Deployment¶

Prerequisites¶

• A shinyapps.io account (free tier is fine).
• The artifacts produced by ./setup.sh (see Local development).

One-time: register your token¶

1. Sign in at shinyapps.io and go to Account → Tokens → Show.
2. Copy the rsconnect::setAccountInfo(...) snippet.

3. Run it in R:

   Rscript -e "rsconnect::setAccountInfo(name='<acct>', token='<TOKEN>', secret='<SECRET>')"
   

4. Verify:

   Rscript -e "rsconnect::accounts()"
   

Deploying¶

./deploy.sh

What it does:

1. Verifies Rscript, the rsconnect package, and your configured account.
2. Reads dictionary.parquet out of config.yml so it knows which release ships.
3. Checks all 15 required deploy artifacts exist (app + config + Python backend + model + the active parquet + both embeddings/metadata pairs + manifest).
4. Confirms data/embeddings/manifest.txt matches config.yml — refuses to ship a bundle that would crash on boot due to drift.
5. Cleans python/__pycache__/ (rsconnect's hardcoded __pycache__/ exclusion has a trailing-slash bug).
6. Calls rsconnect::listDeploymentFiles(), filters out any non-active dictionary parquet and stray release CSVs (.rscignore can't express "all except the active release" — no globs), and prints a bundle preview (top 10 files by size + total). Expected: ~16 files, ~125 MB for the 7.0 build.
7. Asks for confirmation (skipped if stdin isn't a TTY).

8. Calls:

   rsconnect::deployApp(
     appDir = '.',
     appFiles = files,                  # filtered file list
     appName = 'abcd-dictionary',
     appTitle = 'ABCD Dictionary Search',
     account = 'biplabendu',
     python = 'python_env/bin/python',   # <-- captures Python in manifest
     forceUpdate = TRUE,
     launch.browser = FALSE
   )
   

Configurable via env vars:

APP_NAME=abcd-dict-staging APP_TITLE='ABCD (staging)' ./deploy.sh
SHINYAPPS_ACCOUNT=myaccount ./deploy.sh                    # multi-account case

How Python is provisioned on shinyapps.io¶

This is the part most reticulate users get stuck on. The model that works in 2026:

1. deployApp(python = ...) tells rsconnect to write a python section into manifest.json with the local Python version + a pointer to requirements.txt.
2. shinyapps.io's Connect runtime ignores the manifest's Python info at build time (a quirk of the free tier) — so we don't rely on it.

3. At runtime, app.R calls reticulate::py_require(readLines("requirements.txt")) before source_python(). This triggers reticulate's uv-based auto-installer, which:

   • Downloads uv (~1 s)
   • Downloads a pre-built CPython 3.12 tarball from Posit's mirror (~2 s, 32.5 MB)
   • Resolves and installs all packages in requirements.txt (~5 s, ~30 MB total)

The whole boot — R packages, Python install, package install, ONNX model load — takes about 16 seconds on a fresh instance. Subsequent boots reuse the cached Python+packages.

What ships and what doesn't¶

Controlled by .rscignore (top level), data/.rscignore (per-subdir), and an in-deploy.sh filter that drops any non-active dd-abcd-*.parquet plus stray release CSVs (rsconnect uses exact-string setdiff() against directory listings, not gitignore-style globs — so trailing slashes, nested paths, and *.csv don't match). See Troubleshooting.

Top-level .rscignore excludes:

• sanity-chks/, .claude/, .dockerignore, Dockerfile, fly.toml
• setup.sh, run.sh, deploy.sh, clean_artifacts.sh
• *.Rproj, .RData
• Root-level duplicate / dummy CSVs

data/.rscignore excludes the source CSVs for the currently active build (build inputs only; the runtime uses Parquet and .npy/.npz).

deploy.sh additionally filters out:

• Any dd-abcd-*.parquet other than the one named in config.yml (e.g. an old 6.0 parquet left in data/ won't ship alongside 7.0).
• Any stray dd-abcd-*.csv at the repo root or under data/.

Final deploy bundle: ~16 files, ~125 MB (7.0), with the largest items being the imaging-corpus embedding (~69 MB), the ONNX model (~23 MB), and the noimag embedding (~27 MB).

Watching the deploy¶

After ./deploy.sh finishes, tail server logs:

Rscript -e "rsconnect::showLogs(appName='abcd-dictionary', account='biplabendu', streaming=TRUE)"

A healthy first boot looks like:

shinyapps[...]: Shiny application starting ...
shinyapps[...]: Attaching package: 'dplyr'
shinyapps[...]: Attaching package: 'bslib'
shinyapps[...]: Downloading uv...Done!
shinyapps[...]: Downloading cpython-3.12.13-linux-x86_64-gnu (32.5MiB)
shinyapps[...]:  Downloaded cpython-3.12.13-linux-x86_64-gnu
shinyapps[...]: Downloading numpy (15.9MiB)
shinyapps[...]: Downloading tokenizers (3.2MiB)
shinyapps[...]: Downloading onnxruntime (17.3MiB)
shinyapps[...]: Installed 27 packages in 5.40s
shinyapps[...]: Listening on http://127.0.0.1:35131

If the log streams stop before "Listening on…", see Troubleshooting.

Sanity-checking the bundle before deploying¶

deploy.sh already prints a preview, but you can also run it standalone (this won't replicate the in-script "drop non-active parquet" filter — deploy.sh is the canonical view):

Rscript -e '
  files <- rsconnect::listDeploymentFiles(".")
  sizes <- file.info(files)$size
  cat(sprintf("%d files, %.1f MB\n", length(files), sum(sizes)/1e6))
  for (i in order(-sizes)) cat(sprintf("  %7.2f MB  %s\n", sizes[i]/1e6, files[i]))
'

Confirm no raw CSVs, sanity-chks/, python/__pycache__/, or .RData appear, and only the active parquet is present. If something leaks, add the exact directory entry name (no trailing slash, no nested path) to .rscignore, or extend the filter inside deploy.sh.

Updating the deployed app¶

Deployments are versioned by shinyapps.io. Every ./deploy.sh invocation:

• Uploads a new bundle
• Builds a new container image (~1 min for unchanged R deps; longer if renv.lock changed)
• Performs a rolling restart (deploying: Starting instances → terminating: Stopping old instances)

Users connected to the old version stay on it until they reload.