| Parts of a large system | Load balancer, app servers, cache, database, queue and workers, CDN; each added when a requirement needs it | The parts of a system |
| Functional vs non-functional | What it must do / how well it must do it | Requirements |
| Constraint | A fixed fact (budget, region, team, law) that removes options | Constraints |
| Quality attribute | Needs a number: "p99 under 300 ms", not "fast" | Quality attributes |
| Requests per second | Daily requests ÷ 86,400 (about 100,000 s); peak is a few times that | Back-of-envelope estimates |
| Storage growth | Rows per day × row size × days; size for year three | Storage growth |
| Latency order | Memory ≪ SSD ≪ same-zone network ≪ cross-region round trip | Latency numbers |
| Latency budget | Split the target between the hops | Latency budget |
| p50, p99 | Typical request / the slow one in a hundred; the average hides it | Percentiles |
| Tail at scale | Wait on many services and the rare slow reply becomes common | The tail at scale |
| Decision record | What you chose, what it gives up, and why | Trade-offs |
| Exam question | A requirement plus one deciding phrase | How the exam asks |
| Scale up (vertical) | Bigger machine: simple, has a ceiling, one failure loses all | Vertical scaling |
| Scale out (horizontal) | More machines: no ceiling, a failure costs one share | Horizontal scaling |
| Stateless service | State moved out, so any server can be copied or removed | Stateless services |
| Session store or signed token | Keep sessions off the server | Sessions |
| Sticky sessions | Ties a user to one server; avoid | Sessions |
| Auto scaling group | Follows load between a minimum and a maximum | Auto scaling |
| Target tracking, step, scheduled, predictive | Ways to decide when to scale | Scaling policies |
| Target tracking arithmetic | Servers × metric ÷ target, rounded up; the metric must fall as servers are added | Target tracking |
| Scale set, managed instance group | Azure’s and Google’s Auto Scaling group | Auto scaling on three clouds |
| Scaling lag | Reactive scaling arrives late; schedule known peaks | Scaling lag |
| Scale-in: draining and lifecycle hook | Deregistration delay 300 s, then Terminating:Wait for up to an hour by default | Scale-in |
| Load balancer | One address in front, only healthy servers behind | Load balancers |
| Cross-zone load balancing | Every node uses every zone’s targets; always on for ALB, off by default for NLB | Cross-zone balancing |
| Layer 4 vs layer 7 | Forwards connections (NLB) / reads requests (ALB) | Layer 4 and layer 7 |
| Round robin vs least connections | Fair to requests / fair to servers | Balancing algorithms |
| ALB or NLB | HTTP features and WAF, or static IP, UDP, PrivateLink and raw throughput | ALB and NLB |
| Listener rule | Route by host and path to target groups | Listener rules |
| Health check | Unhealthy and healthy thresholds take servers out and back | Health checks |
| DNS routing | Picks the region before the first packet | DNS routing |
| Simple, weighted, latency, failover, geolocation | Route 53 policies and their exam clues | Route 53 policies |
| Anycast | One address worldwide, failover without DNS | Global routing |
| Front Door, global Application LB | Azure’s and Google’s worldwide HTTP entry points | Load balancers on three clouds |
| Choose by access pattern | How data is read and written decides the database | Choosing a database |
| Read replica | More read capacity, a moment behind | Read replicas |
| Scaling order | Fix queries, pool connections, cache, add replicas, scale up; shard last | Scaling order |
| Multi-AZ standby | Synchronous copy in another zone, loses nothing on failover | Multi-AZ |
| Standby vs replica | For staying up / for reading more | Standby or replica |
| Aurora storage | Six copies across three zones, shared by every replica | Aurora storage |
| Sharding | Split by key across machines; pick a key that spreads load | Partitioning and sharding |
| RCU, WCU | DynamoDB read and write units, rounded up by item size | DynamoDB capacity |
| Read-after-write | Read from a copy and you may read the past | Consistency |
