真题 Arena 真题竞技场

Billions of events/day, 24h retry window, per-endpoint ordering, idempotency, DLQ, multi-tenant isolation, cost controls. 日均十亿级事件、24h 重试窗口、每 endpoint 顺序、幂等、DLQ、多租户隔离、成本控制。

REST resource semantics, cache invalidation, DB schema, queue retry semantics. Cache-aside vs write-through trade-offs explicit. REST 资源语义、缓存失效、DB schema、队列重试语义。显式讨论 cache-aside 与 write-through 的权衡。

External dependency (ServiceB) for URL lookup + 24h retry forces you to design a robust state machine, TTL config cache, circuit breaker. URL 需从 ServiceB 查询 + 24 小时重试，把系统推向「状态机 + 外部一致性 + 可靠重试」。

Real-time chat, channels, presence, offline delivery, fan-out. The "2-week MVP" framing is a scope-discipline trap. 实时消息、频道、在线状态、离线投递、fanout 策略。「2 周 MVP」是范围控制的陷阱。

Multi-tenant workflow engine, runner pool, log/artifact storage, lease-based scheduling, "reliability first" per interviewer. 多租户工作流引擎、Runner 池、日志/制品存储、基于 lease 的调度，「可靠性优先」。

Builds on O5 with productization: YAML config parsing/versioning, event integration, secret management, permission tokens, audit. 在 O5 基础上加产品化能力：YAML 解析/版本化、事件集成、Secret 管理、权限 Token、审计。

Minimum distributed-workflow subset: state-driven scheduling via CDC, step table, lease-based serialization. 分布式工作流引擎最小子集：CDC 驱动调度、step 表、lease 化串行。

Hybrid retrieval (BM25 + vector), reranker, where to insert the LLM (query rewrite vs summarize), offline/online eval. 混合检索（BM25 + 向量）、重排序、LLM 插入位置（Query rewrite vs Summarize）、离线/在线评估。

SET/GET/DEL with TTL + range scans; evolve to WAL, snapshots, sharding, replication. Follow-ups add GROUP BY, ORDER BY. SET/GET/DEL + TTL + 范围扫描；演进到 WAL、快照、分片、复制。追问会加 GROUP BY、ORDER BY。

URL frontier, politeness scheduler, robots.txt cache, Bloom-filter dedup, canonicalization, content-hash for idempotent writes. URL frontier、礼貌调度器、robots.txt 缓存、Bloom 去重、URL 规范化、基于内容 hash 的幂等写入。

Frontend wireframe, API layer, thread/message history schema, prompt versioning, streaming output, multi-model selector. 前端线框图、API 层、线程/消息历史 schema、prompt 版本化、流式输出、多模型选择器。

End-to-end scaling: session persistence, GPU fleet management, regional routing, conversation storage, usage metering. 端到端扩展：会话持久化、GPU 集群管理、区域路由、对话存储、用量计费。

Data collection pipeline, model choice (rule vs classifier vs LLM-judge), latency budget, feedback loop, red-team flywheel. 数据收集管道、模型选择（规则 vs 分类器 vs LLM 评判）、延迟预算、反馈循环、red-team 飞轮。

Credit calculator with expiration, FIFO-consume-oldest semantics, prevent double-spend, fair-queue per tenant. 信用计算器带过期、先消费最旧额度、防止双重扣减、按租户公平排队。

SSE/WebSocket, low TTFT, in-stream moderation pipeline, backpressure, reconnect with cursor, server-to-client event model. SSE/WebSocket、低 TTFT、流式 moderation、背压、带游标重连、服务端到客户端事件模型。

RAG pipeline: ingestion → chunking → embedding → vector DB → hybrid retrieval → rerank → LLM w/ citation. Hallucination mitigation. RAG 管道：采集 → 分块 → embedding → 向量库 → 混合检索 → 重排 → 带引用的 LLM 生成。幻觉缓解。

Token bucket vs leaky bucket vs sliding window. Token-based billing for LLM APIs. Distributed coordination via Redis. 令牌桶 vs 漏桶 vs 滑动窗口。LLM API 的 token 级计费限流。基于 Redis 的分布式协调。

Store & search billions of embeddings. ANN algorithms (HNSW, IVF-PQ), sharding, hybrid filter queries, ingestion pipeline. 数十亿 embedding 的存储与搜索。ANN 算法（HNSW、IVF-PQ）、分片、混合过滤查询、采集管道。

