hzerrad@insights:~$

9/23/2026 by Houssem Eddine Zerrad

A 4B-parameter model just reported a 1.81× speedup over PostgreSQL's planner. Before that becomes a reason to put a learned optimizer in front of your database, check whether your bad plan was a search failure at all. In the incident behind this post, the planner expected 1,700 rows and got 16 million. It picked a nested loop, and for 1,700 rows that was the right call. Smarter search cannot fix a plan built on the wrong number. Two checks to run first, and why the fix was `ANALYZE`, not another optimizer.

9/15/2026 by Houssem Eddine Zerrad

Most sharding proposals arrive with a latency graph, not an `EXPLAIN ANALYZE`. A misestimated query and a cluster that's genuinely out of capacity look identical on a dashboard, and only one of them gets fixed by adding nodes.

9/10/2026 by Houssem Eddine Zerrad

Production AI agents are not just smarter chatbots. They are long-running, stateful systems that must survive crashes, retries, partial failures, and human approvals without repeating irreversible actions. Building them reliably requires durable execution, idempotency, authorization, sandboxing, observability, and other principles familiar from distributed systems.