📄 Daily Digest

英语关键词背诵

独立背诵文档，通过 uv run python3 scripts/sync_to_anki.py --vocab 同步到 Anki。
范围来自 07-Interview-Prep/Java后端八股/ 和 07-Interview-Prep/风控面试题/ 的高频 Key Terms 与英文答案。
用法：先背连接词和意群，再背领域词。面试时尽量整块输出：term + purpose + trade-off + production scenario。

0. 背诵优先级

优先级 背什么 目标
P0 连接词、转折、取舍、排查句型 先让回答连贯，不要一个词一个词蹦
P0 大厂面经追问模式 先知道面试官会怎么追，不只背孤立术语
P1 Java 后端核心词 能把八股题讲成英文工程解释
P1 风控系统与模型核心词 能把项目经验讲成业务和技术结合
P2 发音、易混词、补充搭配 减少听不懂、说不准、用词不专业的问题

1. 常用连接词和回答骨架

功能 English 中文 面试短句
开始回答 From my perspective 从我的角度看 From my perspective, the key point is not the API itself, but the trade-off behind it.
分维度 In terms of ... 从某个维度来说 In terms of latency, ZGC is better, but in terms of throughput, G1 may be safer.
取决于 It depends on ... 取决于 It depends on the SLA, traffic pattern, and consistency requirement.
核心区别 The main difference is ... 主要区别是 The main difference is whether the operation is blocking, interruptible, or timeout-aware.
做取舍 The trade-off is ... 取舍是 The trade-off is that stronger reliability usually reduces throughput.
强调重点 The key idea is ... 核心思路是 The key idea is to keep the hot path short and predictable.
举例 For example 例如 For example, I would use Redis Lua to make rate limiting atomic.
补充 On top of that 在此基础上 On top of that, we need monitoring, alerting, and rollback.
对比 Compared with ... 与什么相比 Compared with synchronized, ReentrantLock gives more explicit control.
转折 However 但是 However, this approach increases operational complexity.
让步 Although ... 虽然 Although Redis locks are convenient, they must be designed carefully.
原因 This is because ... 这是因为 This is because Kafka only guarantees ordering within a single partition.
结果 As a result 因此 As a result, the consumer must be idempotent.
总结 In short 简而言之 In short, I would optimize for correctness first, then latency.
实战原则 My rule of thumb is ... 我的经验原则是 My rule of thumb is to use ArrayList by default unless linked-list behavior is required.
常见误区 A common pitfall is ... 常见误区是 A common pitfall is assuming volatile can replace locks.
排查顺序 I would first ..., then ... 我会先，然后 I would first confirm the symptom with metrics, then locate the root cause with logs and traces.
生产经验 In production 在线上环境 In production, I prefer graceful degradation over hard failure.
兜底 If ... fails, we fall back to ... 如果失败，回退到 If the model service fails, we fall back to rules or static lists.
权衡业务 We need to balance ... with ... 需要平衡 We need to balance fraud prevention with user experience.

2. 高频意群：可以直接整块输出

场景 必背意群 中文
描述系统目标 high-throughput, low-latency service 高吞吐、低延迟服务
描述核心路径 on the critical path 在关键链路上
描述非核心路径 move it off the critical path 从关键链路移出去
描述性能 reduce P99 latency 降低 P99 延迟
描述稳定性 prevent cascading failures 防止级联故障
描述降级 degrade gracefully 优雅降级
描述一致性 provide eventual consistency 提供最终一致性
描述强一致 enforce strong consistency 保证强一致性
描述幂等 avoid repeated side effects 避免重复副作用
描述排查 identify the bottleneck 定位瓶颈
描述扩展 scale horizontally 水平扩展
描述取舍 optimize for correctness first 优先保证正确性
描述风控 reduce fraud loss without hurting legitimate users 降低欺诈损失，同时不伤害正常用户
描述模型 improve recall at the same false positive rate 在误报率不变时提高召回
描述策略 compare the challenger against the champion 用挑战策略对比当前冠军策略
描述审计 provide auditability and explainability 提供可审计性和可解释性

