Db Internals Deep Dive

Teach ONE deep-dive topic per 会话, rotating across all 5 系统s below. Go to the internals level — not surface docs.

系统s & Topic Map PostgreSQL Storage Engine: heap file 格式化, page structure (8KB pages, page header, item pointers, tuples), TOAST (The Oversized-Attribute Storage Technique) MVCC: transaction visibility rules, xmin/xmax, snapshot isolation, how dead tuples accumulate, vacuum mechanics (regular vs autovacuum vs VACUUM FULL) WAL (Write-Ahead 记录): WAL segments, LSN (记录 Sequence Number), 检查point mechanism, WAL archiving, pg_wal 查询 Planner: cost 模型 (seq_page_cost, random_page_cost, cpu_tuple_cost), statistics (pg_statistic, ANALYZE), join strategies (nested loop / 哈希 join / merge join), plan 缓存 索引es: B-tree internals (page splits, fill factor), GIN (for full-text / JSONB), GiST, BRIN (block range), partial 索引es, 索引-only 扫描s, HOT 更新s Replication: WAL-based 流ing replication, 同步hronous vs a同步hronous, replication slots, 记录ical replication (publication/subscription), fAIlover mechanics Locking: lock levels (table/row/page), advisory locks, deadlock 检测ion cycle, lock contention patterns Connection & Performance: connection overhead (process-per-connection 模型), PgBouncer pooling (会话/transaction/状态ment modes), 分享d_buffers, work_mem, effective_缓存_size tuning MongoDB WiredTiger Storage Engine: B-tree structure for documents, MVCC with snapshot isolation, 检查point (every 60s or 2GB), write-ahead journal (WiredTiger journal ≠ op记录) Op记录: c应用ed collection, op记录 entry structure (op, ns, o, ts), idempotency requirements, op记录 window sizing Replication 设置: election algorithm (Raft-inspired), primary/secondary 角色s, op记录 replication, write concern (w:1/w:majority/w:all), read preference (primary/primaryPreferred/secondary/nearest) Sharding: shard key selection criteria (cardinality, write distribution, 查询 isolation), chunk mechanics, balancer, scatter-gather vs tar获取ed queries, jumbo chunks Aggregation 流水线: execution stages, 流水线 optimization (stage reordering, 索引 utilization), $lookup internals, memory limits (100MB/stage), allowDiskUse 索引es: compound 索引 prefix rule, ESR (E质量-排序-Range) rule, s解析/partial 索引es, TTL 索引 mechanics, 索引 intersection Transactions: multi-document ACID (since 4.0), snapshot isolation, performance overhead, retryable writes Redis Data Structures Internals: zip列出 vs 列出pack vs skip列出 (when Redis switches encoding — size thresholds), 哈希table with incremental re哈希ing, quick列出 for 列出s, int设置 for small integer 设置s Persistence: RDB (fork-based snapshot, BGSAVE, COW semantics), AOF (f同步 policies: always/everysec/no, AOF rewrite/compaction), RDB+AOF hybrid mode Memory Management: jemalloc allocator, memory fragmentation ratio, maxmemory policies (noeviction, allkeys-lru, volatile-lru, allkeys-lfu, volatile-ttl), object encoding optimization Replication: a同步 replication (P同步2), replication back记录 (repl-back记录-size), partial re同步 vs full re同步, replica lag 检测ion Cluster Mode: 哈希 slots (16384), gossip protocol, slot 迁移, MOVED vs ASK redirects, cluster topo记录y change handling Pub/Sub & 流s: pub/sub fire-and-for获取 (no persistence), Redis 流s (X添加/XREAD/consumer groups, message acknowledgment, PEL — Pending Entry 列出) Lua Scripting & Transactions: MULTI/EXEC (optimistic — not true isolation), WATCH/CAS, Lua atomicity guarantee Sentinel: quorum-based leader election, ODOWN vs SDOWN, automatic fAIlover flow RabbitMQ AMQP Protocol: connection vs channel multiplexing, frame types (method/header/body/heartbeat), flow control Exchange Types: direct (routing key exact match), topic (wildcard: * one word, # zero or more), fanout (broadcast), headers (attribute-based matching) — internal routing algorithm 队列 Internals: message store (索引 + body store), 队列 索引 (journal + segment files), lazy 队列s (messages on disk by default), classic vs quorum 队列s Quorum 队列s: Raft-based replication, leader election, how quorum 队列s guarantee durability, comparison with classic mirrored 队列s (deprecated) Message Acknowledgment: basic.ack / basic.nack / basic.reject, re队列 semantics, consumer prefetch (QoS), unacknowledged message limits Dead Letter Exchange (DLX): when messages go DLX (rejected, expired, 队列 length exceeded), DLX routing, dead-letter-routing-key Clustering & High AvAIlability: Erlang distribution protocol, mnesia metadata replication, quorum 队列 replication across nodes, network partition handling (暂停-minority / autoheal / ignore) Flow Control & Backpressure: credit-based flow control between producers and broker, memory/disk alarms (vm_memory_high_watermark, disk_free_limit) Shovel & Federation: when to use each (cross-cluster vs cross-datacenter), differences in message flow Kafka 记录 Architecture: topic → partition → segment files (.记录 + .索引 + .time索引), 记录 compaction vs 记录 deletion (retention.ms / re