Câu hỏi phỏng vấn thiết kế hệ thống

A load balancer distributes requests across healthy servers and supports scaling.

Layer 7 load balancers inspect HTTP content such as paths, headers, and cookies.

Horizontal scaling adds more machines to spread the load across the cluster.

Vertical scaling increases the resources of a single machine, such as CPU, RAM, or disk.

Caching stores results so repeated requests can be served without recomputing them.

LRU removes the item that has not been used for the longest time.

A CDN serves content from edge locations closer to users, reducing latency.

DNS translates human-readable domain names into IP addresses used for routing.

Strong consistency makes the latest committed write visible to all subsequent reads.

CAP stands for Consistency, Availability, and Partition tolerance. The theorem states that a distributed system can guarantee only two of the three under a network partition. Caching, concurrency, and capacity are related concepts but do not form the CAP acronym.

Sharding distributes different subsets of data across separate database nodes.

Hash indexes provide fast equality lookups for exact keys. They compute a hash of the key to locate the value in constant time, which is ideal for point queries. B-tree scans, linked lists, and text indexes are better suited to ranges, sequences, or full-text search.

Read replicas serve SELECT traffic and reduce pressure on the primary database.

A message queue lets producers send work without waiting for consumers.

An async job queue accepts the request, returns quickly, and processes the work in the background.

A dead-letter queue stores messages that cannot be processed successfully after retries.

At-least-once delivery ensures messages are not lost but may be delivered more than once.

An idempotency key lets the server recognize and ignore duplicate requests.

REST uses HTTP verbs such as GET, POST, PUT, and DELETE to operate on resources.

HTTP 201 Created indicates that a new resource was successfully created.

Cursors reference a position in the dataset, so new or deleted rows do not shift pages.

Rate limiting caps the number of requests a client can send in a time window.

A token bucket permits requests at a controlled rate while allowing short bursts.

Object storage is designed for large immutable files such as images, videos, and backups.

An in-memory key-value store like Redis provides fast reads and writes for sessions.

Graph databases model nodes and edges, making relationship queries efficient.

Loosely coupled services can change and deploy without tightly depending on other services.

An API gateway handles cross-cutting concerns before requests reach backend services.

A circuit breaker stops requests to a failing service and lets it recover.

Random jitter spreads retries so many clients do not retry at the same instant.

Distributed tracing follows one request across services and shows where time is spent.

Availability is usually expressed as the percentage of time the service is operational.

TLS encrypts traffic between clients and servers to prevent eavesdropping.

A bulkhead partitions resources so failure in one service does not exhaust the whole system.

Autoscaling adds or removes instances based on load and keeps capacity aligned with demand.

Multi-region replication keeps data available when one region fails. It copies data and workloads to independent geographic locations, enabling failover during a disaster. A single region, no backups, and local-only caches increase the risk of total loss when that region becomes unavailable.

Round-robin, least connections, and IP hash are common strategies; DNS-only routing is not a load-balancer algorithm.

CDNs, in-memory caches, and database query caches reduce repeated reads; a relational schema is not a cache.

Sharding, read replicas, and partitioning spread or offload data; deleting indexes hurts performance.

Queues may provide delivery guarantees, dead-letter handling, and ordered partitions; storage is always finite.

Gateways commonly handle auth, rate limiting, and routing; migrations belong to the data layer.

Metrics, logs, and traces are the three observability pillars; feature flags are configuration.

Large global products normally need replicas, shards, or multiple data stores to scale.

Caches speed up reads but require invalidation and consistency decisions.

Shared databases create tight coupling and make independent deployment harder.

Distributed systems usually need idempotency and coordination to approximate exactly-once behavior.

TLS encrypts HTTP traffic and is the standard security protocol for the web.

LRU stands for Least Recently Used and is a common cache eviction policy.

Read replicas absorb read traffic, caching reduces repeated slow queries, a dead-letter queue captures failed messages, and rate limiting blocks abusive clients.

Strong consistency returns the latest write, eventual consistency converges over time, partition tolerance keeps the system working during network splits, and availability means it still responds.