We gave terabytes of CI logs to an LLM

4 hours ago

Agent traced a flaky test by scanning hundreds of millions of log lines in seconds.
Agent uses a SQL interface to query job metadata (63%) and raw log lines (37%).
Analyzed 8,534 agent sessions with 52,312 queries, averaging 4.4 queries per session.
Agent starts broad with job metadata queries, then drills into raw logs for detailed investigations.
ClickHouse stores 5.31 TiB of data (uncompressed) in 154 GiB (compressed) with a 35:1 compression ratio.
Denormalized metadata columns in ClickHouse compress efficiently, making queries fast.
Query latency scales linearly with rows scanned, from 10ms for <1K rows to 31s for 1B+ rows.
Ingestion pipeline throttles GitHub API requests to stay within rate limits while maintaining fresh data.
Durable execution engine (Inngest) handles rate limits by suspending and resuming workflows without crashes.
Mendral (YC W26) automates CI log analysis to identify failures and changes.

Hasty Briefsbeta