2A. 大厂面经高频考察模式

这一节来自 2026-06-08 新增数据源：大厂面经数据源。
面经不是官方题库，价值在于反推“面试官会怎么追问”。背的时候按 公司/场景 -> 追问 -> 英文骨架 记。
公司 / 方向 面经高频追问 必背关键词 English 回答骨架
字节 / 后端 Redis 限流在多机房多集群下怎么保证一致性？ multi-datacenter rate limiting, eventual consistency, local quota, global quota, fallback For multi-datacenter rate limiting, I would split the problem into local quota, global coordination, and fallback when cross-region communication is slow.
字节 / 后端 AOP、Bean 初始化、循环依赖能不能讲到底层？ AOP proxy, Bean lifecycle, three-level cache, early reference I would explain Spring from the bean lifecycle first, then show where proxy creation and early references fit in.
字节 / 后端 HTTP 到 HTTPS、公钥私钥、链路安全怎么讲？ TLS handshake, public key, private key, certificate, symmetric encryption HTTPS uses asymmetric encryption for key exchange and symmetric encryption for efficient data transfer.
字节 / 后端 项目里的延迟优化是不是能量化？ P99 latency, bottleneck, metrics, tracing, critical path I would not only say we optimized latency; I would show the original P99, the bottleneck, and the final P99 after optimization.
美团 / 零售营销 Redis 库存扣减 + 写 DB 的代码有什么问题？ atomic stock deduction, Lua script, overselling, reliable message, eventual consistency The first issue is atomicity in Redis, and the second issue is consistency between Redis and the database.
美团 / 零售营销 接口超时怎么排查？ timeout diagnosis, metrics, logs, tracing, downstream dependency I would start from metrics to identify whether it is a client-side, gateway, service, database, or downstream dependency problem.
美团 / 零售营销 水平分表后跨分片分页怎么做？ sharded pagination, scatter-gather, cursor pagination, global ordering Cross-shard pagination is expensive because it requires scatter-gather and global ordering; I prefer cursor-based pagination when possible.
美团 / 零售营销 MQ 怎么保证消息不丢？ producer ACK, broker persistence, manual ACK, retry, DLQ, idempotency I would cover producer confirmation, broker persistence, consumer acknowledgment, retries, dead letter queues, and idempotent consumers.
蚂蚁 / 金融后端 金融系统为什么不能让 Agent 直接动钱？ human approval, permission boundary, audit trail, risk control, irreversible action For money-moving operations, the agent should propose actions, but final execution must go through permission checks, human approval, and audit trails.
蚂蚁 / 金融后端 分布式一致性和事务怎么选？ strong consistency, eventual consistency, TCC, Saga, local message table For fund-related flows, I would prioritize correctness and auditability, then choose TCC or reliable messages based on the business boundary.
蚂蚁 / Java 后端 虚拟线程、JMM、锁、事务会不会结合项目问？ virtual threads, JMM, synchronized, ReentrantLock, transaction boundary I would connect the mechanism to production scenarios, such as blocking I/O, transaction boundaries, and thread pool isolation.
腾讯 / AI 后台 微服务熔断降级、Kafka 高并发、SQL 注入防御怎么讲？ circuit breaker, degradation, Kafka partitioning, SQL injection prevention I would explain the failure mode first, then the protection mechanism, and finally the monitoring signal.
百度 / AI 后端 ConcurrentHashMap、CAS、JVM、Redis 分布式锁、慢 SQL concurrent map, CAS, GC, distributed lock, slow query I would answer from correctness first, then performance, and finally production diagnosis.
快手 / 后端 Redis 集群、HashMap、线程同步、JVM、操作系统底层 Redis Cluster, hash slot, lock, scheduler, virtual memory I would connect data structure details with system-level cost, such as memory locality, scheduling, and blocking.
金融风控 / 策略 误杀率和欺诈损失怎么权衡？ false positive, false negative, cost matrix, threshold selection, profit curve I would choose the threshold based on a cost matrix, not only model accuracy.
金融风控 / 模型 PSI 漂移、样本偏、模型解释怎么处理？ PSI, concept drift, covariate shift, SHAP, model governance I would separate population shift, concept drift, and feature quality issues before deciding whether to retrain or adjust rules.
风控系统设计 规则、模型、特征平台、反馈闭环怎么串起来？ rules engine, feature store, model serving, feedback loop, shadow mode I would design a layered pipeline: cheap rules first, feature fetching, model scoring, decision orchestration, and a feedback loop.

2B. 面经驱动必背追问句

面试官追问 English 回答开头 适用场景
你这个项目最难的点是什么？ The hardest part was not implementing the API, but making the system reliable under peak traffic. 项目深挖
线上出问题你怎么排查？ I would first reproduce the symptom from metrics, then narrow it down with logs, traces, and runtime evidence. 超时、慢 SQL、GC、MQ 积压
有没有量化收益？ The measurable impact was a reduction in P99 latency from X to Y, while keeping the error rate stable. 性能优化
这个方案有什么缺点？ The main drawback is operational complexity, so I would add monitoring, rollback, and a fallback path. 架构取舍
为什么不用另一个方案？ I would compare them across latency, consistency, implementation complexity, and operational risk. 技术选型
如果流量翻 10 倍怎么办？ I would first identify the bottleneck, then scale the stateless layer horizontally and protect stateful dependencies with backpressure. 高并发
如果依赖挂了怎么办？ If the dependency fails, we should fail fast, open the circuit breaker, and fall back to a conservative decision. 熔断降级
如何防止重复消费？ I would make the consumer idempotent with a business idempotency key, a unique constraint, or a state machine. MQ
如何避免超卖？ I would make stock deduction atomic with Lua or a transactional inventory service, and then use reliable messages to sync the database. Redis + DB
如何保证训练和线上一致？ I would define features once in the feature platform and serve them through both offline and online stores. 风控特征平台
如何处理欺诈样本极少？ I would avoid accuracy as the main metric and use precision-recall, cost-sensitive learning, and business loss metrics. 风控模型
如何解释拒绝原因？ I would return reason codes from rules and use SHAP values for model-level explanations. 模型解释 / 审计

