Corrections & changelog

2026-07-20 — review pass

A round of factual corrections following a technical review of the published articles.

• Home · Recent work fix

  "Velostrata-based workload transfers" listed under Azure was wrong — Velostrata is Google Cloud migration technology. Replaced with Azure Migrate's actual agentless/agent-based replication, and corrected the stale GCP program names (MCMA → Migration Center, Migrate to Virtual Machines).

• AI on Microsoft Fabric fix

  The AI Functions code used a non-existent AIFunctions.classify(df, …) module call. Replaced with the documented accessor df.ai.classify(labels=…, input_col=…, output_col=…), cited to Microsoft Learn. Also replaced "safe by construction" with the precise inherits-controls-but-still-validate wording.

• Fast GraphRAG Tools Compared · Build-Time Knowledge Graphs fix

  Both claimed adding a document to Microsoft GraphRAG "effectively requires a rebuild." That is outdated — graphrag update (v1.0+) does incremental indexing. Corrected, with the honest caveat that it still pays LLM extraction per new doc and can degrade to a full community recompute.

• Fabric vs Databricks clarify

  Distinguished classic Databricks compute (DBU + a separate Azure VM charge) from serverless (VM cost already inside the DBU rate). Narrowed "no data movement" between Lakehouse and Warehouse: cross-database reading is zero-copy, but materialising a Warehouse table still copies.

• Azure AI Foundry & Fabric Data Agent clarify

  Added the Fabric Data Agent's documented operational limits — 5 data sources, 25×25 result cap, same-region requirement, Fabric-native sources only, English-first, bounded analysis.

• Testing Data Pipelines fix

  Corrected "dbt unit tests run with no warehouse compute" — the compiled SQL still executes on the target platform over tiny synthetic input. It's minimal, controlled compute, not zero.

• VertiPaq Internals fix

  Repaired a confusing 2022/2025 chronology (the article claimed to be period-accurate while discussing 2025 Fabric). Added a provenance block with separate Published / Updated / Verified dates, and labelled the 200 MB→40 MB and V-Order figures as dataset-specific rather than universal.

• Real-Time AML on Kafka and Flink clarify

  Cited the specific FinCEN SAR filing timeline (30 days, 60 with no suspect); separated the AML stages (alert / investigate / escalate / file); reframed the sub-300 ms latency point as a property of this pipeline's shape rather than a general Flink limitation; and disclosed the "mid-sized bank" as an explicit composite.

• Build-Time Knowledge Graphs clarify

  Labelled the 1,000 / 100,000-document thresholds as heuristics, and softened over-absolute claims (a parser recovers structure "exactly" → noted dynamic dispatch and reflection are invisible to it).