Orchestrate training across thousands of GPUs: DP/TP/PP, ZeRO, checkpointing, fault-tolerance, job scheduler, bandwidth-aware routing. 数千 GPU 的训练编排：DP/TP/PP、ZeRO、Checkpoint、容错、作业调度、带宽感知路由。

One of OpenAI's five SD-pool classics. Base62 vs hash-of-URL, redirect path cache, write/read ratio, analytics, custom aliases, expiration. OpenAI 五大 SD 题池之一。Base62 vs URL 哈希、重定向路径缓存、读写比、分析、自定义别名、过期。

Narrower than Slack (O4): single/multi chat rooms, WebSocket + pub/sub, presence, message ordering, history pagination. No DMs, workspaces, or threads. 比 Slack（O4）窄：单/多聊天室、WebSocket + pub/sub、在线状态、消息顺序、历史翻页。不含 DM、工作区、Thread。

Language-runtime flavor: lexer → parser → AST → tree-walking evaluator. Sandbox, memory/time limits, stdlib surface area, REPL vs script mode, error UX. 语言运行时：词法 → 语法 → AST → 树遍历求值。沙箱、内存/时间上限、标准库边界、REPL vs 脚本模式、错误提示。

"Strong Hire" real question. Four-layer: ingestion/chunking → retriever → evaluator → generator. Traceable citations, per-doc ACLs, P95 < 2s, EN/中 multilingual, online learning from feedback. 「Strong Hire」真题。四层架构：采集/分块 → 检索 → 评估 → 生成。可追溯引用、按文档 ACL、P95 < 2s、中英双语、基于反馈的在线学习。

Heterogeneous GPU jobs scheduled across a pool with priority, preemption, at-least-once + idempotent handlers. GPU 池上调度异构任务；含优先级、抢占、at-least-once 与幂等 handler。

Async endpoint: upload JSONL, poll for completion. 24h SLA at 50% discount. Scheduler fills idle GPU capacity. 异步端点：上传 JSONL、轮询完成；24h SLA、5 折价；调度器填充 GPU 空闲。

Server-side conversational state: persistent threads, tool calls, file search, SSE streaming, run lifecycle. 服务端维持会话：持久 threads、工具调用、文件检索、SSE 流式、run 生命周期。

Dataset upload → validation → scheduled training → model artefact → private-tag inference. 数据集上传 → 校验 → 排程训练 → 模型制品 → 私有 tag 投入推理。

Targeting rules, percentage rollout, SDK, low-latency evaluation, audit trail, A/B analysis. 定向规则、百分比灰度、SDK、低延迟判定、审计日志、A/B 分析。

Bidirectional audio with interruption, ASR+LLM+TTS streamed, sub-300 ms first audio. 双向音频 + 打断；ASR+LLM+TTS 全流式；首音 < 300 ms。

Author graders, run suites, compare across runs, catch regressions pre-rollout. 编写打分器、跑 suite、跨次对比、上线前捕获回归。

Detect shared prefixes, cache KV states, serve subsequent tokens cheaper. 识别共享前缀，缓存 KV，后续 token 低成本生成。

Input + output moderation at API boundary; classifier ensemble; block/flag/review tiers; appeals. 入站与出站均审核；分类器集成；阻断/打标/人工复核；申诉流程。

IDE-triggered completions: p95 < 200 ms, huge cancellation rate, context with repo retrieval, privacy tiers. IDE 触发补全：p95 < 200 ms，取消率极高，上下文含仓库检索，隐私分级。

Text-to-image path, variable step counts, upscaling, NSFW filters, CDN for outputs. 文生图路径、可变步数、上采样、NSFW 过滤、输出走 CDN。

Per-request token counts, aggregation by user/org, daily billing, rate-limit enforcement, audit-grade records. 按请求 token 计数、按用户/组织聚合、日结计费、限流执行、审计级记录。

High-cardinality logs/traces from GPU fleet, cost-controlled sampling, fast search, retention tiers. GPU fleet 高基数日志/trace、成本可控采样、快速检索、分级保留。

Execute untrusted code/tool calls produced by LLMs in isolation: filesystem, network, time limits. 在隔离环境执行 LLM 产出的不可信代码/工具调用：文件系统、网络、时间限制。

Global front door for API traffic, auth, rate limit, routing to regional fleet, automatic failover. 全球前门：认证、限流、路由到区域 fleet、自动故障切换。