3. Java 基础关键词

English 中文 常用意群 面试短句
HashMap 哈希表 hash table implementation HashMap is a hash table implementation with array buckets and linked lists or red-black trees.
hash collision 哈希冲突 when multiple keys map to the same bucket Hash collision happens when multiple keys map to the same bucket.
red-black tree 红黑树 reduce lookup cost under heavy collisions HashMap uses red-black trees to reduce lookup cost when collisions become severe.
resize mechanism 扩容机制 introduce latency spikes The resize mechanism can introduce latency spikes, so I pre-size the collection when possible.
dynamic array 动态数组 contiguous memory and fast random access ArrayList is backed by a dynamic array, which gives fast random access.
doubly-linked list 双向链表 efficient insertion and deletion by node reference LinkedList is a doubly-linked list, but it has poor cache locality.
immutability 不可变性 safe to share across threads String immutability makes strings safe to share and enables hashCode caching.
String constant pool 字符串常量池 reuse identical string literals The string constant pool reuses identical string literals to reduce memory usage.
checked exception 受检异常 must be declared or handled Checked exceptions must be declared or handled at compile time.
unchecked exception 非受检异常 usually indicates programming errors Unchecked exceptions usually indicate programming errors or runtime failures.
type erasure 类型擦除 remove generic type information at runtime Java generics are implemented through type erasure, so generic type information is mostly removed at runtime.
bridge method 桥方法 preserve polymorphism after type erasure The compiler generates bridge methods to preserve polymorphism after type erasure.
reflection 反射 inspect and invoke at runtime Reflection allows runtime inspection and invocation, but it has performance and encapsulation costs.
dynamic proxy 动态代理 add cross-cutting behavior Dynamic proxy is commonly used to add cross-cutting behavior without changing business code.
InvocationHandler 调用处理器 intercept method calls In JDK dynamic proxies, InvocationHandler intercepts method calls.
MethodInterceptor 方法拦截器 intercept CGLIB proxy calls In CGLIB proxies, MethodInterceptor intercepts target method calls.
Optional 可选值 represent absence explicitly Optional represents absence explicitly, but it should not be overused in entity fields.
Stream API 流式 API declarative data processing Stream API supports declarative data processing, but debugging can be harder than loops.
Record 记录类 immutable data carrier A record is a concise immutable data carrier with generated equals and hashCode.
Virtual Threads 虚拟线程 lightweight threads for blocking I/O Virtual threads are useful for high-concurrency blocking I/O workloads.
equals contract equals 契约 reflexive, symmetric, transitive, consistent The equals contract requires reflexivity, symmetry, transitivity, and consistency.
object layout 对象布局 object header, instance data, padding Java object layout includes the object header, instance data, and alignment padding.
GC Roots GC 根 starting points for reachability analysis GC Roots are starting points for reachability analysis during garbage collection.

4. Java 并发关键词

English 中文 常用意群 面试短句
visibility 可见性 changes made by one thread become visible to another Volatile guarantees visibility and ordering, but it does not guarantee atomicity.
atomicity 原子性 complete as one indivisible operation For compound operations like increment, we need atomic classes or locks.
happens-before 先行发生 basis for reasoning about memory visibility The happens-before rule is the basis for reasoning about memory visibility in Java.
memory barrier 内存屏障 prevent reordering and flush memory Memory barriers prevent unsafe reordering and help enforce visibility.
synchronized 同步锁 JVM-level monitor lock synchronized is a JVM-level monitor lock based on monitor enter and monitor exit.
ReentrantLock 可重入锁 API-level lock backed by AQS ReentrantLock is an API-level lock backed by AQS.
lock contention 锁竞争 many threads compete for the same lock Under high lock contention, explicit locks give more control.
AQS 抽象队列同步器 framework for locks and synchronizers AQS provides a queue-based framework for building locks and synchronizers.
CLH queue CLH 队列 queue for waiting threads AQS uses a CLH-style queue to manage waiting threads.
CAS 比较并交换 compare and swap atomically CAS updates a value only when the current value matches the expected value.
ABA problem ABA 问题 value changes from A to B and back to A The ABA problem can be solved with a version stamp or AtomicStampedReference.
ThreadPoolExecutor 线程池执行器 core-max-queue model ThreadPoolExecutor follows a core-max-queue model before triggering rejection.
rejection policy 拒绝策略 when the pool and queue are saturated A rejection policy is required when the pool and queue are both saturated.
backpressure 背压 slow down producers when consumers cannot keep up CallerRunsPolicy can provide simple backpressure by making the caller execute the task.
CountDownLatch 倒计数器 wait until several tasks finish CountDownLatch lets one or more threads wait until several tasks finish.
CyclicBarrier 循环栅栏 wait until all parties reach the barrier CyclicBarrier is reusable and waits until all parties reach the barrier.
Semaphore 信号量 limit concurrent access with permits Semaphore limits concurrent access by controlling permits.
ThreadLocal 线程本地变量 store per-thread context ThreadLocal can leak memory in thread pools if remove is not called.
WeakReference 弱引用 cleared when only weakly reachable ThreadLocalMap uses weak references for keys, but stale values can still leak.
CompletableFuture 异步编排 non-blocking composition CompletableFuture supports non-blocking composition with thenCompose and allOf.
optimistic read 乐观读 read first and validate later StampedLock provides optimistic reads, but the stamp must be validated.
work stealing 工作窃取 idle workers steal tasks ForkJoinPool improves parallelism through work stealing.
token bucket 令牌桶 allow bursts while limiting average rate Token bucket allows short bursts while controlling the average rate.
leaky bucket 漏桶 smooth outgoing rate Leaky bucket smooths traffic but does not preserve bursts.

