{targets} + {vetiver} (and other kinds of continuous deployment)

[wlandau]

@philbowsher suggested that I explore the power of vetiver in targets pipelines. (@juliasilge, thanks for the package.) The integration seems like another successful instance of pipeline-based continuous deployment, which also comes up for Shiny apps and literate programming documents. The following _targets.R file extends the example from https://github.com/wlandau/targets-four-minutes with a vetiver model and and a pin:

# _targets.R file
library(targets)
tar_option_set(packages = c("dplyr", "parsnip", "pins", "vetiver", "workflows"))

pin_model <- function(model) {
  local_options(list(pins.quiet = TRUE))
  board <- board_folder(path = "example_board")
  suppressMessages(vetiver_pin_write(board, model))

  # Return the hash so any target downstream of the pin can update in response to changes:
  pin_meta(board, "air_quality")$pin_hash
}

list(
  tar_target(
    name = data,
    command = filter(airquality, !is.na(Ozone))
  ),
  tar_target(
    name = regression,
    command = lm(Ozone ~ Temp, data)
  ),
  tar_target(
    name = model,
    command = vetiver_model(model = regression, model_name = "air_quality")
  ),
  tar_target(
    name = pin,
    command = pin_model(model)
  )
)

The dependency graph shows the overall flow of the pipeline.

# R console
tar_mermaid()

graph LR
  subgraph legend
    direction LR
    x0a52b03877696646([""Outdated""]):::outdated --- xbf4603d6c2c2ad6b([""Stem""]):::none
    xbf4603d6c2c2ad6b([""Stem""]):::none --- xf0bce276fe2b9d3e>""Function""]:::none
  end
  subgraph Graph
    direction LR
    xb7119b48552d1da3(["data"]):::outdated --> x2af8164e6b30c875(["regression"]):::outdated
    xe1eeca7af8e0b529(["model"]):::outdated --> xd4981b1ef5777538(["pin"]):::outdated
    x7b4395794557f035>"pin_model"]:::outdated --> xd4981b1ef5777538(["pin"]):::outdated
    x2af8164e6b30c875(["regression"]):::outdated --> xe1eeca7af8e0b529(["model"]):::outdated
  end
  classDef outdated stroke:#000000,color:#000000,fill:#78B7C5;
  classDef none stroke:#000000,color:#000000,fill:#94a4ac;
  linkStyle 0 stroke-width:0px;
  linkStyle 1 stroke-width:0px;

Loading

And the targets in the pipeline run in topological order.

# R console
tar_make()
#> • start target data
#> • built target data [0.406 seconds]
#> • start target regression
#> • built target regression [0.005 seconds]
#> • start target model
#> • built target model [0.019 seconds]
#> • start target pin
#> • built target pin [0.101 seconds]
#> • end pipeline [0.67 seconds]

The pin only updates when the vetiver model object changes.

# After nothing changed:
tar_make()
#> ✔ skip target data
#> ✔ skip target regression
#> ✔ skip target model
#> ✔ skip target pin
#> ✔ skip pipeline [0.079 seconds]

This cuts down on the number of superfluous versions in the model history. In production, the pipeline could run on Connect at regular intervals, maybe in a scenario where the data updates on a schedule and the model needs to be retrained. targets would strategically reduce both computation time and the number of unimportant saved models.

Model cards would work well with the literate programming integration at https://books.ropensci.org/targets/literate-programming.html#literate-programming-within-a-target. As for plumber APIs, another target could redeploy the API with a short-lived synchronous command if needed, and then downstream targets could run predictions with the updated API using predict(vetiver_endpoint("https://..."), ...).

[juliasilge]

This is looking great @wlandau! 🙌 Thanks for sharing. Let me know if you have questions as you work on this in the future.