英文面试意群速查

按场景组织的背诵材料。每个意群是一个完整的表达单元（3-8 个词），面试时可以整块输出。

目标：不需要临场组织语言，靠肌肉记忆直接说出来。

关联：Java 八股 · 风控面试题 · 刷题进度

一、自我介绍 & 项目描述

开场白

• "I'm a Java backend developer with X years of experience, primarily focused on risk control systems."
• "I've been working on high-throughput, low-latency microservices for the past three years."
• "My main tech stack includes Java, Spring Boot, Kafka, Redis, and MySQL."

描述项目角色

• "I was the tech lead on the real-time risk engine project."
• "I owned the end-to-end delivery of the fraud detection pipeline."
• "I worked cross-functionally with the data science and compliance teams."
• "I was responsible for the core decision engine that processes 100 million transactions per day."

描述项目规模（必须量化）

• "Our system handles 10 billion daily requests with a P99 latency under 50 milliseconds."
• "We reduced the false positive rate by 30% while maintaining a 99.9% pass rate for legitimate transactions."
• "The platform serves 200 million users across three data centers."

二、描述技术决策（Why we chose X）

选型表达

• "We evaluated several options and ultimately went with X because..."
• "The main trade-off was between X and Y. We chose X because..."
• "We migrated from X to Y to achieve better throughput / latency / scalability."
• "X didn't meet our SLA requirements, so we switched to Y."

量化表达

• "This gave us a 3x improvement in throughput."
• "We brought the tail latency down from 200ms to 50ms."
• "This optimization reduced our infrastructure cost by 40%."
• "After the migration, we saw a significant reduction in GC pauses."

做选择时的话术

• "The tipping point was when we realized X couldn't scale beyond..."
• "We did a proof of concept with both options and X came out ahead in..."
• "One concern was X's operational complexity, but the performance gain was worth it."

三、架构设计讨论

描述系统架构

• "The system follows a layered architecture — gateway, feature service, rule engine, and model serving."
• "We use an event-driven architecture with Kafka as the central message bus."
• "The design follows the separation of concerns principle — each service has a single responsibility."
• "We implemented a progressive filtering strategy: rules first, then scorecards, then ML models."

描述高可用

• "We run across multiple availability zones with automatic failover."
• "We have a tiered degradation strategy — if the ML model is down, we fall back to rules."
• "The system is designed to be fault-tolerant at every layer."
• "We use circuit breakers to prevent cascading failures."

描述数据流

• "The data pipeline looks like this: events come in through Kafka, get processed by Flink, and the results are stored in Redis for real-time serving."
• "We use Change Data Capture to keep the cache in sync with the database."
• "Features are computed in both batch and streaming pipelines and merged at serving time."

四、Java 八股高频意群

并发

• "synchronized is a JVM-level keyword that uses monitor enter/exit bytecodes"
• "ReentrantLock is an API-level lock backed by AQS (AbstractQueuedSynchronizer)"
• "The thread pool uses a core-max-queue model — tasks go to core threads first, then the queue, then extra threads up to max, then the rejection handler kicks in"
• "We use a CallerRunsPolicy as the rejection handler to provide backpressure"
• "ThreadLocal can cause memory leaks in thread pools because threads are reused, not destroyed"
• "CompletableFuture gives us non-blocking composition with thenCombine, thenCompose, and exceptionally"

JVM

• "The heap is divided into young generation — which includes Eden and two survivor spaces — and old generation"
• "G1 uses a region-based layout and prioritizes collecting regions with the most garbage — hence 'garbage first'"
• "ZGC uses colored pointers and load barriers to achieve sub-millisecond pauses"
• "We tuned the GC by setting Xms equal to Xmx to avoid heap resizing overhead"
• "Metaspace stores class metadata and uses native memory, so it can grow beyond the heap limit"

Spring

• "Spring uses a three-level cache to resolve circular dependencies"
• "The third level holds an ObjectFactory that defers AOP proxy creation until it's actually needed"
• "AOP proxies are created in the postProcessAfterInitialization phase of the BeanPostProcessor"
• "Transaction self-invocation doesn't work because the call bypasses the proxy"
• "By default, @Transactional only rolls back on unchecked exceptions — you need rollbackFor = Exception.class for checked exceptions"

