Repository navigation
Investigate comment and hyperlink loading strategies before 4.0 #783
Description
Comments and hyperlinks are currently loaded on demand and cached per worksheet. This avoids collecting things the caller doesn't need, but for XLS and SpreadsheetML it means scanning the sheet twice if both are requested. For XLSX/XLSB, comments are stored separately, so only hyperlinks require another sheet scan.
We could collect these during the initial scan in two-pass mode instead. But unlike most of what we collect there today, comments and links can contain quite a lot of text. Doing this automatically could add unnecessary allocations. Making it opt-in seems more reasonable.
SinglePassMode would still need lazy loading, since the metadata can appear after the cells it belongs to.
Activity
Sinkleberg commentedon Nov 10, 2026
I tested the opt-in approach on a synthetic 2,048-row SpreadsheetML sheet with dense comments and hyperlinks. When the caller requested both, collecting them during the initial scan saved two later scans (3.16 MB read from the input stream). The returned cells, comments and links matched the lazy reader.
The tradeoff showed up when the caller requested neither. Always collecting both raised managed allocations from 1.09 MB to 9.22 MB. Collecting only what was requested kept it at 1.09 MB.
That supports keeping lazy loading as the default and offering an opt-in for two-pass reads. One detail worth documenting is that an opted-in reader can report a malformed comment during initialization. Single-pass reads stayed lazy in this experiment.
I used Codex for the local prototype and tests on Windows. The comparison covers SpreadsheetML only.