A reader opens a URL and gets the pipeline itself: the built store, the values, the code, the docs — and can change the code and re-run it, with correct incremental rebuild. No R, no Docker, no server.
It works. Everything below was executed in headless Chromium against webR 0.6.0 (R 4.6.0), not inferred from documentation.
repo.r-wasm.org carries targets 1.12.0 for R 4.6 and every package tardoc
imports — a 48-package closure with nothing missing. Installing it takes about
four seconds in the browser.
One thing stopped it, and the same thing every time. targets reads
_targets.R in a subprocess by default, which keeps the pipeline’s environment
out of the caller’s session. wasm cannot spawn processes, so callr looks for
R.home("bin")/R, finds nothing, and fails:
| Call | Default | callr_function = NULL |
|---|---|---|
tar_manifest() |
✗ Cannot find R executable |
✓ |
tar_network() |
✗ | ✓ |
tar_outdated() |
✗ | ✓ |
tar_make() |
✗ | ✓ |
tar_meta(), tar_progress() |
✓ | — |
Nothing else about targets needed changing. The virtual filesystem, the store,
serialisation and the metadata all behave.
Yes — and the bundle is nothing. The example pipeline, source and built store together, is 26 KB gzipped. Untarred into webR’s filesystem:
files unpacked 14
tar_read(station_summary) 6 rows, 2 cols
tar_read(clean) station, temp_c, day_of_year
tar_outdated() none - store is current
tar_meta() rows 13
tar_make() no-op 7 skipped
So a visitor arrives at a pipeline that is already built, and can read any target’s value.
That is the part that matters, and yes. Editing a threshold in R/functions.R
from inside the browser:
before: clean rows 400
now outdated report, clean, station_summary, drift_model, drift_flags
rebuild 5 completed, 2 skipped
after: clean rows 326
now outdated again? none - rebuilt and current
targets invalidated exactly the five downstream targets, left the two upstream
ones alone, re-ran, and settled. Then document_targets() regenerated the docs
on top of the new state — 9 files, a 77 KB viewer.
This is not a demo of R in a browser. It is a targets pipeline behaving like a targets pipeline, published as static files.
The obvious worry: webR wants SharedArrayBuffer, which needs
Cross-Origin-Opener-Policy and Cross-Origin-Embedder-Policy headers, and
GitHub Pages cannot set headers.
Tested on a server sending neither:
crossOriginIsolated: false
SharedArrayBuffer : undefined
Everything above still ran — the shipped store, the edit, the rebuild, the docs. webR 0.6’s PostMessage channel covers this case. Cross-origin isolation makes it faster; it is not required.
| The project bundle | 26 KB |
R.wasm + BLAS/LAPACK + webR JS |
20.0 MB |
| R packages for tardoc’s stack (49) | 31.4 MB |
| Total, cold | 51.5 MB |
The publisher hosts the 26 KB and one HTML file. The 51.5 MB comes from r-wasm’s CDN on demand and caches, so hosting is free and the repository stays small.
A pipeline’s own packages add to it. Measured against the same index:
| Pipeline stack | Packages | Available |
|---|---|---|
| tardoc alone | 48 | all |
| + tidyverse core (ggplot2, tidyr, readr, …) | 70 | all — about 16 MB more |
| + modelling (broom, glmnet, randomForest) | 67 | all |
+ sf |
60 | all |
+ data.table |
48 | all |
+ arrow |
— | arrow unavailable |
+ rstan |
— | rstan unavailable |
So a tidyverse pipeline lands around 68 MB, and the practical rule is that most pipelines work while heavyweight compiled ones do not.
The right comparison is not tardoc’s 300 KB viewer — it is the other ways to hand someone a runnable pipeline. Binder needs a server and cold-starts in minutes. Docker needs Docker. “Install R and these fourteen packages” needs R. Against those, 51.5 MB of cached static files with no server is a good trade, and the 26 KB bundle means the project itself costs nothing to publish.
It is a bad trade only if all the reader wanted was the docs, which the existing viewer already delivers for 300 KB. These are two products, not one, and the existing one should not grow 51.5 MB to become the other.
Worth recording on its own, because it affects every tardoc user and nothing to do with webR.
load_targets_data() called targets::tar_config_set(store = cfg$targets_store).
cfg$targets_store is absolute, and tar_config_set() writes _targets.yaml
in the user’s project. So running tardoc once left a line like
main:
store: /home/someone/analysis/_targets
committed in their repository — breaking the project on every other machine, in
CI, and under webR, where it is what made tar_read() fail on a shipped store.
Reproduced on the desktop against a project that had no _targets.yaml before.
It was gratuitous as well as harmful: everything in load_targets_data() already
runs inside the project directory via withr::with_dir(), where the default
relative _targets is what targets would have used anyway. The store is now
passed per call — tar_network(), tar_meta(), tar_progress() and
tar_outdated() all take store; tar_manifest() needs none, reading
_targets.R alone — and no _targets.yaml is written.
Keep both changes. .callr_fn() and the store fix cost nothing on the
desktop, and the second is a bug fix regardless of any of this.
A publishing command is worth building, as its own thing. Something like
publish_webr() writing a directory of static files: an HTML page that boots
webR, the project tarball, and the viewer. Not a change to
document_targets() — the existing tiers stay as they are.
Two things to settle before building it, neither answered here: what the page should look like when webR is still loading twenty seconds in, and whether the reader gets an editor or only a re-run button. The second is a product question, not a technical one — everything needed for either already works.
The container blocks browser egress, so webR and the packages were vendored and
served locally. tardoc is pure R, so it installs into webR as a binary by
tarring an R CMD INSTALL tree — no compilation. Two gotchas cost time and are
worth knowing: untar() shells out by default and system() is unsupported
under Emscripten, so tar = "internal" is required; and webR’s browser entry
point is dist/webr.js, not dist/webr.mjs, which imports Node built-ins.