If we're on Mac and Windows, we're using boot2docker, and we need to run the
equivalent of $(boot2docker shellinit).
Shamelessly stolen from from Winston Chang's harbor
boot2docker_shellinit <- function() {
if (!(Sys.info()["sysname"] %in% c("Darwin", "Windows")))
return()
if (Sys.which("boot2docker") == "")
return()
if (boot2docker_ver() < "1.3.0")
stop("Running boot2docker locally requires boot2docker >= 1.3.0")
Run shellinit and capture the output, which are comands setting env vars for sh. We need read them in and set them from R.
envvars <- system2("boot2docker", "shellinit", stdout = TRUE)
if (length(envvars) != 0) {
envvars <- sub("^ +export +", "", envvars)
envvars <- strsplit(envvars, "=")
envvars <- setNames(pluck(envvars, 2), pluck(envvars, 1))
do.call(Sys.setenv, envvars)
}
}
boot2docker_ver <- function(){
out <- system2("boot2docker", "version", stdout = TRUE)
ver <- gsub("^.*?v([0-9\\.]+).*", "\\1", out[1])
as.package_version(ver)
}
#' Build a docker image with server that serves model$predict and push it to a registry.
#'
#' Available routes:
#'
#' GET/POST / "OK"
#' GET/POST /ping "pong"
#' GET/POST /predict serialized output of predict as JSON
#'
#' @importFrom productivus pp
#' @param model tundraContainer. A tundra container that contains necessary munge
#' procedures and the predict function.
#' @param name character. name of the resulting docker image
#' @param registry character. Where will you push your image? Leave blank for
#' pushing to docker hub, or specify your private registry.
#' @param dockerfile character optional. You can specify a custom dockerfile
#' instead of the default one shipped with kunteynir.
#' @param server_script character optional. You can specify a custom server script
#' that will be used to start serving the model inside the docker container.
#' @export
build_image <- function(model, name, registry = '', dockerfile = NULL, server_script = NULL) {
The goal of this package is to provide an easy way for analysts to convert their models into deployable applications. This is achieved by creating a Docker image that would serve the model's predictions as a RESTful service. By default the server will listen on port 8103.
We insist that the model is a tundraContainer. Check out https://github.com/robertzk/tundra
stopifnot(is(model, "tundraContainer"))
if(missing(name)) stop("You need to name your image.")
dir <- tempdir(); on.exit(unlink(dir))
saveRDS(model, paste0(dir, "/model"))
build_args <- list(dir = dir)
You can provide your custom dockerfile and server script, instead of the ones
in inst/templates.
if (!is.null(dockerfile)) build_args$dockerfile <- dockerfile
if (!is.null(server_script)) build_args$server_script <- server_script
dockerfile <- do.call(write_dockerfile, build_args)
if(identical(registry, '')) separator <- '' else separator <- "/"
cat("Preparing to build: ", registry, separator, name, "\n", sep="")
Build and push to the registry.
system2("docker", paste0("build -t ", registry, separator, name, " ", dir))
system2("docker", paste0("push ", registry, separator, name))
}
write_dockerfile <- function(dir,
dockerfile = system.file("templates", "Dockerfile", package = "kunteynir"),
server_script = system.file("templates", "start_server.R", package = "kunteynir")) {
template <- readLines(dockerfile)
data <- list(
model_path = "model",
start_server_path = "start_server.R",
timestamp = Sys.time()
)
writeLines(whisker.render(template, data), paste0(dir, "/Dockerfile"))
writeLines(readLines(server_script), paste0(dir,"/start_server.R"))
}
#' Convert a tundraContainer into a dockerized REST server.
#'
#' @name kunteynir
#' @import stringr methods whisker
#' @docType package
NULL
pending <- function() {
TRUE
}
pluck <- function(x, name, type) {
if (missing(type)) {
lapply(x, "[[", name)
} else {
vapply(x, "[[", name, FUN.VALUE = type)
}
}
.onLoad <- function(libname, pkgname) {
if (Sys.info()["sysname"] %in% c("Darwin", "Windows")) {
boot2docker_shellinit()
}
}