Architecture Glossary
63 terms, defined the way they are actually used in production — each with a worked example and the mistake it usually causes. Definitions are revised when practice moves, not left at whatever was true the year they were written.
Distributed Systems13Quality Attributes9Agent Engineering8Consistency & Transactions8Data Stores8Performance & Latency7Architecture Practice6Delivery & Operations4
A
- ADR (Architecture Decision Record)An ADR is a short, immutable document capturing one architecturally significant decision: its context, the options considered, the choice, and the consequences.
- Amdahl's LawAmdahl's Law bounds the speedup available from parallelism: if a fraction s of the work is inherently serial, maximum speedup is 1/s no matter how many processors you add.
- Anti-Corruption LayerAn anti-corruption layer is a translation boundary that converts between your domain model and an external or legacy one, so foreign concepts and naming do not leak inward.
- At-Least-Once DeliveryAt-least-once delivery guarantees a message reaches its consumer one or more times, never zero.
- Availability NinesNines is shorthand for an availability target expressed as a percentage of successful time or requests.
B
- BackpressureBackpressure is the mechanism by which an overloaded component signals upstream to slow down, rather than accepting work it cannot complete.
- Blast RadiusBlast radius is the extent of damage a single failure or change can cause — how many users, tenants, regions or services are affected when one component fails.
- Bounded ContextA bounded context is an explicit boundary within which a domain model and its terms have one consistent meaning.
- Burn RateBurn rate is how fast an error budget is being consumed relative to the rate that would exactly exhaust it over the SLO window.
C
- Cache StampedeA cache stampede occurs when a popular cache entry expires and every concurrent request for it misses simultaneously, sending the full uncached load to the origin at once.
- Canary ReleaseA canary release routes a small fraction of production traffic to a new version, compares its error and latency metrics against the stable version, and promotes or rolls back based on the result.
- CAP TheoremThe CAP theorem states that when a network partition occurs, a distributed system must choose between consistency (every read sees the latest write) and availability (every request gets a non-error response).
- Circuit BreakerA circuit breaker wraps calls to a dependency and stops making them once the failure rate crosses a threshold, failing fast for a cool-down period before letting a trial request through.
- Connection PoolA connection pool keeps a fixed set of open database connections and lends them to requests, avoiding per-request handshake cost and bounding concurrency at the database.
- Consistent HashingConsistent hashing maps keys and nodes onto the same circular hash space, assigning each key to the next node clockwise.
- Context WindowThe context window is the maximum number of tokens a model can attend to in one request — system prompt, conversation history, retrieved documents, tool definitions, tool results and the response combined.
- Contract TestingContract testing verifies that a provider's API satisfies the expectations its consumers actually rely on, by running each consumer's recorded expectations against the provider in isolation.
- Conway's LawConway's Law observes that a system's structure mirrors the communication structure of the organisation that builds it.
- Coordinated OmissionCoordinated omission is a measurement error in which a load generator waits for a slow response before sending the next request, so the requests that would have been issued during the stall are never recorded.
- CouplingCoupling is the degree to which one component depends on the internals of another, measured by how much of it must change when the other changes.
E
- EmbeddingAn embedding is a dense vector representation of text, an image or other content, positioned so that semantically similar items sit close together under a distance measure such as cosine similarity.
- Error BudgetAn error budget is the amount of unreliability an SLO permits — the complement of the target.
- Eval HarnessAn eval harness is the automated test suite for a non-deterministic system: a fixed dataset of inputs, a scoring method per case, and a scoreboard run on every change.
- Eventual ConsistencyEventual consistency guarantees only that if writes stop, all replicas eventually converge to the same value.
- Exactly-Once SemanticsExactly-once semantics means each message produces its effect precisely once.
F
- Feature FlagA feature flag is a runtime switch that changes behaviour without a deploy, separating release from deployment.
- Fitness FunctionAn architectural fitness function is an automated, objective test of a non-functional requirement — coupling, latency, security posture, cost — run continuously in CI.
G
H
- HallucinationA hallucination is model output that is fluent and confident but not supported by the input or by fact — an invented citation, a non-existent API method, a fabricated figure.
- Hot PartitionA hot partition is a single shard receiving a disproportionate share of traffic, so it saturates while the rest of the cluster idles.