5. JVM 关键词

English 中文 常用意群 面试短句
heap 堆 shared object allocation area Objects are usually allocated on the heap, which is shared across threads.
stack 栈 thread-private stack frames The stack stores thread-private stack frames and local variables.
Metaspace 元空间 native memory for class metadata Metaspace stores class metadata and uses native memory.
program counter 程序计数器 thread-private execution pointer The program counter records the current bytecode execution position for each thread.
Stop-The-World pause STW 停顿 all application threads are paused For latency-sensitive services, GC tuning focuses on reducing Stop-The-World pauses.
garbage collector 垃圾收集器 reclaim unreachable objects The garbage collector reclaims unreachable objects automatically.
G1 G1 收集器 region-based collector G1 uses a region-based layout and is a good general-purpose collector.
ZGC ZGC 收集器 low-latency collector ZGC is designed for very large heaps and strict latency targets.
SATB 快照标记 snapshot-at-the-beginning G1 uses SATB to handle concurrent marking safely.
colored pointers 染色指针 encode metadata in object pointers ZGC uses colored pointers and load barriers to reduce pause time.
read barrier 读屏障 barrier on object read Read barriers help concurrent collectors maintain correctness.
write barrier 写屏障 barrier on object write Write barriers track reference changes during concurrent marking.
heap dump 堆转储 snapshot of heap objects When old generation usage keeps growing, I would take a heap dump and analyze GC roots.
GC log GC 日志 evidence for GC behavior GC logs help identify allocation rate, pause time, and promotion failures.
promotion failure 晋升失败 object cannot move to old generation Promotion failure often indicates old generation pressure or fragmentation.
escape analysis 逃逸分析 determine whether an object escapes a method Escape analysis enables scalar replacement and lock elision.
scalar replacement 标量替换 replace object fields with scalar variables Scalar replacement can reduce heap allocation.
JIT compiler 即时编译器 compile hot code at runtime The JIT compiler optimizes hot code paths at runtime.
classloader leak 类加载器泄漏 classloader cannot be garbage collected Dynamic class generation can cause classloader leaks and eventually Metaspace OOM.
parent delegation 双亲委派 delegate loading to parent first Parent delegation prevents duplicate class loading and protects core classes.

6. Spring 关键词

English 中文 常用意群 面试短句
BeanDefinition Bean 定义 metadata for creating beans BeanDefinition stores metadata that Spring uses to create beans.
bean lifecycle Bean 生命周期 from definition to destruction The bean lifecycle goes from definition parsing to instantiation, injection, initialization, proxy creation, and destruction.
instantiation 实例化 create the object Instantiation creates the bean object before dependency injection.
dependency injection 依赖注入 inject dependencies into the bean Dependency injection wires required collaborators into the bean.
BeanPostProcessor Bean 后处理器 customize beans before and after initialization BeanPostProcessor can wrap beans with proxies after initialization.
circular dependency 循环依赖 beans depend on each other Spring resolves singleton setter-based circular dependencies through a three-level cache.
three-level cache 三级缓存 singleton pool, early references, factories The three-level cache exposes early references when resolving circular dependencies.
AOP proxy AOP 代理 proxy-based cross-cutting behavior Spring AOP works through runtime proxies, so self-invocation can bypass transactional behavior.
JDK dynamic proxy JDK 动态代理 interface-based proxy JDK dynamic proxy requires an interface.
CGLIB proxy CGLIB 代理 subclass-based proxy CGLIB creates a subclass proxy, so final methods cannot be advised.
transaction propagation 事务传播行为 how nested transactional calls behave REQUIRED joins the current transaction, while REQUIRES_NEW starts an independent one.
rollback rule 回滚规则 which exception triggers rollback By default, Spring rolls back RuntimeException and Error, so checked exceptions need rollbackFor.
self-invocation 自调用 method call bypasses the proxy Transaction self-invocation fails because the call bypasses the proxy.
auto-configuration 自动配置 configure beans based on classpath and conditions Spring Boot auto-configuration creates beans based on classpath, properties, and conditions.
conditional assembly 条件装配 register beans only when conditions match Conditional assembly keeps auto-configuration flexible.
TaskExecutor 任务执行器 abstraction for async execution @Async uses a TaskExecutor to run methods asynchronously.
application event 应用事件 publish and listen to domain events Application events decouple publishers from listeners inside the application.
transactional event listener 事务事件监听 listen after transaction phase @TransactionalEventListener runs listeners according to transaction phases.
AOT compilation 提前编译 optimize at build time Spring Boot 3 supports AOT compilation and GraalVM native images.

7. MySQL 关键词

English 中文 常用意群 面试短句
B+ tree B+ 树 high fan-out and range scans InnoDB uses B+ trees because they provide high fan-out and efficient range scans.
fan-out 扇出 many child pointers per node High fan-out reduces tree height and disk I/O.
clustered index 聚簇索引 leaf nodes store full row data In InnoDB, the clustered index stores the full row data in leaf nodes.
secondary index 二级索引 leaf nodes store primary key values A secondary index stores primary key values in leaf nodes.
bookmark lookup 回表 lookup full row by primary key A secondary index lookup may require a bookmark lookup if columns are not covered.
covering index 覆盖索引 satisfy query from the index only A covering index avoids bookmark lookups and can reduce P99 latency.
Index Condition Pushdown 索引下推 filter at storage engine layer ICP lets the storage engine filter rows earlier using index conditions.
leftmost prefix rule 最左匹配原则 composite index matching starts from the left Composite indexes must follow the leftmost prefix rule.
MVCC 多版本并发控制 consistent snapshot reads MVCC uses undo logs and Read View to provide consistent snapshot reads.
Read View 读视图 snapshot of active transactions Read View determines which row versions are visible to a transaction.
dirty read 脏读 read uncommitted changes Read Uncommitted may allow dirty reads.
phantom read 幻读 new matching rows appear InnoDB uses next-key locks to prevent phantom reads in locking reads.
record lock 记录锁 lock an index record A record lock locks a specific index record.
gap lock 间隙锁 lock the gap between records Gap locks prevent inserts into a locked index range.
next-key lock 临键锁 record lock plus gap lock A next-key lock combines a record lock and a gap lock.
deadlock 死锁 transactions wait for each other Deadlocks should be reduced through consistent lock ordering and short transactions.
EXPLAIN 执行计划 inspect query execution plan EXPLAIN helps identify access type, chosen index, rows scanned, and extra operations.
filesort 文件排序 sort outside index order Filesort usually means MySQL cannot use index order for sorting.
slow query log 慢查询日志 record slow SQL The slow query log is the first evidence source for SQL optimization.
deep pagination 深分页 large offset pagination Deep pagination is expensive because MySQL still scans and skips many rows.
sharding 分片 split data across databases or tables Sharding improves scalability but increases query and transaction complexity.
read-write splitting 读写分离 route reads to replicas Read-write splitting improves read scalability but introduces replication lag.
WAL 预写式日志 write log before data page WAL guarantees durability by writing redo logs before flushing data pages.
two-phase commit 两阶段提交 coordinate redo log and binlog InnoDB uses two-phase commit to keep redo log and binlog consistent.

