34 termos que caem em entrevista, em inglês

Definição em inglês de propósito: na entrevista ninguém vai traduzir pra você. Se algum desses te pegar de surpresa numa call, volta aqui e anota no caderno.

ANN

Approximate Nearest Neighbor: fast "close enough" similarity search over vectors (HNSW is the common index).

ATS

Applicant Tracking System: software that screens resumes before a human sees them; mirror the job post's keywords.

Back-of-envelope

quick rough math to size a system before designing it.

Backoff + jitter

retry waiting longer each time, plus randomness so clients don't retry in sync.

Backpressure

when a data source produces faster than the consumer can handle; must pause/resume.

CDN

servers around the world caching static content close to users.

Circuit breaker

after N failures, stop calling a broken service for a while; fail fast.

Cold start

first-call latency while a serverless function's environment spins up.

Context window

max tokens a model can see at once.

Correlation ID

unique ID attached to a request and passed to every service, to trace it in logs.

DataLoader

batches many individual lookups into one query per tick; the fix for GraphQL's N+1.

DLQ

Dead Letter Queue: where messages land after failing N retries.

Eventual consistency

data becomes consistent across nodes "eventually", not instantly.

Fan-out

one event copied to many consumers.

Golden dataset

fixed set of question/expected-answer pairs used to test LLM quality on every change.

Idempotent

repeating the operation produces the same result (safe to retry).

Isolation level

how much one running transaction can see of another (Read Committed is the Postgres default).

Latency

time for ONE request. Throughput — requests handled per second.

LLM-as-judge

using a strong model to grade another model's outputs against a rubric.

MRR

how high the first correct result ranks, on average.

Multi-tenant

one system serving many customers whose data must stay isolated.

N+1

fetching N rows and then running one extra query per row; fix with joins or batching.

p50 / p95 / p99

latency percentiles; p95 = 95% of requests are faster than this.

PII

personally identifiable information (names, documents, emails).

Recall@k

of all correct documents, the fraction that appeared in the top k results.

Resolver

the function that fetches the data for one GraphQL field.

RPS

requests per second; 1M/day ≈ 12 RPS average, x3-5 for peak.

Sharding

splitting rows across multiple databases by a key.

SSE

Server-Sent Events: one long-lived HTTP connection where the server pushes to the client (used for LLM streaming).

Sticky session

the load balancer sends the same client to the same server (needed for WebSockets).

TDZ

Temporal Dead Zone: the gap where a let/const exists but can't be accessed yet.

Token

the unit LLMs read and bill by (~4 chars of English).

Top-k

the k best/most similar results.

TTL

Time To Live: how long a cache entry survives before expiring.