I
L
- LinearizabilityLinearizability is the strongest single-object consistency model: every operation appears to take effect instantaneously at some point between its invocation and its response, and once a write is visible to any client it is visible to all.
- Little's LawLittle's Law states that the average number of items in a stable system equals the average arrival rate times the average time each item spends there: L = λW.
- LSM Tree (Log-Structured Merge Tree)An LSM tree buffers writes in an in-memory table, flushes them to immutable sorted files, and merges those files in the background.
M
- Materialized ViewA materialized view stores the result of a query physically rather than recomputing it on each read.
- MCP (Model Context Protocol)The Model Context Protocol is an open standard for connecting AI applications to external tools, data and prompts through a uniform client/server interface.
N
O
- OLAP (Online Analytical Processing)OLAP describes workloads made of a small number of large queries that scan and aggregate many rows over few columns — revenue by region by month.
- OLTP (Online Transaction Processing)OLTP describes workloads made of many small, short-lived transactions that read and write a few rows each — placing an order, updating a profile.
P
- p99 (99th Percentile)p99 is the latency value below which 99% of requests complete — one request in a hundred is slower.
- PACELCPACELC extends CAP: if there is a Partition, choose Availability or Consistency; Else — when the network is healthy — choose Latency or Consistency.
- Prompt InjectionPrompt injection is an attack in which instructions embedded in content the model processes — a web page, a document, a ticket, a tool result — are followed as if they came from the operator or user.
Q
R
- RAG (Retrieval-Augmented Generation)RAG is the pattern of retrieving relevant documents at query time and placing them in the model's context so the answer is grounded in your data rather than in the model's parameters.
- Read-Your-Writes ConsistencyRead-your-writes consistency guarantees a client always sees its own completed writes, even while other clients may still see stale data.
- Replication LagReplication lag is the delay between a write committing on the primary and becoming visible on a replica.
S
- Saga PatternA saga replaces a distributed ACID transaction with a sequence of local transactions, each publishing an event that triggers the next.
- Serializable IsolationSerializable isolation guarantees that concurrent transactions produce the same result as some serial execution of them.
- ShardingSharding splits a dataset across independent database instances by a partition key, so each shard holds a disjoint subset.
- SLA (Service Level Agreement)An SLA is a contract with a customer that specifies a service level and the consequence of missing it — usually a service credit.
- SLI (Service Level Indicator)An SLI is a quantitative measure of one aspect of service behaviour, expressed as the ratio of good events to valid events — successful requests over total requests, say.
- SLO (Service Level Objective)An SLO is a target value for an SLI over a window — for example, 99.9% of requests succeed over 28 days.
- Snapshot IsolationSnapshot isolation gives each transaction a consistent view of the database as of its start time, so reads never block writes and writes never block reads.
- Split BrainSplit brain is the state where a network partition leaves two halves of a cluster each believing it is the authoritative one, so both accept writes.
- Strangler Fig PatternThe strangler fig pattern replaces a legacy system incrementally: a facade routes traffic, new functionality is built alongside the old, and routes are cut over one at a time until the original can be removed.
T
- Tail LatencyTail latency is the response time of the slowest requests — typically p99 and beyond.
- Thundering HerdA thundering herd is a surge of simultaneous requests to one resource, triggered by an event that releases many waiting clients at once — a cache expiry, a service restart, a synchronised retry, a cron.
- Tool-Use LoopThe tool-use loop is the core agent pattern: the model receives a goal and tool definitions, chooses a tool call, the runtime executes it and returns the result into the context, and the model decides again — repeating until it produces a final answer or hits a stop condition.
- Transactional OutboxThe transactional outbox pattern writes an outgoing message into an `outbox` table inside the same database transaction as the state change it describes.
- Trunk-Based DevelopmentTrunk-based development is the practice of integrating every change into a single shared branch at least daily, with short-lived branches measured in hours.
V
W
- Write AmplificationWrite amplification is the ratio of bytes physically written to storage to bytes logically written by the application.
- Write SkewWrite skew is the anomaly where two concurrent transactions each read an overlapping set of rows, each decides an invariant still holds, and each writes a different row — leaving the invariant violated.