mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-11-09 10:39:41 -06:00
Vijay Janapa Reddi
38439425a6
The repository carried the book twice. publishing/quarto/contents/vol{1..4}/
was what every build and every check read; books/vol{1..4}/ was a second
tracked copy that fed nothing and had drifted badly, holding 3 vol3 chapters
against 30 in the live tree. A `book` symlink pointed at publishing/, so
book/quarto/ and publishing/quarto/ were one directory under two names, next
to a books/ that was a third thing. Editorial work kept landing in the copy
nothing rendered.
books/ is now the Quarto project root and the only place book sources live:
books/vol1..vol4 one directory per volume
books/shared/ frontmatter, backmatter, _partials, assets, tex,
filters, scripts, audits, publish, calc
books/config/ the per-volume quarto configs
books/_extensions/ Quarto needs these at the project root
books/index-vol{1..4}.qmd, references*.bib, _variables.yml, 404.qmd
books/_build/ render output, gitignored and disposable
The toolchain stays in publishing/ and is now unambiguous, since it no longer
contains any of the book. The `book` symlink is gone and the CLI is reached as
./binder from the repository root; publishing/binder resolves through the
symlink so invoking it from either place works.
Config paths are relative to the project root, so a chapter entry goes from
contents/vol1/introduction/introduction.qmd to vol1/introduction/introduction.qmd,
and shared material picks up a shared/ prefix. 5,204 path references were
rewritten across 401 files, plus 127 segment-built paths in Python that string
matching could not see, plus 795 ./book/binder references.
Verified rather than assumed: a single-chapter HTML render produces output
equivalent to the same build on dev, 432,901 bytes against 433,158, the same
18 images, zero unresolved cross-references on both. Scoped checks return
identical findings on both trees.
One behaviour genuinely changes. While the sources sat under contents/, the
checks resolved their default scan root to the Quarto project directory rather
than the content directory, so a bare `./binder check tables` matched no files
and always reported success. Now that the project root and the content root
are the same directory the default works, and it immediately surfaces 31
pre-existing findings in vol3 and vol4 that were always present and never
reported. Those are content issues to fix separately, not regressions from
this move.
docs/REPO_LAYOUT.md describes the result. docs/BOOKS_MIRROR_RETIREMENT.md
records what the retired mirror held, which commits landed only there, and how
to recover any of it from history.
9 lines
231 B
YAML
9 lines
231 B
YAML
email:
|
|
contact: vj@eecs.harvard.edu
|
|
subject: [MLSys Book]
|
|
info: mailto:vj@eecs.harvard.edu?subject="CS249r%20MLSys%20with%20TinyML%20Book%20-%20"
|
|
|
|
title:
|
|
long: "Machine Learning Systems"
|
|
short: "Machine Learning Systems"
|