8. Redis 关键词

English 中文 常用意群 面试短句
SDS 简单动态字符串 Redis string implementation Redis String is implemented with SDS rather than a raw C string.
quicklist 快速列表 linked list of listpacks Redis List uses quicklist to balance memory efficiency and update cost.
skiplist 跳表 ordered structure for range queries ZSet can use skiplist for efficient range queries.
cache penetration 缓存穿透 nonexistent keys hit the database Cache penetration happens when nonexistent keys repeatedly hit the database.
cache breakdown 缓存击穿 hot key expires under high concurrency Cache breakdown is a hot-key expiration problem under high concurrency.
cache avalanche 缓存雪崩 many keys expire together Cache avalanche happens when many keys expire together or the cache layer fails.
bloom filter 布隆过滤器 probabilistic membership test A Bloom filter can block requests for definitely nonexistent keys.
null caching 空值缓存 cache empty results briefly Null caching protects the database from repeated nonexistent-key requests.
distributed lock 分布式锁 lock across processes A Redis lock must use a unique value and release through Lua.
watchdog 看门狗续期 automatically renew lock expiration Redisson uses a watchdog to renew the lock while the owner is still alive.
RDB snapshot RDB 快照 point-in-time persistence RDB is compact and fast to restore, but it may lose recent writes.
AOF AOF 追加日志 append-only persistence AOF provides better durability by appending write commands.
mixed persistence 混合持久化 combine RDB snapshot and AOF tail Mixed persistence balances recovery speed and data safety.
Sentinel 哨兵 monitor and fail over master Sentinel provides monitoring and automatic failover for Redis masters.
Redis Cluster Redis 集群 shard by hash slots Redis Cluster shards data by hash slots.
hash slot 哈希槽 cluster partition unit Keys are assigned to hash slots in Redis Cluster.
Cache Aside 旁路缓存 read cache, load DB on miss Cache-aside usually provides eventual consistency rather than strong consistency.
delayed double delete 延迟双删 delete cache twice around DB update Delayed double delete reduces stale cache windows after database writes.
BigKey 大 Key key with too much data BigKeys can block Redis operations and cause network pressure.
HotKey 热 Key key with extremely high traffic HotKeys can overload a single Redis node.
SCAN 游标扫描 incremental key iteration SCAN is safer than KEYS for production key scanning.
UNLINK 异步删除 delete keys asynchronously UNLINK deletes keys asynchronously and reduces blocking time.
Lua script Lua 脚本 execute atomically in Redis Redis Lua scripts are atomic because Redis executes them in a single-threaded event loop.
PEL 待处理列表 pending entries list Redis Stream uses PEL to track delivered but unacknowledged messages.
LFU 最不经常使用 evict least frequently used keys LFU is useful when access frequency is more important than recency.
LRU 最近最少使用 evict least recently used keys LRU evicts keys that have not been used recently.

9. 消息队列关键词

English 中文 常用意群 面试短句
throughput 吞吐量 messages processed per second Kafka is optimized for high throughput through sequential writes and batching.
latency 延迟 time from produce to consume For real-time systems, we monitor both average latency and P99 latency.
message ordering 消息顺序 preserve order of messages Kafka guarantees ordering only within a partition.
partition key 分区键 route same business key to same partition For ordered delivery, the same business key must route to the same partition.
consumer group 消费者组 scale consumption horizontally Consumer groups allow multiple consumers to share partitions.
consumer lag 消费滞后 messages waiting to be consumed When consumer lag grows, I first scale consumers up to the partition count.
manual acknowledgment 手动确认 acknowledge after successful processing The consumer should acknowledge only after business logic succeeds.
dead letter queue 死信队列 store messages that cannot be processed A dead letter queue prevents bad messages from blocking normal consumption.
idempotent producer 幂等生产者 avoid duplicate production on retry Kafka idempotent producers prevent duplicate messages caused by producer retries.
idempotent consumer 幂等消费者 handle duplicate delivery safely Because retries can cause duplicate delivery, the consumer must be idempotent.
transactional API 事务 API atomic write across topics and offsets Kafka transactional APIs support exactly-once processing under strict conditions.
exactly-once semantics 精确一次语义 process each message exactly once Exactly-once semantics depend on idempotent producers, transactions, and committed reads.
at-least-once delivery 至少一次投递 retry until delivered At-least-once delivery is easier to guarantee, but consumers must handle duplicates.
sequential write 顺序写 append data sequentially to disk Sequential writes are much faster than random disk writes.
zero-copy 零拷贝 avoid copying data between user and kernel space Zero-copy reduces CPU overhead during message transfer.
page cache 页缓存 operating system file cache Kafka relies heavily on the operating system page cache.
rebalance 重平衡 redistribute partitions across consumers Rebalance can interrupt consumption, so it should be monitored carefully.
heartbeat 心跳 signal consumer liveness Heartbeats tell the coordinator that a consumer is still alive.
half message 半消息 transactional message before commit RocketMQ uses half messages before confirming local transaction status.
transaction check 事务回查 check uncertain local transaction status Transaction checks resolve uncertain transaction message states.
time wheel 时间轮 efficient timer structure Time wheels are commonly used to implement delay messages.

