Model Fallback
A useful fallback needs tested quality, independent availability, ready capacity, and a clear policy for timeouts and partial responses.
-
-
System Retirement
Retire a system completely: confirm no dependencies, keep retained records accessible under their obligations, then remove credentials, jobs, and costs.
-
Recovery Time and Recovery Point Objectives
Separate the recovery time objective, how long recovery may take, from the recovery point objective, how much recent data may be lost, and test both.
-
Forward Recovery
Know when rollback is no longer feasible after a cutover and how forward recovery repairs the new system in place with controlled, reviewed corrections.
-
Cutover
Plan a cutover with a rehearsal, a timed runbook naming owners and evidence, go/no-go criteria agreed in advance, a stop rule, and a decision owner.
-
Coexistence
Run old and new systems together with an authority rule, a shared comparison point, no duplicated side effects, and a plan for changes after the copy.
-
Business Continuity
Write business continuity requirements for a system change: what must keep working, the tolerated interruption, and how continuity is measured.
-
System Dependency
Find system dependencies beyond application code: jobs, files, procedures, reports, service accounts, manual steps, and rare month-end work.
-
Legacy System
Understand a legacy system through support, changeability, dependencies, and operating risk rather than age, using the GOV.UK criteria.
-
Double Counting
Recognize double counting when one benefit or cost appears under several criteria, check what each criterion includes, and keep weights honest.
-
Sensitivity Analysis
Use sensitivity analysis to test whether plausible changes in weights or inputs change a ranking, locate the reversal point, and name deciding evidence.
-
Weighted Scoring
Learn how weighted scoring adds anchored scores, why a weight is the value of a swing rather than abstract importance, and what the model assumes.
-
Scoring Anchors
Define scoring anchors that map raw measurements to 0 and 100, handle values outside the range, and see how fixed anchors differ from relative scaling.
-
Mandatory Gate and Preference
Separate mandatory gates with pass, fail, and unverified states from preferences that are scored, and see why a high total cannot offset a failed gate.
-
Evaluation Scorecard
Understand an evaluation scorecard as gates, anchored scores, weights, evidence, and a conditional recommendation, and what its total cannot claim.
-
Traceability
Understand requirements traceability as the link from a need to its requirement, tests, evidence, and owner, and what it reveals when something changes.
-
Total Cost of Ownership
Estimate total cost of ownership over a stated period and workload: license, integration, operation, review effort, and exit, with sensitivity shown.
-
Portability and Exit Plan
Define portability and an exit plan as tested requirements: what must move, what an export proves, and what leaving would cost in time and money.
-
Interoperability
Understand interoperability as verified exchange of fields, links, and permissions between systems, and why a connector label is not evidence.
-
Concurrency and Request Rate
Learn why 20 concurrent users do not define a load, how think time links concurrency to requests per second, and what a performance test must state.
-
Workload
Describe a workload as the demand a system must handle: users, concurrency, request rate, data volume, sizes, update timing, and cache state.
-
Acceptance Criteria
Write acceptance criteria that name the population, the check, the pass rule, and the person who judges the evidence before results are seen.
-
Constraint
Understand a constraint as a condition imposed from outside the design trade-off, with a source and an owner, and how to identify who may approve a change.
-
Functional and Quality Requirements
Distinguish what a solution must do from how well it must do it, and write quality requirements with a population, boundary, and conditions.
-
Need, Requirement, and Design Choice
Separate a user need, a testable requirement, and a design choice so that a preferred product is not written down as an unavoidable condition.
-
Technology Selection
Understand technology selection as choosing an approach and tools against written requirements, with evidence states, gates, cost, and exit in view.
-
Request Tracing
Learn how request tracing connects operations within a request and helps locate failures without proving answer quality.
-
Continuous Delivery and Deployment
Distinguish a release kept ready for deployment from eligible changes deployed automatically.
-
Continuous Integration (CI)
Understand how continuous integration checks combined changes and why a passing build is not a release approval.
-
Rollback
Learn what rollback restores, why compatibility matters, and which effects need separate recovery actions.
-
Shadow Testing
Learn how shadow testing evaluates a candidate using copied inputs without using its result for the user decision.
-
Canary Deployment
Understand how a canary deployment limits initial exposure while comparing a candidate with the current release.
-
Release Gate
Learn how a release gate turns acceptance criteria and evidence into a promotion decision.
-
Training-Serving Skew
Understand how differences between training and serving inputs can break predictions without changing the model.
-
Experiment Tracking
Learn how experiment tracking connects runs, inputs, settings, metrics, and artifacts for meaningful comparison.
-
LLMOps
Learn what LLMOps manages when prompts, retrieval, model services, and tools change an LLM application.
-
MLOps
Understand how MLOps connects model development, evaluation, deployment, monitoring, and recovery.
-
Latency Guardrails
Define latency guardrails with measurement boundaries, percentiles, realistic load, and a clear response when a limit is exceeded.
-
Cloud Inference
Distinguish cloud inference from training and deployment, and compare online, asynchronous, batch, and on-device execution.
-
Caching in Machine Learning
Learn when ML caches can safely reuse results, with model versions, data freshness, permissions, expiry, and cache failures in view.
-
Authorization
Understand authorization as deciding permitted actions on resources, including ownership, current policy, and cached results.
-
Authentication
Understand authentication as verifying identity and distinguish it from authorization and request-level access decisions.
-
Response Latency
Measure response latency with a clear boundary and distinguish it from model time, throughput, and data freshness.
-
Deployment
Understand deployment as putting a selected release into an execution environment, with validation, rollout, and rollback.
-
Model Registry
Learn how a model registry tracks versions and metadata without replacing evaluation, deployment, or inference.
-
Model Artifact
Understand the saved model representation and the dependencies needed to load and use it correctly.
-
Endpoint
Learn what an endpoint identifies and why a reachable address does not guarantee a usable model service.
-
API
Understand an API as a software interface, including inputs, outputs, errors, and its distinction from an endpoint.
-
Data Plane
Understand the data plane as the application processing path, including runtime checks and dependencies.
-
Control Plane
Learn how a control plane manages desired configuration and resources, and how it relates to the running data path.