In jhrcook/mustasher: Stash and Load Objects

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(digest)

library(mustashe)

Overview

The following is the actual code for the stash() function, the main function of the 'mustashe' package. I have only added a few more comments for clarification.

stash <- function(var,
                  code,
                  depends_on = NULL,
                  functional = NULL,
                  verbose = NULL) {
  if (is.null(functional)) functional <- mustashe_functional()
  if (is.null(verbose)) verbose <- mustashe_verbose()

  # Make sure the stashing directory ".mustashe" is available.
  check_stash_dir()

  # Deparse and format the code.
  deparsed_code <- deparse(substitute(code))
  formatted_code <- format_code(deparsed_code)

  # Make sure the `var` and `code` are valid.
  var <- validate_var(var, functional)
  if (formatted_code == "NULL") stop("`code` cannot be NULL")

  # The environment where all code is evaluated and variables assigned.
  target_env <- parent.frame()
  # Make a new hash table.
  new_hash_tbl <- make_hash_table(formatted_code, depends_on, target_env)

  # if the variable has been stashed:
  #     if the hash tables are equivalent:
  #         load the stored variable
  #     else:
  #         make a new stash
  # else:
  #     make a new stash
  if (has_been_stashed(var)) {
    old_hash_tbl <- get_hash_table(var)
    if (hash_tables_are_equivalent(old_hash_tbl, new_hash_tbl)) {
      if (verbose) {
        message("Loading stashed object.")
      }
      res <- load_variable(var, functional, target_env)
    } else {
      if (verbose) {
        message("Updating stash.")
      }
      res <- new_stash(
        var, formatted_code, new_hash_tbl, functional, target_env
      )
    }
  } else {
    if (verbose) {
      message("Stashing object.")
    }
    res <- new_stash(var, formatted_code, new_hash_tbl, functional, target_env)
  }

  invisible(res)
}

Overall, I believe the logic is quite simple. The steps that the stash() function follows, further explained in the following sections, are:

Deparse and format the code.
Make a hash table based on the code and dependencies.
If the object has previously been stashed, then the new hash table and the stashed one are compared. If they are the same, then the object is loaded from memory.
If the hash tables are different or the object has never been stored, then the code is evaluated and the object and its hash table are stashed.

The stashing process

Deparsing and formatting code

The first step taken by the stash() function is to deparse and format the code.

Deparsing the code means to turn the unevaluated expression into a string. The deparsing is done by passing the code immediately to substitute() and deparse(). This must be done immediately, else the code will be evaluated. The substitute() function "returns the parse tree for the (unevaluated) expression expr, substituting any variables bound in env."

substitute(x <- 1)

The deparse() function "Turn[s] unevaluated expressions into character strings." Paired with substitute(), it returns a string of the unevaluated code.

deparse(substitute(x <- 1))

With the code now as a string, it is formatted using the tidy_source() function from 'formatR'. An internal function in 'mustashe', format_code() handles this process:

format_code <- function(code) {
  fmt_code <- formatR::tidy_source(
    text = code,
    comment = FALSE,
    blank = FALSE,
    arrow = TRUE,
    brace.newline = FALSE,
    indent = 4,
    wrap = TRUE,
    output = FALSE,
    width.cutoff = 80
  )$text.tidy
  paste(fmt_code, sep = "", collapse = "\n")
}

format_code("x <- 2")

The purpose of formatting the code is so any stylistic changes to the code input do not affect the hash table. To demonstrate this, notice how the output from format_code() is the same between the two different code examples.

format_code("x=2")
format_code(("x <- 2  # a comment"))

Making a hash table

The hash table is a two-column table with the name and hash value of the code and any (optional) dependencies.

The hashing is handled by the 'digest' package. It takes a value and reproducibly produces a unique hash value.

digest::digest("mustashe")

A hash value is made for the code and for any of the dependencies linked to the object. This process is handled by the make_hash_table(code, depends_on) internal function.

Comparing hash tables

To tell if the code or dependencies have changed, the new hash table and stashed hashed table are compared. The function underlying this process is all.equal() from base R. This function compares two objects and "If they are different, [a] comparison is still made to some extent, and a report of the differences is returned."

Here is an example of using all.equal() to compare two data frames.

# Two data frames with a small difference     *
df1 <- data.frame(a = c(1, 2, 3), b = c(5, 6, 7))
df2 <- data.frame(a = c(1, 2, 3), b = c(5, 6, 8))

# When the two data frames are equivalent.
all.equal(df1, df1)

# When the two data frames are not equivalent.
all.equal(df1, df2)

A word of caution, if using all.equal() for a boolean comparison (like in an if-statement), make sure to wrap it with isTRUE, otherwise it will return TRUE or comments on the differences, but not FALSE. Below is the actual function stash() uses to compare hash tables.

hash_tables_are_equivalent <- function(tbl1, tbl2) {
  isTRUE(all.equal(tbl1, tbl2, check.attributes = TRUE, use.names = TRUE))
}

Evaluating code and making a stash

If the hash tables are different, that means the code must be evaluated, the new object be assigned to the desired name (var), and the new hash table and value stashed. This is handled by the internal function new_stash().

# Make a new stash from a variable, code, and hash table.
new_stash <- function(var, code, hash_tbl, functional, target_env) {
  val <- evaluate_code(code, target_env)
  write_hash_table(var, hash_tbl)
  write_val(var, val)
  if (functional) {
    return(val)
  } else {
    assign_value(var, val, target_env = target_env)
    return(NULL)
  }
}

The first step is to evaluate the code with the evaluate_code(code) function. It uses the parse() and eval() functions and returns the resulting value.

evaluate_code <- function(code, target_env) {
  eval(parse(text = code), envir = new.env(parent = target_env))
}

This value is then assigned the desired name in the global environment using the internal assign_value(var, val) function, where .TargetEnv is a variable in the package pointing to .GlobalEnv.

assign_value <- function(var, val, target_env) {
  assign(var, val, envir = target_env)
}

Lastly, the hash table and value are written to file using wrapper functions around qs::qsave().

Some details

Below are some implementation details that are useful to know about.

Environment management

When I first created this library, I just made all assignments to and searched for dependencies in the global environment (.GlobalEnv). This generally worked for standard use of the package, but could easily result in strange behavior if stash() was used within functions. In early 2022, Magnus Torfason (GitHub @torfason) addressed this shortcoming by capturing the parent environment in stash() and conducting code evaluation and dependency search within that R environment.

Functions as dependencies

As described in Issue #8, using functions as dependencies can be tricky because their hashes change after evaluation.

fxn <- function(x) {
  x**2
}

print(digest(fxn))
print(digest(fxn))
a <- fxn(2)
print(digest(fxn))
a <- fxn(2)
print(digest(fxn))
a <- fxn(4)
print(digest(fxn))

You can actually see a difference in the print-out of a function when in an interactive R session.

fxn <- function(x) {
  x**2
}

fxn

a <- fxn(2)
a <- fxn(2)
fxn

Therefore, stash() uses a specific process for hashing functions. As described in the section on "Functions" by Hadley Wickham in Advanced R, "R functions have three parts:

the body(), the code inside the function.
the formals(), the list of arguments which controls how you can call the function.
the environment(), the "map" of the location of the function’s variables."

Therefore, stash() specifically deparses the code of a function passed as dependency and uses the hash of the resulting string (of the code of the function) as the hash.

fxn_str <- deparse(fxn)
fxn_str

digest(fxn_str)

Contact

Any issues and feedback on 'mustashe' can be submitted here. Alternatively, I can be reached through the contact form on my website or on Twitter \@JoshDoesa