Guardrails for Analytics Agents That Do More Than Answer Questions
The risk isn't agents going rogue, it's agents acting correctly on bad input at machine speed.
The archive
Writing, page 14
652 posts on Apache Iceberg, lakehouse architecture, data engineering and applied AI.
Building Agent Telemetry Tables in Iceberg That Survive an Audit
A practical guide to building agent decision traces in Apache Iceberg that support audit reconstruction, governance review, and cost attribution.
What Agentic Analytics Actually Costs, and How to Keep It Bounded
Agent analytics generates two cost streams that scale on different variables. Here's the arithmetic, the levers that actually move the number, and how.
Running an Apache Iceberg Lakehouse With No Internet Connection
A practical guide to deploying an Iceberg lakehouse in air-gapped environments: component choices, artifact pipelines, identity without a cloud.
When the Query Optimizer Starts Managing Its Own Materializations
Autonomous materialized view management replaces quarterly review meetings with workload-driven scoring, and it's essential when AI agents generate.
Why AI Agents Fail on Raw Data, and What to Give Them Instead
Agents fail on raw lake data because business rules live in people's heads. Data products with semantic contracts fix this at the source.
Why Iceberg V4 Wants to Retire Equality Deletes, and What Streaming Teams Should Do About It
Equality deletes made streaming upserts into Iceberg practical at the cost of read performance.
The Five Layers Between Your Lakehouse and a Trustworthy Agent
Agent reliability is a property of the stack the model sits on. Five layers with distinct owners and failure modes turn the agent is unreliable.
Apache Fluss and Kafka Solve Different Problems in an Iceberg Pipeline
Fluss puts a columnar, indexed hot tier between Kafka and Iceberg. Here's what it changes structurally, what Kafka still does better, and how to benchmark.
Serving Sub-Second Queries Over an Iceberg Lakehouse With a Hot Tier
A lakehouse cannot serve sub-second queries over seconds-old data. A hot tier in front solves it, with consequences for consistency, governance.