| Tight coupling | A direct call ties two services' speed and uptime | Coupling |
| Queue | Absorbs a spike; workers catch up at their pace | Queues |
| Standard vs FIFO | Scale, best-effort order / order per message group | Standard and FIFO |
| Visibility timeout | Message hides while worked on, returns if the worker fails | Message lifecycle |
| Dead-letter queue | Where messages go after too many failed tries | Message lifecycle |
| Peek-lock, ack deadline | Service Bus’s and Pub/Sub’s visibility timeout: 1 minute and 10 s by default | Queues on three clouds |
| Redrive policy | deadLetterTargetArn plus maxReceiveCount on the source queue; visibility timeout six times the function timeout | Redrive policy |
| Idempotency key | Messages can arrive twice; handle them safely twice | Idempotency |
| Transactional outbox | Event stored in the same transaction as the data, published afterwards by a relay | Transactional outbox |
| Queue vs topic | One worker per message / every subscriber gets a copy | Publish and subscribe |
| Fan-out | Topic delivering to several queues | Publish and subscribe |
| Event bus | Routes each event by its content (EventBridge) | Event buses |
| Queue, topic, bus, stream, workflow | Which to pick | Choosing the tool |
| Monolith, microservices, serverless | Where the boundaries fall | Application shapes |
| Concurrency | Requests per second × duration | Functions |
| Cold start | A new instance pays for starting; warm ones do not | Cold starts |
| Reserved and provisioned concurrency | A free guarantee and cap; and paid, pre-initialised instances | Concurrency controls |
| Serverless cost | Nothing when idle; steady heavy traffic is cheaper on servers | Serverless cost |
| API gateway | Auth, throttling, routing, before your code runs | API gateways |
| REST API or HTTP API | REST has caching, usage plans, WAF and private endpoints; HTTP is cheaper and simpler | REST and HTTP APIs |
| Fargate, Cloud Run, Container Apps | Managed containers without servers | Managed containers |
| Container Apps, Cloud Run | Azure’s and Google’s serverless containers; both scale to zero | Compute on three clouds |
| Service boundary | A business capability with its own data | Service boundaries |
| Monolith first | Split when team and scale need it | Monolith first |
| Timeout | Without one, a slow dependency holds every thread | Timeouts |
| Bulkhead | Each dependency gets its own pool | Bulkheads |
| Retry storm | Retries multiply through layers | Retries |
| Exponential backoff with jitter | Wait longer each try, with randomness | Backoff and jitter |
| Circuit breaker | Closed, open, half-open: stop calling a failing dependency | Circuit breakers |
| Graceful degradation | Serve a smaller answer instead of an error | Graceful degradation |
| Load shedding | Reject the excess so the rest succeed | Load shedding |
| Chaos engineering | Break things on purpose, with a hypothesis | Chaos engineering |
| Canary, blue-green | Release in small steps you can roll back | Safe deployments |
| Right-sizing | Match instance size to measured use | Elasticity and right-sizing |
| Data transfer cost | Free in; charged out to the internet, across zones and across Regions | Data transfer |
| On-demand, Savings Plans, Spot | Peaks / committed floor / work that can wait | Purchase options |
| Storage tiers | Rarely read data in infrequent-access or archive classes | Storage classes |
| Savings Plan or Reserved Instance | Commit a spend (flexible) or a configuration (rigid); only zonal RIs and Capacity Reservations hold capacity | Commitments compared |
| Metrics, logs, traces | Something is wrong / what happened / where the time went | Observability |
| SLO and error budget | "Reliable enough" as a number to spend | SLOs and error budgets |
| CloudTrail, Config, Budgets | Who changed it, what it looked like, and whether spend will pass the limit | Tools on three clouds |
| Six pillars | Operational excellence, security, reliability, performance, cost, sustainability | The Well-Architected Framework |
| Design review | Turns "is it good?" into a ranked list of risks | Design review |
Nothing matches. Try a shorter word.