Pipeline views
In pipeflow: Fast Interactive Data Analysis Pipelines

knitr::opts_chunk$set(
    comment = "#",
    prompt = FALSE,
    tidy = FALSE,
    cache = FALSE,
    collapse = TRUE
)

old <- options(width = 100L)

Pipelines can get long, and often you want to focus on a subset of steps: a topic, a stage, or the steps that produce the outputs you care about. Views are {pipeflow}'s way of working on a subset of steps without copying anything. A view references the underlying pipeline, so every operation applied to a view (running it, updating parameters, tagging, locking, ...) writes through to the original pipeline, restricted to the steps covered by the view.

This vignette shows how to create and combine views, how to select steps with the [ operator, how views behave when the pipeline changes, and how to run only part of a pipeline.

Setup

Again, we use a very simplified example of a pipeline.

library(pipeflow)

pip <- pip_new("my-pip") |>
    pip_add("load", \(n = 5) seq_len(n), tags = c("io", "daily")) |>
    pip_add("clean", \(x = ~load) x * 2, tags = c("io", "core")) |>
    pip_add("fit", \(x = ~clean) sum(x), tags = c("model", "core")) |>
    pip_add("report", \(x = ~fit) paste("result:", x), tags = "report")

As you see above, besides step and fun, pip_add() also allows to set tags. We will use these tags as meta information to filter certain steps by topic and/or output type. Let's do a first run before we move on.

(pip_run(pip, lgr = NULL))

Creating views

pip_view() returns a view of the pipeline that contains only the steps matching the given filters:

pip_view(pip, tags = "core")

Filters can be combined. By default, steps must match all filters (logical AND), while the values within a single filter are treated as alternatives (OR):

pip_view(pip, tags = "core", state = "done")

pip_view(pip, step = c("clean", "fit"))

join = "union" keeps steps that match any filter:

pip_view(pip, tags = "report", step = "clean", join = "union")

With fixed = FALSE, filter values are interpreted as regular expressions:

pip_view(pip, step = "^f", fixed = FALSE)

The available filters are step, params, state, exec, tags, and depends. For example, to find all steps that depend on load and are still new:

pip_reset(pip) # reset to initial state

pip_view(pip, depends = "load", state = "new")

Selecting steps with `[`

The extract operator [ provides a data.table-like way of selecting steps. It returns a view by default:

pip_run(pip, lgr = NULL)
pip[c("load", "fit")]

pip[1:3]

Boolean filters are evaluated in the context of the step table, so the same columns as in pip_view() are available as variables:

pip[tags %like% "core"]

pip[step %in% c("clean", "fit") & state == "done"]

Note that %like% matches the tags of each step as a single string, such as 'c("model", "core")', or "character(0)" for a step without tags. A pattern can therefore match more steps than intended, so to select steps by their tags, pip_view(tags = ) is the safer choice.

Negative indices select all steps except the excluded ones:

pip[-2]

p[] returns a copy of the pipeline, and using two indices (p[i, j]) extracts the given rows and columns as a data.table:

pip2 <- pip[]

pip[, c("step", "tags")]

pip[c("load", "fit"), "out"]

While [ returns a view by default, view = FALSE builds a new, self-contained pipeline containing the selected steps together with all their upstream dependencies:

pip[c("fit", "report"), view = FALSE]

The printed message tells you how many steps were pulled in as upstream dependencies.

Composing views

Views can be nested: applying pip_view() (or [) to a view narrows the view further.

v1 <- pip_view(pip, tags = "io") # load, clean
v1

v2 <- v1 |> pip_view(tags = "core") # clean only
v2

The `view` meta field

Under the hood, a view stores the steps it covers in the view meta field of the pipeline object (NULL for a full pipeline). Reading the field returns the names of the covered steps, so in principle you can also manipulate the view directly by assigning step names to the meta field^[ Direct manipulation of the view usually is not needed and probably mostly useful for debugging. ].

pip[["view"]] # NULL — not a view

w <- pip
w[["view"]] <- c("load", "clean")
w

w[["view"]] <- NULL # back to the full pipeline
w

When the pipeline changes

A view keeps covering its steps when the pipeline changes later, no matter whether the change is made through the view, another view, or the pipeline itself. Renamed steps stay in the view under their new name, and removed steps drop out of it. New steps are not added to an existing view, even if they match the view's filters.

q <- pip_clone(pip)
v <- pip_view(q, tags = "core")
v

# A new "core" step is not added to the existing view
pip_add(q, "check", \(x = ~clean) all(x > 0), tags = "core", after = "clean")

# A renamed step stays in the view under its new name
pip_rename(q, from = "fit", to = "model")
v

# Removed steps drop out of the view
pip_remove(q, "model", force = TRUE)
v

Running views

Running a view executes the covered steps together with any upstream dependencies that are not up to date. The run log marks steps that belong to the view as [view] and steps that were pulled in as dependencies as [upstream]:

pip_reset(pip)
pip_run(pip_view(pip, step = "report"))

Afterwards, the original pipeline is up to date for the covered steps.

Pipeline views are not only useful to inspect or run certain parts of the pipeline but also to filter and collect the final output of your analysis run. For more details on this see the next vignette Collect and group output.