10. 分布式系统关键词

English 中文 常用意群 面试短句
CAP theorem CAP 定理 consistency, availability, partition tolerance Since network partitions are unavoidable, the real trade-off is usually between consistency and availability.
consistency 一致性 all nodes see the same data Strong consistency is critical for fund-related operations.
availability 可用性 every request receives a response Availability means the system can still respond even when some nodes fail.
partition tolerance 分区容错 tolerate network partitions Partition tolerance is required in distributed systems because networks can fail.
eventual consistency 最终一致性 converge after a delay Eventual consistency is acceptable when temporary inconsistency can be tolerated.
soft state 软状态 state may change over time without input BASE accepts soft state and eventual consistency.
2PC 两阶段提交 prepare and commit phases 2PC provides strong consistency but can block when the coordinator fails.
TCC Try-Confirm-Cancel reserve, confirm, or cancel resources TCC is suitable when business resources can be explicitly reserved and confirmed.
Saga 长事务编排 sequence of local transactions with compensation Saga relies on compensating transactions, but it does not provide isolation.
compensating transaction 补偿事务 undo a previous business action A compensating transaction reverses the effect of a completed local transaction.
local message table 本地消息表 persist event with local transaction The local message table pattern is simple and reliable for eventual consistency.
idempotency 幂等性 repeated calls have the same effect Idempotency prevents retries from producing repeated side effects.
idempotency key 幂等键 unique key for deduplication An idempotency key prevents duplicate requests from producing repeated side effects.
Snowflake ID 雪花算法 ID timestamp, worker ID, sequence Snowflake IDs are sortable and decentralized, but they depend on clock stability.
clock drift 时钟回拨 system clock moves backward Clock drift can cause duplicate Snowflake IDs if not handled.
rate limiting 限流 control request rate Rate limiting protects downstream services from overload.
sliding window 滑动窗口 rate limit over moving time window Sliding windows are smoother than fixed windows.
circuit breaker 熔断器 stop calling failing dependency Circuit breakers stop cascading failures.
fallback 兜底 use backup behavior when dependency fails Fallback keeps the core service available with reduced functionality.
half-open 半开状态 trial state after circuit opens Half-open state allows limited trial requests before closing the circuit.
service registry 服务注册中心 discover service instances Service registry lets clients discover available service instances.
load balancing 负载均衡 distribute traffic across instances Load balancing distributes traffic and improves availability.
serialization 序列化 convert objects to bytes RPC frameworks depend on efficient serialization and protocol design.
failover 故障转移 switch to a healthy node Automatic failover reduces downtime when an instance fails.

11. 风控业务关键词

English 中文 常用意群 面试短句
risk control 风控 control risk without blocking good users Risk control should reduce losses without hurting legitimate users.
fraud detection 欺诈检测 identify suspicious behavior Fraud detection combines rules, models, graph signals, and manual review.
payment fraud detection 支付欺诈检测 detect fraudulent payment attempts I designed a payment fraud detection engine that balances fraud catch rate with user experience.
coupon abuse 羊毛党 abuse promotions and coupons Coupon abuse should be detected through device, account, IP, behavior, and graph signals.
marketing anti-fraud 营销反作弊 prevent promotion abuse Marketing anti-fraud focuses on account farms, device farms, and abnormal behavior patterns.
application fraud 申请欺诈 fraudulent loan or credit application Application fraud often involves fake identities, forged documents, or organized intermediaries.
account takeover 账户盗用 attacker controls a legitimate account For ATO detection, I combine device fingerprinting, behavior sequence analysis, and step-up authentication.
AML 反洗钱 anti-money laundering AML systems monitor suspicious transactions and support investigation workflows.
suspicious transaction monitoring 可疑交易监测 monitor abnormal transaction patterns Suspicious transaction monitoring uses rules, thresholds, and network analysis.
sanctions screening 制裁名单筛查 screen against sanctions lists Cross-border payment risk requires sanctions screening, KYC checks, and compliance controls.
CDD 客户尽职调查 customer due diligence CDD verifies customer identity and risk level before or during service usage.
KYC 了解你的客户 know your customer KYC is the identity verification foundation for financial risk control.
liveness detection 活体检测 verify a real live person Liveness detection reduces fake identity and replay attacks.
device fingerprint 设备指纹 identify device characteristics Device fingerprinting helps detect account farms and account takeover.
behavioral biometrics 行为生物特征 behavior-based identity signals Behavioral biometrics can detect subtle changes in login and transaction behavior.
step-up authentication 阶梯式验证 require stronger verification for risky actions Step-up authentication adds friction only when risk is high.
MFA 多因素认证 multi-factor authentication MFA reduces account takeover risk but increases user friction.
first-party fraud 第一方欺诈 user abuses service under own identity First-party fraud is difficult because the user may look legitimate at onboarding.
friendly fraud 友好欺诈 chargeback abuse by real buyer Friendly fraud is hard because the buyer may be legitimate before abusing chargebacks.
chargeback abuse 退款滥用 abusive dispute or refund behavior Chargeback abuse requires transaction, logistics, and buyer behavior analysis.
fraud ring 欺诈团伙 connected group of risky entities Graph analytics helps uncover fraud rings invisible from isolated user-level features.
false positive rate 误报率 legitimate users incorrectly blocked Every false positive blocks a legitimate user, so risk strategy must control user friction.
false negative rate 漏报率 fraud cases missed by the system False negatives create direct fraud loss.
pass rate 通过率 percentage of approved transactions A good strategy improves fraud control while keeping the pass rate stable.
fraud rate 欺诈率 proportion of fraudulent cases Fraud rate should be monitored by product, channel, and user segment.
graduated limits 阶梯额度 increase limits as trust grows Graduated limits are useful for cold-start users with limited history.

