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.
Joining features only by entity and nearest timestamp can still leak data that arrived after the label into the training set. This article freezes event time, availability time, TTL, and snapshot watermarks while ensuring that offline training and online serving consume the same feature-definition generation.
Uncontrolled concurrent updates can cause dirty reads, phantom reads, and lost updates. This article explains ACID, the four isolation levels, PostgreSQL's MVCC implementation, the anomalies possible at each level, and how to choose: Read Committed is usually enough; use Serializable only for strong consistency.
SSH in and set up the environment by hand? Slow, unreproducible, error-prone. Ansible automates server management with declarative YAML over SSH, no agent. Covers core concepts (inventory/playbook/task/role), common modules, a hands-on Nginx playbook, and the split with Terraform (infrastructure vs config).
GPU capacity cannot be derived directly from request counts. This guide builds a load-testable, reviewable inference capacity model using input and output token distributions, concurrent sequences, KV Cache, batching efficiency, and failure redundancy.
Are queries and maintenance slowing down on a table with hundreds of millions of rows? Partitioning splits it into smaller pieces, scans only relevant partitions, and removes old data quickly. This guide covers RANGE, LIST, and HASH partitioning, setup and maintenance, partition pruning, and when to partition by time or category.
Reusing an Idempotency-Key does not mean the requests are identical, and caching one response cannot prevent concurrent double writes. This article defines a contract for operation identity, canonical payload fingerprints, atomic reservation, result replay, and reconciliation of unknown side effects.
"99.9% availability" is not a slogan. SLOs and error budgets should drive engineering decisions. This article explains the relationship between SLIs, SLOs, and SLAs, how to calculate and use error budgets, and how to manage reliability with data rather than intuition.
CPU metrics do not promptly reflect AI queue backlogs or model warm-up. This article implements backlog-based scaling with autoscaling/v2 external metrics, covering metric semantics, startup latency, stabilization windows, failure testing, and rollback.
A schedulable GPU count does not guarantee correct device, CPU, and memory topology. This guide separates Device Plugin and DRA responsibilities and covers API discovery, NUMA alignment, tenant isolation, validation, and rollback.
Inference requests may run for tens of seconds or stream tokens over long-lived connections. preStop or a longer grace period alone can still lose requests. This article connects readiness, EndpointSlice, SIGTERM, in-flight tracking, PDBs, failure testing, and rollback.
Storing everything in VARCHAR(255)? Using FLOAT for money? Wrong types lose precision, weaken indexes, and waste storage. This guide explains how to choose PostgreSQL types for numbers, text, UUIDs, time, booleans, enums, arrays, and JSONB, and why the right type improves both performance and correctness.
Large-model weights should not be downloaded without verification every time a Pod starts. This article compares init containers, shared volumes, and DaemonSet node caches, and provides validation workflows for immutable digests, atomic publication, capacity reclamation, and rollback.