Coredump
of some of my work
Blogs detailing the design, implementation, and evaluation of my systems, research, and open-source work.
Benchmarking Sirius, a GPU-native SQL engine
I benchmarked Sirius against DuckDB on 113 join-heavy queries, then fixed the empty-result and string-aggregation bugs that were blocking the full suite.
- 6× average hot-run speedup on JOB
- 113 join-heavy queries over 74M rows
- Two fixes merged upstream into Sirius
Comparing Linux's EEVDF and CFS schedulers under contention
I put Linux's EEVDF and CFS schedulers under controlled contention and found that prompt scheduling is not the same thing as useful progress.
- 4–6× lower p99.9 latency with EEVDF
- Up to 65% higher Redis GET throughput
- 10–15% more 7-Zip throughput with CFS
Privilege separation for a Gradescope autograder
I inherited a Gradescope autograder image that could write scores to Canvas, and found that the student notebooks it graded could reach the same authority. The investigation, the proof of concept, and the privilege-separated redesign that closed it.
- Grade-write credentials reachable by untrusted student code
- One notebook cell could leak secrets and hidden tests
- Closed with least privilege and trusted/untrusted process isolation
Four concurrency models for serving TUIs over SSH
textish serves Textual TUI apps over SSH. Getting there meant building three concurrency models — subprocess, subinterpreter, and free-threading — and discovering that the most powerful primitive was the wrong one.
- Three concurrency models built and compared
- Subprocess and subinterpreter, then a shared interpreter
- Thousands of SSH sessions, none with its own process or thread