Join us

LLMs Are Good at SQL. We Gave Ours Terabytes of CI Logs.

LLMs Are Good at SQL. We Gave Ours Terabytes of CI Logs.

Mendral's agent runs ad‑hoc SQL against compressed ClickHouse logs. It traces flaky tests across months and scans up to 4.3B rows per investigation.

They denormalize 48 metadata columns per log line. They compress 5.31 TiB down to ~154 GiB (~21 bytes/line) — a 35:1 ratio. That turns arbitrary filters into column predicates.

The pipeline uses materialized views, bloom and ngram indexes, and Inngest for durable execution. GitHub API throttling (~3 req/s, 4k spare/hr) keeps P95 ingest under 5 minutes.

System shift: Granting LLM agents direct SQL access to denormalized, columnar CI logs moves debugging out of fixed tool APIs and into ad‑hoc, queryable data stores.


Give a Pawfive to this post!


Only registered users can post comments. Please, login or signup.

Start writing about what excites you in tech — connect with developers, grow your voice, and get rewarded.

Join other developers and claim your FAUN.dev() account now!

Avatar

DevOpsLinks #DevOps

FAUN.dev()

@devopslinks
DevOps Weekly Newsletter, DevOpsLinks. Curated DevOps news, tutorials, tools and more!
Developer Influence
13

Influence

1

Total Hits

132

Posts