12. 风控架构与数据关键词

English 中文 常用意群 面试短句
real-time risk engine 实时风控引擎 make decisions in milliseconds A real-time risk engine must keep feature fetching, rules, and scoring within the SLA.
decision engine 决策引擎 execute risk decisions The decision engine evaluates each request through rules, scorecards, models, and post-processing.
rule engine 规则引擎 evaluate business rules The rule engine should support versioning, hot reload, rollback, and audit logs.
decision flow orchestration 决策流编排 orchestrate rules, models, and actions Decision flow orchestration coordinates rules, scorecards, ML models, and manual review decisions.
feature service 特征服务 serve real-time features The feature service must provide low-latency and consistent feature values.
feature platform 特征平台 manage feature definitions and serving A feature platform is the data contract between rules, models, and the decision engine.
feature store 特征存储 store and serve model features A feature store supports both offline training and online serving.
online serving 在线服务 serve data for live decisions Online serving requires low latency, high availability, and strict version control.
offline computation 离线计算 batch compute features Offline computation usually handles historical aggregates and training datasets.
real-time feature 实时特征 feature computed from fresh events Real-time features capture recent behavior such as failed logins or payment attempts.
offline feature 离线特征 feature computed from historical data Offline features capture long-term behavior and stable risk profiles.
point-in-time correctness 时间点正确性 no future information leakage Point-in-time correctness prevents future information from leaking into model training.
training-serving skew 训练-推理偏差 mismatch between training and production features Shared feature definitions reduce training-serving skew between offline training and online inference.
feature join 特征关联 combine labels and features Feature joins must respect event time to avoid data leakage.
data pipeline 数据管道 move and transform data The data pipeline should support streaming, batch, quality checks, and lineage.
event stream 事件流 ordered stream of events Risk systems consume event streams from login, payment, device, and behavior events.
CEP 复杂事件处理 detect patterns across events Flink CEP detects patterns such as repeated failed logins followed by a new-device payment.
sliding window 滑动窗口 compute metrics over recent time Sliding windows are useful for velocity features and rate limits.
hot path 热路径 latency-critical path The hot path should include only operations required for real-time decisions.
cold path 冷路径 offline or asynchronous path The cold path can handle heavy analysis, reporting, and model retraining.
SLA 服务等级协议 latency and availability target The risk engine must meet the SLA even during traffic spikes.
degradation 降级 reduce functionality under failure Degradation keeps core decisions available when dependencies fail.
graceful degradation 优雅降级 fallback without hard failure If model inference is unavailable, the system should gracefully degrade to rules.
circuit breaker 熔断器 isolate failing dependency Circuit breakers protect the decision engine from cascading failures.
fallback policy 兜底策略 backup decision logic Fallback policies must be conservative and auditable.
multi-datacenter 多机房 deploy across data centers Multi-datacenter deployment improves availability but complicates consistency.
disaster recovery 容灾 recover from major failures Disaster recovery requires backup data, failover drills, and clear RTO and RPO targets.
shadow mode 影子模式 run without affecting live decisions In shadow mode, the new strategy runs in parallel but does not affect live decisions.
canary release 金丝雀发布 release to small traffic first Canary release reduces risk by exposing a new strategy to a small traffic slice first.
traffic splitting 流量分桶 split traffic for experiments Traffic splitting must be stable and unbiased for valid experiments.
graph analytics 图分析 analyze relationships between entities Graph analytics reveals fraud networks across users, devices, IPs, and accounts.
community detection 社区发现 detect dense connected groups Community detection helps identify fraud rings and account farms.
centrality 中心性 importance of a node in a graph Centrality can identify key nodes in a fraud network.
device fingerprint clustering 设备指纹聚类 group similar devices Device fingerprint clustering helps find device farms.

