Notes and talks
Write-ups of papers I have presented or read closely, and reports from course and research projects. Each one tries to say what the work establishes and where I think it stops short.
-
Column stores: from C-Store to modern analytical databases
On C-Store (2005) and the two Abadi papers (2006, 2008). Which execution features explain the speedups, why the layout alone does not, and what Vertica, Snowflake, DuckDB and ClickHouse kept.
-
When is one query per template enough? Adapting learned cost models to new hardware
1,162 PostgreSQL queries timed with identical plans in six environments. One measured query per template, with a per-template offset, comes close to retraining; nothing helps on templates never measured.
-
Scaling laws for neural language models
On Kaplan et al. (2020). The fitted power laws, the batch-size correction that the compute result depends on, and why the paper's most quoted conclusion rests on its least robust number.