`officer::body_import_docx()` causes multi-minute hang on large documents with many tables

[scarlettsijiaw]

For a document size of 6.2 MB with 26 tables (it's an appendix, hence the number of tables), the render took ~11 minutes between when bookdown render completed and when csasdown::render() actually completed.

Looks like it's because fix_table_cell_styles_xml() is O(tables x doc_size).

[seananderson]

Thanks for reporting. I've fixed that now, but it seems that officer::body_import_docx() is also very slow on large documents. We'll have to switch to officer::body_add_docx() as is done in other parts of the package, but that will take a bit of fiddling. On my test, that took the rendering from ~60 seconds to < 1 second.

[seananderson]

@scarlettsijiaw can you let me know how much faster this document builds now? If you're able, I'd also be interested if you could send a copy of this larger document so I can speed test. Avoiding officer::body_import_docx() over here will take a bit of work, because all the page numbering gets messed up without it.

[scarlettsijiaw]

The speed doesn't seem improved, still 10+ mins, the number of tables didn't change but the file size is slightly bigger (6.8MB) from some figure edits. Would you be able to speed test with just the docx file or do you need the rmd and data files?

[seananderson]

If possible, a complete building project would be ideal. Doesn't have to be a real project — just something that demos the issue. I guess with a .docx I could at least reproduce some Rmd files with similar properties.

[scarlettsijiaw]

I invited you to a speed_test_example repo I made. It replicates the appendix but uses pre-generated figures and csvs instead of generating figures and tables in the report. The pre-generated version is faster but still takes about 4-7 minutes (still only counting the time between when rmarkdown render finishes and csasdown render finishes).

[seananderson]

Thanks so much for putting this together! This will be very helpful for seeing if I can optimize this. For reasons that I don't yet understand, the example repository only takes ~68 seconds to run csasdown::render() from start to finish on my laptop (M2 Macbook Pro) with the latest csasdown.

sessionInfo()
R version 4.5.2 (2025-10-31)
Platform: aarch64-apple-darwin20
Running under: macOS Sequoia 15.7.4

Matrix products: default
BLAS:   /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib 
LAPACK: /Library/Frameworks/R.framework/Versions/4.5-arm64/Resources/lib/libRlapack.dylib;  LAPACK version 3.12.1

[seananderson]

OK, it looks like it was building faster for me because I already had a version of csasdown loaded that used officer::cbody_add_docx() intead of officer::body_import_docx(). The test repo you supplied, @scarlettsijiaw , takes 65 seconds to build for me now from start to finish. Can you try it out with the latest version? Note that the one formatting issue I'm aware of that this creates is that all pages get Arabic numbering whereas the front matter is supposed to get separate Roman numerals. This means currently one would have to manually fix this at the final publishing stage. I've been unable to quickly fix this. Perhaps we will figure out a way. But increasing build speed by an order of magnitude seems more important for now.

Also, landscape pages seem to be working now. Where they already working in this test repo?

[scarlettsijiaw]

The latest version is significantly faster! Landscape didn't already work in the test repo.