cellforge
A spreadsheet that stays fast at 100M rows
- Role
- Personal project
- Year
- 2026
- Built with
- TypeScript · Canvas · Web Workers
Overview
Excel stops at about a million rows and Google Sheets at 10 million cells. I wanted to know whether that ceiling is really necessary, or just a side effect of how spreadsheets are built. cellforge is a spreadsheet that runs entirely in the browser and scrolls, sorts and computes over 100 million rows. No server is involved, so your data never leaves your machine.
The question
Once a dataset is too big for a spreadsheet, people move it into SQL or pandas. I wanted the spreadsheet to keep up instead. That meant scrolling 100M rows smoothly, opening big CSVs without freezing the page, and making formulas, fill-down, sort, copy/paste and undo work at that size.
Columns, not cells
Data is stored by column, not as a grid of cell objects. A column of 100M numbers is one Float64Array, and text columns are dictionary-encoded. Edits sit in a sparse overlay on top of the imported data, so changing a few cells copies nothing. The grid is a single virtualized canvas that only draws what's on screen. CSV parsing streams the file in a Web Worker.
Compiling formulas
A formula filled down a column is really one vector operation, so cellforge compiles it into a single loop over typed arrays. The interpreter is still the source of truth. Compiled results are bit-identical to it, and a fuzz test over 250K random formulas checks that. In the README benchmark at 16M rows, compiled formulas ran 6–22× faster than interpreted ones.
What didn't work
I also tried an experimental WebGPU backend. Its compute is very fast, but f32 precision isn't safe for spreadsheets: cancellation produced errors of up to 32%. I wrote the results up in the repo instead of shipping it.
Tech Stack
- TypeScript
- Canvas
- Web Workers
- WebGPU
- Vite
- Vitest