starts there |
| STRIDE | Spoofing, Tampering, Repudiation, Information disclosure, Denial of service, Elevation of privilege — each violates authentication, integrity, non-repudiation, confidentiality, availability or authorisation respectively |
| LINDDUN | Privacy threat categories: Linking, Identifying, Non-repudiation, Detecting, Data disclosure, Unawareness and unintervenability, Non-compliance (KU Leuven) |
| RBAC / ABAC | Role-based access decides by role; attribute-based access decides by attributes of user, resource and context — used here for tenant isolation |
| Break-glass access | A designed, time-limited, audited and alerted exception to normal permissions |
| Envelope encryption | Data encrypted with a data key (DEK), itself wrapped by a key-encryption key (KEK) held in a KMS |
| BYOK / HYOK | Bring your own key: imported into the provider's KMS. Hold your own key: kept outside the provider, which calls out to it for each use |
| Pseudonymisation | Replacing identifiers so data needs separately held information to identify a person; still personal data under the GDPR |
| WORM | Write once, read many — storage that refuses change or deletion until retention ends |
| SBOM | Software bill of materials — the components in a build (CycloneDX, SPDX) |
| Zero trust | No implicit trust from network location; every access decided per session by policy (NIST SP 800-207) |
Unit 17 of 18
RFPLumen DiagnosticsOpen anytime - every module Task uses this brief