8 million rows in. 600,000 out. 5.7 seconds. On a 16GB RAM laptop.
Duckle joined 4 sources at 2M rows each - an ADBC (Arrow) source, a CSV file, a MySQL table, and a second ADBC source - through one visual mapper: a 3-way join, 9 expressions, and a filter, straight to Parquet.
No cloud. No servers. Just Duckle on your laptop/desktop.
This is what local-first data engineering looks like now.
It is becoming impossible to work with Opus 4.7. The calculations Opus 4.7 is making is not adding up and hence we feel like it is hallucinating. May be they tuned down RAG, context secretively.