Package management and reproducibility are a nightmare

I try my best. I really do.

I use renv for package management. My blog is built with Quarto. I use caching and freezing so posts don’t get recomputed every time. I hadn’t touched the blog in a few months. Today, I just wanted to fix a couple of typos on my publications page. Two minutes of work. Run quarto render. Done.

Except, of course, not done.

I open the project and renv tells me the project is out of sync. Cannot render because package Rmarkdown is missing. Fine. I run renv::restore(), whose entire purpose is to restore the package environment to its previous state.

It immediately fails while trying to install Matrix from source.

I’ve never had that problem before. A bit of googling reveals that installing Matrix requires a Fortran compiler (???). The official instructions are here: https://mac.r-project.org/tools/. I do not remember ever needing this in the past. I try to force renv to use binaries instead of building from source. That also fails.

Alright. Fine. I install a Fortran compiler. Matrix installs successfully.

One minute later, renv::restore() fails again—this time because it can’t install stringi. The error message is completely opaque. I paste it into an AI, which helpfully explains:

“This error occurs because of a conflict between the modern [[noreturn]] attribute in R 4.5 and an older NORET declaration within the stringi package, specifically on ARM64 macOS.”

Jesus fucking christ.

Yes, I updated R last month from 4.4.2 to 4.5. And now one of the most fundamental string manipulation packages in the entire R ecosystem—on which half the world depends—can’t be installed?

Okay. Fine. Surely I can just install a newer version of stringi. I stop the restore process and run:

renv::install("stringi")

That works.

But now I have a problem: if I run renv::restore() again, it will just try to install the older version and fail all over again. So I need to snapshot the current state.

I run renv::snapshot(), and it tells me:

The following required packages are not installed:
- ambient
- bmm
- fields
- ggdark
- ggplot2
- here
- httpgd
- kableExtra
- patchwork
- rmarkdown
- tidyr
- viridis
- wesanderson
Packages must first be installed before renv can snapshot them.

Why are they not installed? I’m on the same machine. Same project. I haven’t changed anything.

Oh. Right. I updated R. Since R is a global installation, renv considers this a fresh library and wants everything reinstalled.

Fine. I choose option 2: install the packages, then snapshot.

Everything installs—except ggdark.

Error: package 'ggdark' is not available

What.

A quick google search leads me to CRAN, which cheerfully explains:

Package ‘ggdark’ was removed from the CRAN repository.
Formerly available versions can be obtained from the archive.
Archived on 2025-11-07 as issues were not corrected despite reminders.

Oh for fuck’s sake.

I recognize this exact message. It recently bit me with another ggplot utility package, gghalves. ggplot just went through a major version bump to v4, and it broke a bunch of extension packages. If the maintainer doesn’t fix things quickly enough, CRAN just removes the package.

Unfortunately, R’s entire package ecosystem is built on the assumption that every package should work with the latest version of every other package. That assumption collapses completely the moment you care about reproducibility.

The whole point of renv is that I’ve already specified which versions of ggplot2 and ggdark this project used. When I chose “install the packages” during snapshotting, I assumed renv would install the versions recorded in the lockfile.

But something clearly didn’t work.

What exactly failed?

• Did renv try to download the correct version of ggdark, but CRAN refused because it’s archived?
• Or did renv ignore the lockfile at that point and try to install the latest version instead?

CRAN does link to the archived sources, so in principle the old version exists. But apparently not in a way that renv can use automatically.

I snapshot without installing the missing packages. That technically works, but I’m stuck with an inconsistent environment.

I’m not even using ggplot v4. This isn’t a usage problem. It’s an infrastructure problem: R’s dependency system forces compatibility with the present, and fights you when you try to live in the past.

Fine. I try installing from the archive manually:

renv::install("ggdark@0.2.1")

That works.

I try renv::snapshot() again.

Now it fails on httpgd.

Same story. Package removed from CRAN. Archived. The version in my lockfile is httpgd@2.0.2.

renv::install("httpgd@2.0.2")

Works.

Why am I doing this manually? Isn’t this exactly what renv is supposed to manage for me?

At this point I start wondering whether this has something to do with Quarto. Some posts use post-specific packages declared via renv::use("packagename@version"), as recommended in the documentation. Maybe those dependencies don’t fully participate in the main restore/snapshot logic? Maybe I’ve built myself a beautiful little Rube Goldberg machine of reproducibility.

One last attempt:

renv::snapshot()

Finally:

renv::status()
No issues found -- the project is in a consistent state.

What time is it?

Two hours have passed. I wanted to fix two typos.

I run quarto render.

Hallelujah—it works.

I need a drink.

Never mind. I forgot I don’t drink.