Packaging an inlabru model API in a container

Once a model is pinned to a board (see vignette("vetiver")), the same /predict API can run in a container. This vignette shows a Dockerfile that works for inlabru fits, and why it differs from what vetiver_prepare_docker() would generate. The commands work with Docker and Podman alike.

Why not `vetiver_prepare_docker()`?

For most models vetiver_prepare_docker() writes a Dockerfile, a plumber.R and an renv lockfile in one go. It does not work for inlabru fits:

It looks up system requirements for every package in the lockfile on Posit Package Manager, which rejects packages it doesn’t know. INLA is not on CRAN, so the call fails with “unrecognized package id”.
vetiver never loads the packages a model needs inside the API process. The methods that serve bru fits live in inlaops, so the container has to run library(inlaops) before building the API.
inlabru imports sf, so every image needs GDAL, GEOS and PROJ, spatial model or not.

So write plumber.R with vetiver and the Dockerfile by hand.

1. Write `plumber.R`

Pin the model to a board the container can reach, then let vetiver write the API file:

library(vetiver)
library(pins)

board <- board_folder("model-store")   # or board_s3(), board_connect(), ...
vetiver_pin_write(board, v)

vetiver_write_plumber(board, "gaussian_glm")

The generated plumber.R reads the model from the board when the API starts, so deploying a new version means pinning it and restarting the container — no rebuild.

2. Write the `Dockerfile`

# inlabru imports sf, so every image needs GDAL/GEOS/PROJ:
# rocker/geospatial ships them prebuilt.
FROM docker.io/rocker/geospatial:4.6.1

# INLA is not on CRAN.
RUN Rscript -e "install.packages('INLA', repos = c(getOption('repos'), \
  INLA = 'https://inla.r-inla-download.org/R/stable'))"

# sn: INLA's posterior sampler needs it but only lists it in Suggests.
RUN Rscript -e "install.packages(c('inlabru', 'sn', 'vetiver', 'pins', \
  'plumber', 'rapidoc', 'remotes'))"
RUN Rscript -e "remotes::install_github('novica/inlaops', upgrade = 'never')"

WORKDIR /opt/ml
# Only for a local board_folder(); remote boards are read at startup.
COPY model-store model-store
COPY plumber.R plumber.R

EXPOSE 8000
# Load inlaops first: it provides the handlers that serve bru fits.
ENTRYPOINT ["R", "-e", "library(inlaops); pr <- plumber::plumb('/opt/ml/plumber.R'); pr$run(host = '0.0.0.0', port = 8000)"]

Notes:

Pin the image tag. rocker images for past R releases install packages from a dated Posit Package Manager snapshot, which keeps rebuilds reproducible. inlaops needs inlabru (>= 2.15.0), so use an image whose snapshot is at least that recent.
Check the INLA repository URL at https://www.r-inla.org/download/index.html before deploying to a new environment.
For a remote board (S3, Posit Connect, …), drop the COPY model-store line and pass credentials as environment variables at docker run time (see vignette("versioning")).

3. Build and test locally

docker build -t gaussian_glm .
docker run --rm -p 8000:8000 gaussian_glm

The API answers once INLA has loaded, after a few seconds:

curl http://localhost:8000/ping

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '[{"x1": 0.5, "x2": 1.0}, {"x1": -1.2, "x2": 0.3}]'
#> [{".pred":1.9943,".pred_low":1.893,".pred_high":2.1142,".pred_sd":0.057},
#>  {".pred":0.9361,".pred_low":0.8352,".pred_high":1.0505,".pred_sd":0.0543}]

Values differ slightly between calls because they come from posterior sampling.

4. Spatial models

Point-referenced fits (sf / sp points, SPDE fields, LGCP intensity surfaces) use the same image. The request carries coordinates, so the pin metadata needs coord_names and crs:

v <- vetiver_model(
  fit_sf, "sf_points",
  metadata = list(
    pred_formula = "~ Intercept + x1",
    coord_names  = c("X", "Y"),   # sf fits: sf::st_coordinates() names
    crs          = 4326L          # EPSG integer or WKT string
  )
)

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '[{"X": 10.3, "Y": 59.3, "x1": 0.8}]'

See vignette("spatial") for the full request contract.

Polygon / areal fits (sf polygons, SpatialPolygonsDataFrame) are rejected at pin time. Until areal serving lands, use the custom-plumber pattern in vignette("lgcp") (section 3); its plumber.R goes into the same image.

5. Image size

The image is about 5 GB: rocker/geospatial plus INLA, which bundles binaries for several platforms. The first build takes several minutes; later builds reuse the cached layers until the package lines change.

Why not vetiver_prepare_docker()?