MySQL

• "InnoDB uses B+ tree indexes — leaf nodes form a doubly-linked list for efficient range scans"
• "MVCC implements Read Committed and Repeatable Read using undo logs and Read Views"
• "Gap locks prevent phantom reads by locking the intervals between index records"
• "We use covering indexes to avoid table lookups — the query is satisfied entirely from the index"
• "For sharding, we chose hash-based partitioning by user ID to ensure even data distribution"

Redis

• "We use the Cache Aside pattern — read from cache first, on miss read from DB, then populate the cache"
• "For writes, we update the database first, then delete the cache — not update the cache"
• "Distributed locking uses Redisson with a watchdog that automatically renews the lock every 10 seconds"
• "We switched from RDB to AOF with mixed persistence for better data safety"
• "Hot keys are handled with local caching using Caffeine as the first layer"

消息队列

• "Kafka achieves high throughput through sequential writes, zero-copy, and page cache utilization"
• "We guarantee at-least-once delivery by using manual offset commits after successful processing"
• "Exactly-once semantics are achieved through idempotent producers and transactional APIs"
• "We handle message backlog by scaling up consumers and using temporary topics with more partitions"
• "For ordered delivery, we route messages with the same partition key to ensure per-key ordering"

分布式

• "In practice, we choose between CP and AP since network partitions are inevitable"
• "We use TCC (Try-Confirm-Cancel) for financial transactions because it provides strong consistency"
• "For eventual consistency, we use the local message table pattern — it's simple and reliable"
• "Idempotency is achieved through a dedup table using the request ID as a unique key"
• "Our rate limiter uses a sliding window algorithm implemented in Redis with Lua scripts"

五、风控专业术语意群

风控系统核心

• "The decision engine evaluates each transaction through a multi-stage pipeline: pre-processing, feature computation, rule matching, model scoring, and post-processing"
• "We use a progressive filtering approach — simple rules first to filter obvious cases, then ML models for nuanced decisions"
• "The feature platform merges real-time features from Flink and offline features from Spark batch jobs"
• "Our rule engine supports hot reload — rules can be updated without restarting the service"

模型 & 策略

• "We built a credit scorecard using WOE encoding and logistic regression with a KS of 0.38"
• "Model drift is monitored using PSI (Population Stability Index) — we trigger retraining when PSI exceeds 0.25"
• "We run champion-challenger experiments with a 90/10 traffic split and measure statistical significance"
• "SHAP values provide per-transaction explainability — we can tell exactly which features drove the decision"
• "We handle cold-start for new users by relying on device fingerprinting and graph-based risk propagation"

欺诈检测

• "Fraud ring detection uses graph neural networks on a heterogeneous graph of users, devices, and IPs"
• "Account takeover (ATO) detection monitors login patterns, device changes, and behavioral biometrics"
• "Synthetic identity fraud is detected through entity resolution and cross-referencing multiple data sources"
• "We use Flink CEP for complex event processing — detecting patterns like 'three failed logins followed by a new device within 10 minutes'"

六、行为面试 & 软技能

描述挑战

• "One of the biggest challenges I faced was..."
• "The hardest part was balancing X and Y — we needed both low latency and high accuracy"
• "We hit a bottleneck in X, which was causing Y. I investigated and found that..."

描述解决方案

• "I proposed a solution that involved..."
• "The key insight was to..."
• "I worked with the team to refactor the critical path, which brought latency down by..."
• "We implemented a two-tier caching strategy that eliminated the bottleneck"

描述结果

• "As a result, we saw a significant improvement in..."
• "This reduced the false positive rate by 30% while keeping the pass rate above 99.5%"
• "The solution was well-received by stakeholders and became the standard pattern for similar services"
• "We also open-sourced the component and it was adopted by three other teams"

描述失误/教训

• "Looking back, I would have done X differently because..."
• "The main lesson learned was the importance of..."
• "We post-mortemed the incident and put safeguards in place to prevent recurrence"

七、反问面试官

• "What does the tech stack look like for the risk control team?"
• "How does the team balance model accuracy with system latency?"
• "What's the maturity level of the current risk platform — are you building from scratch or improving an existing system?"
• "How does the team handle model deployment — is there an ML platform team or does each team manage their own?"
• "What's the on-call rotation like?"

八、关键名词发音速查

关联

• Java 后端八股 — 对应的面试题文件
• 风控面试题 — 对应的面试题文件
• 刷题进度 — 学习进度跟踪