Comment by simonh | Hacker News Reader

Comment by simonh | Hacker News Reader

Parent Full thread

I used to work at a bank that built a huge integrated application development and deployment stack on Python. One aspect of it was a custom library implementing a Directed Acyclic Graph, basically a way to ink up properties so that when the results of one property changed, any dependent properties on the DAG would automatically recalculate. This was explicitly implemented so that we could directly translate the logic from Excel sheets into Python code.

A lot of our projects were built to accept an excel sheet as input, then generate another Excel sheet as the output. The idea was the logic and data manipulation and transformation should be in source control, but the Quants would still have Excel available to build graphs, pivot tables and do statistical analysis on the results.

The DAG also turned out to be pretty handy for implementing web applications and all sorts of other apps though.

We also found using Excel as an output to be a powerful design pattern. Excel is a great language for explaining the flow of calculations. We used to have WPF screens where a portfolio manager / trader / analyst could right-click on a number and select "show details" and he/she would get an Excel workbook where that same number would be selected and they could see the derivation by following the cell's precedents. It was not always feasible to show a complete derivation, but we'd try to show a useful amount of elaboration. This was a heavily used and loved feature, often leading to the users making edits or additions to these detail sheets and giving it back to us as a spec. Building the 'details' Excel books programmatically was a lot of work, but part of the payoff was that we also used them as a nightly check on our primary system, i.e., we'd automatically run 'show details' on everything and make sure the 'detail' result matches the primary system result. That comparison caught many bugs. Kent Beck: "More important than tests vs types is the principle of double checking. If you say something twice in independently derived ways, you're more likely to be correct than if you just say it once. Tests are a form of double checking. So are types."

Sounds like Athena at JP, or Quartz at BAML. Those off graph errors drove me nuts.

Reply on news.ycombinator.com