Core dump epidemiology: fixing an 18-year-old bug
OpenAI's own account: epidemiological core-dump analysis surfaces two unrelated crash causes.
2 items · 2 sources · 2 days
Operational story trace
Follow in this browser to see new updates on your Live feed.
Latest change
InfoQ's write-up details the fix: an 18-year-old GNU libunwind race condition with a roughly 100-picosecond vulnerable window, made real by OpenAI's frequent per-query timer signals, has been patched upstream by reordering the instruction- and stack-pointer updates.
OpenAI engineers spent months chasing mysterious crashes in Rockset, the C++ data infrastructure behind ChatGPT search, where functions returned to invalid addresses and stack pointers shifted unexpectedly. Instead of debugging crashes one at a time, they built a pipeline to mine a year of production core dumps for population-level patterns — "epidemiological debugging."
Arc
OpenAI's own account: epidemiological core-dump analysis surfaces two unrelated crash causes.
InfoQ adds the technical specifics: the exact race-condition window and how it was closed.
OpenAI's own account: epidemiological core-dump analysis surfaces two unrelated crash causes.
InfoQ adds the technical specifics: the exact race-condition window and how it was closed.
What to watch — open questions
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.