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.
A PostgreSQL row is limited to 8KB, so how can TEXT or JSONB hold megabytes of data? TOAST automatically compresses oversized values and moves them to a separate table. This guide explains its four strategies (PLAIN/EXTERNAL/EXTENDED/MAIN), query-performance costs, when to tune storage, and best practices for designing large fields.
What was this price last week? Ordinary tables keep only current values; temporal tables retain history. This article explains audit logs, versioned rows, valid-time ranges, PostgreSQL temporal extensions, the SQL:2011 standard, CDC alternatives, and when temporal tables are warranted versus when an audit log is enough.
HTTP 200 does not prove that AI output remains usable. This article turns fixed tasks into low-risk black-box probes, covering scoring contracts, isolated execution, OpenTelemetry metrics, Prometheus alerts, and false-positive rollback.
Head-based random sampling often drops low-frequency model failures and long-running Agent anomalies. This article covers Trace aggregation, tail-sampling policies, Collector capacity sizing, sampling validation, and overload rollback procedures.
Deduplicating only after a random split leaves content from the same source on both sides of the training and evaluation sets. This article uses stable fingerprints, near-duplicate connected components, and evaluation isolation watermarks to control duplication and benchmark leakage.
Successful field deserialization does not mean training semantics are compatible. This article uses stable field identities, read/write matrices, shadow transformations, and distribution gates to control feature and label Schema evolution.
Deleting records from a source table does not automatically alter training shards, checkpoints, or released models. This article uses irreversible tombstones, reverse lineage, and artifact disposition gates to prevent revoked data from continuing to enter training.
A dataset can keep changing under the same name during training. Immutable membership manifests, source watermarks, transform digests, and derived lineage make each training input set exactly reproducible.
Even when every shard validates successfully, an entire file may still be missing or workers may read different versions. This article uses descriptors, set commitments, publication barriers, and generation pointers to keep training inputs complete and consistent.
Writing business rows and events separately to a database and message system creates a dual-write window, while a send timeout cannot prove that the message was not delivered. This article combines outbox events, dispatcher claims, unknown-delivery reconciliation, and consumer deduplication into an auditable protocol.
A 2xx response proves only that an endpoint successfully answered an HTTP request, not that the subscriber completed business processing. This article uses subscription generations, a signing-key state machine, immutable delivery evidence, and optional processing receipts to govern retries and rotation.