13. 风控模型与策略关键词

English 中文 常用意群 面试短句
scorecard 评分卡 interpretable baseline model A scorecard is interpretable and works well as a baseline for credit risk decisions.
logistic regression 逻辑回归 linear model for probability estimation Logistic regression is simple, stable, and easy to explain.
WOE 证据权重 weight of evidence WOE transforms binned variables into values suitable for logistic regression.
IV 信息值 information value IV is commonly used for feature screening in scorecard development.
PSI 群体稳定性指数 population stability index PSI monitors whether the population distribution has shifted.
CSI 特征稳定性指数 characteristic stability index CSI monitors stability at the individual feature level.
KS KS 统计量 measure model discrimination KS measures how well the model separates good and bad samples.
concept drift 概念漂移 relationship between features and labels changes Concept drift means the relationship between features and fraud labels has changed.
covariate shift 协变量偏移 feature distribution changes Covariate shift often happens when marketing brings in a new customer segment.
champion-challenger 冠军挑战者 compare new strategy with current production strategy Champion-challenger testing compares a new strategy against the current production strategy with controlled traffic.
A/B testing AB 实验 controlled experiment A/B testing measures the real business impact of a strategy change.
statistical significance 统计显著性 result is unlikely due to chance We need statistical significance before rolling out a new strategy.
threshold selection 阈值选择 choose decision cutoff Threshold selection should be based on a cost matrix, not only model accuracy.
cost matrix 代价矩阵 cost of each decision outcome A cost matrix makes the trade-off between false positives and false negatives explicit.
profit curve 利润曲线 business value across thresholds Profit curves help choose a threshold that maximizes expected business value.
precision 精确率 share of predicted positives that are true positives Precision matters when manual review capacity is limited.
recall 召回率 share of true positives caught Recall matters when missing fraud is expensive.
PR-AUC PR 曲线面积 area under precision-recall curve PR-AUC is more informative than ROC-AUC for highly imbalanced fraud data.
XGBoost 梯度提升树模型 gradient boosting decision trees XGBoost performs well on tabular risk-control features.
LightGBM 轻量梯度提升树 efficient gradient boosting LightGBM is efficient for large-scale tabular datasets.
feature importance 特征重要性 contribution of features to model Feature importance helps identify useful signals, but it is not a full explanation.
SHAP SHAP 特征贡献值 per-feature contribution explanation SHAP values help explain why a transaction or application was rejected.
LIME 局部可解释模型无关解释 local model-agnostic explanation LIME explains a single prediction by fitting a local surrogate model.
model interpretability 模型可解释性 explain model behavior Model interpretability is required for auditability, manual review, and regulatory compliance.
Explainable AI 可解释 AI AI that can explain decisions Explainable AI helps risk teams justify automated decisions.
GNN 图神经网络 graph neural network GNNs are useful when fraud risk propagates through user-device-IP-account relationships.
heterogeneous graph 异构图 graph with multiple node and edge types A heterogeneous graph can model users, devices, cards, IPs, and merchants together.
GraphSAGE 图采样聚合模型 inductive graph representation learning GraphSAGE is useful when new nodes appear continuously.
cold start 冷启动 little or no historical data For cold-start users, I rely on device signals, expert rules, transfer learning, and graduated limits.
transfer learning 迁移学习 reuse knowledge from another domain Transfer learning can help when the new product line has limited labels.
semi-supervised learning 半监督学习 learn from labeled and unlabeled data Semi-supervised learning is useful when fraud labels are sparse.
federated learning 联邦学习 train without centralizing raw data Federated learning can reduce privacy risk when data cannot be centralized.
MLOps 机器学习运维 manage model lifecycle in production MLOps covers model training, validation, deployment, monitoring, and rollback.
model governance 模型治理 control model risk and compliance Model governance ensures models are validated, monitored, documented, and auditable.

14. 易混词和发音提示

English 中文 注意点
idempotent 幂等 发音近似 eye-dem-po-tent；不要说成 independent
latency 延迟 指响应时间，不是吞吐量
throughput 吞吐量 指单位时间处理量，不是单次请求速度
consistency 一致性 分布式里通常指数据一致，不是代码风格一致
availability 可用性 强调系统能响应请求
reliability 可靠性 强调长期正确运行、少故障
scalability 可扩展性 强调流量或数据增长时能扩容
durability 持久性 数据落盘后不丢
isolation 隔离性 事务之间互不干扰
authentication 认证 你是谁
authorization 授权 你能做什么
fraud 欺诈 泛指欺诈行为
risk 风险 范围更广，可以是欺诈、信用、合规、运营风险
explainability 可解释性 偏“能解释单次决策”
interpretability 可解释性 偏“模型整体容易理解”
drift 漂移 分布或关系发生变化
skew 偏差 训练和线上口径不一致时常说 training-serving skew
recall 召回率 抓到了多少真实坏样本
precision 精确率 抓出来的坏样本里有多少是真的
false positive 误报 好人被拦
false negative 漏报 坏人漏过

15. 快速口语模板

中文提示 English 模板
这个问题我会分三层讲 I would explain this in three layers: the mechanism, the trade-off, and the production practice.
先讲原理，再讲线上怎么用 I would first explain the underlying mechanism, then discuss how I would use it in production.
这里最重要的是取舍 The most important part here is the trade-off.
这不是一个单纯的技术问题 This is not purely a technical problem; it also depends on business constraints.
如果是金融场景，我会优先正确性 For financial scenarios, I would prioritize correctness and auditability first.
如果是高并发场景，我会关注 P99 For high-concurrency scenarios, I would focus on P99 latency and backpressure.
如果依赖服务不可用，我会降级 If a dependency becomes unavailable, I would degrade gracefully with a fallback policy.
我会先看指标，再看日志，最后抓现场 I would start with metrics, then check logs, and finally collect runtime evidence such as traces or dumps.
我的经验是不要只看平均值 My experience is that average latency is not enough; we need to look at tail latency.
最后要有监控、报警和回滚 Finally, we need monitoring, alerting, and rollback.

16. 关联材料

  • 英文面试意群：按场景背诵整句表达。
  • Java 后端八股：Java 基础、并发、JVM、Spring、MySQL、Redis、消息队列、分布式系统。
  • 风控面试题：业务风控、技术架构、模型策略。
  • 刷题进度：刷题完成度跟踪。
  • 2026-06-08 大厂面经数据源：面经聚合、JavaGuide、风控系统设计资料。

17. 外部数据源链接

来源 链接 本文使用方式
面灵面经聚合页 mj.mianlingai.com 识别 2026 年 Java/AI 后端面经中的公司追问模式
字节跳动后端开发面经 字节 Java 后端 2026-04 补充 Redis 限流、多机房一致性、AOP、Spring、HTTP/HTTPS 追问
美团小象超市后端面经 美团后端实习 2026-03 补充库存扣减、接口超时、跨分片分页、MQ 不丢、项目深挖
JavaGuide 面试指南 javaguide.cn 校准 Java 后端高频知识覆盖面
GankInterview 风控/反欺诈 AI 面试 gankinterview.com 补充风控面试的 features / rules / models / evaluation / online strategy 结构
TechInterview Fraud Detection System Design techinterview.org 补充 rule engine、feature store、feedback loop、precision/recall/cost function
System Design Interview Handbook systemdesigninterview.com 补充 feature store、point-in-time correctness、training-serving skew、data leakage