Insights

Notes on the decisions that get made by default

These are working notes on the parts of platform work that get settled quickly and cost time later: how change data capture handles deletes, what reconciliation does to a migration date, and how a row access policy behaves once data crosses an account boundary. We write them for the engineers who will operate the result.

24 October 202610 min read
Before upgrading an Iceberg table to format v3
Snowflake's in-place upgrade to Iceberg format v3 cannot be reversed, and engine support for v3 is still uneven. Leave a table on v2 until it needs a v3 feature, and confirm every reader and writer before upgrading it.
- Data Engineering
- Snowflake
- Iceberg
24 October 20268 min read
Where metric definitions live for AI agents
Snowflake, Databricks and Microsoft each let an agent read governed metric definitions, and each tests the agent's answers in its own way. Choose one system of record per metric and keep a question set of your own.
- Governance
- Semantic Layer
- AI
24 October 20268 min read
Snowflake AI features and the rights they run with
Cortex Agents run generated SQL with the caller's roles, but a Cortex Search service returns what its owner's role could read. Record the rights model of every tool on an agent before the agent reaches governed data.
- Governance
- Snowflake
- AI
12 August 20264 min read
Row access policies and the sharing boundary
A policy that filters on the current user behaves differently once the data is reached through a share. Confirm which execution context applies on every access path before relying on the policy to separate tenants.
- Governance
- Snowflake
- Security
6 August 20265 min read
Choosing a change data capture strategy
Log-based, query-based and full-refresh capture each fail in a different way. The fastest way to narrow the choice is to ask what happens in the warehouse when a row is deleted at source.
- Data Engineering
- CDC
- Pipelines
28 July 20265 min read
A migration timeline is a reconciliation timeline
Conversion work can be estimated from a count of objects. Reconciliation is estimated from the differences it turns up, and it is the work that sets the cutover date.
- Migration
- Snowflake
- Reconciliation

Notes on the decisions that get made by default

Before upgrading an Iceberg table to format v3

Where metric definitions live for AI agents

Snowflake AI features and the rights they run with

Row access policies and the sharing boundary

Choosing a change data capture strategy

A migration timeline is a reconciliation timeline