DevOps、Docker、Kubernetes、可观测性
This collection features articles and community discussions in the 云原生 space, focusing on real-world engineering scenarios, tool selection, practical paths, and risk boundaries.
The page is continuously updated and serves as a learning roadmap, technology selection reference, and technical review entry point.
When one Redis instance can no longer meet memory or throughput needs, Redis Cluster scales horizontally. This article explains its 16,384-slot and hash-tag sharding model, primary-replica replication, automatic failover, MOVED/ASK client routing, cross-slot limits, and when to choose Cluster over Sentinel.
Indexes require ongoing maintenance: they bloat, become obsolete, and consume storage. Learn to detect bloat with pgstattuple, find unused indexes through pg_stat_user_indexes, rebuild online with REINDEX CONCURRENTLY, and establish regular index health checks.
A request's cost can only be estimated at admission and known after completion. If gateways deduct independently and refund on timeout, capacity can be spent twice. This article uses reservation receipts, settlement, and expiry races to maintain rate-limit conservation.
K8s allows all Pods to communicate by default, which is a production security risk. NetworkPolicy lets you declare which Pods may communicate. This article explains ingress and egress rules, default-deny mode, and zero-trust K8s networking practices.
Large tables slowing down? Partitioning splits one physical table into multiple child tables so queries scan only relevant partitions. This article explains PostgreSQL Range, List, and Hash strategies, partition pruning, when to partition at tens of millions of rows, how to choose a partition key, and how partitioning relates to indexes.
Already use PostgreSQL? Advisory Locks provide application-level coordination without Redis. This guide covers session- and transaction-level locks, blocking vs. try-lock APIs, key design, task scheduling, migrations, idempotency, row-lock differences, and production pitfalls.
A compatible result from the registry does not mean every production consumer can interpret new fields correctly. This article binds the real consumer inventory, Avro reader/writer rules, historical data samples, and upgrade watermarks, then switches producers according to compatibility direction.
Configuration should not be hard-coded into images: ConfigMaps store nonsensitive settings, while Secrets store credentials. This article covers creation and mounting, env vs. volume, base64 and encryption with etcd, KMS, and external operators, ConfigMap hot updates and the subPath trap, and production secret management with External Secrets Operator.
Setting a five-second timeout on every service lets the full call chain exceed the user's budget, while a client disconnect does not mean downstream work has stopped. This article defines trusted ingress deadlines, hop-by-hop remaining budgets, shared retry allowances, and verifiable cancellation convergence.
Why does PostgreSQL choose a Seq Scan over an Index Scan? Its planner uses statistics to estimate which is faster, and inaccurate statistics produce bad plans. This article explains how ANALYZE collects statistics, how they guide planning, when stale data causes slow queries, and how to intervene with statistics targets and extended statistics.
A watermark crossing a window only means the system believes most events have arrived; late facts cannot be discarded. This article versions timestamp extraction, idle partitions, allowed lateness, and correction semantics so reports and external effects can be recomputed without being executed twice.
After read/write splitting, users "can't read what they just wrote"—this is replication lag. This article explains why lag occurs (async replication network/disk/transactions), how to monitor it (pg_stat_replication), how to mitigate it (read/write consistency window/synchronous replication/reading the primary on critical paths), and the application-level