Robust and Efficient Code

R can be a robust, fast and efficient programming language, but some coding practices can be very unfortunate. Here are some suggestions.

Guiding Principles

  1. The primary principle is to make sure your code is correct. Use identical() or all.equal() to ensure correctness, and unit tests to ensure consistent results across code revisions.

  2. Write robust code. Avoid efficiencies that do not easily handle edge cases such as 0 length or NA values.

  3. Know when to stop trying to be efficient. If the code takes a fraction of a second to evaluate, there is no sense in trying for further improvement. Use system.time() or a package like microbenchmark to quantify performance gains.

Common Advice


Vectorize, rather than iterate (for loops, lapply(), apply() are common iteration idioms in R). A single call y <- sqrt(x) with a vector x of length n is an example of a vectorized function. A call such as y <- sapply(x, sqrt) or a for loop for (i in seq_along(x)) y[i] <- sqrt(x[i]) is an iterative version of the same call, and should be avoided. Often, iterative calculations that can be vectorized are “hidden” in the body of a for loop

for (i in seq_len(n)) {
    ## ...
    tmp <- foo(x[i])
    y <- tmp + ## ...

and can be ‘hoisted’ out of the loop

tmp <- foo(x)
for (i in seq_len(n)) {
    ## ...
    y <- tmp[i] + ##

Often this principle can be applied repeatedly, and an iterative for loop becomes a few lines of vectorized function calls.

‘Pre-allocate and fill’ if iterations are necessary

Preallocate-and-fill (usually via lapply() or vapply()) rather than copy-and-append. If creating a vector or list of results, use lapply() (to create a list) or vapply() (to create a vector) rather than a for loop. For instance,

n <- 10000
x <- vapply(seq_len(n), function(i) {
    ## ...
}, integer(1))

manages the memory allocation of x and compactly represents the transformation being performed. A for loop might be appropriate if the iteration has side effects (e.g., displaying a plot) or where calculation of one value depends on a previous value. When creating a vector in a for loop, always pre-allocate the result

x <- integer(n)
if (n > 0) x[1] <- 0
for (i in seq_len(n - 1)) {
    ## x[i + 1] <- ...

Never adopt a strategy of ‘copy-and-append’

not_this <- function(n) {
    x <- integer()
    for (i in seq_len(n))
        x[i] = i

This pattern copies the current value of x each time through the loop, making n^2 / 2 total copies, and scale very poorly even for trivial computations:

> system.time(not_this(1000)
   user  system elapsed
  0.004   0.000   0.004
> system.time(not_this(10000))
   user  system elapsed
  0.169   0.000   0.168
> system.time(not_this(100000))
   user  system elapsed
 22.827   1.120  23.936

Avoid 1:n style iterations

Write seq_len(n) or seq_along(x) rather than 1:n or 1:length(x). This protects against the case when n or length(x) is 0 (which often occurs as an unexpected ‘edge case’ in real code) or negative.

Re-use existing functionality

For common input formats see common Bioconductor methods and classes

If there are problems, e.g., in performance or parsing your particular file type, ask for input from other developers on the bioc-devel mailing list. Common disadvantages to ‘implementing your own’ are the introduction of non-standard data representations (e.g., neglecting to translate coordinate systems of file formats to Bioconductor objects) and user bewilderment.

Re-use existing classes

Re-use enhances interoperability between Bioconductor packages while providing robust code for data manipulation.

Use GenomicRanges::GRanges (and GRangesList) to represent 1-based, closed-interval genomic coordinates.

Use SummarizedExperiment::SummarizedExperiment (with or without ranges as row data) to coordinate rectangular feature x sample data (e.g., RNAseq count matrix) with feature and sample description. Use SummarizedExperiment rather than the older ExpressionSet, especially for sequence data.

For more existing classes see common Bioconductor methods and classes

Essential S4 interface

Remember to re-use common Bioconductor methods and classes before implementing new representations. This encourages interoperability and simplifies your own package development.

If your data requires a new representation or function, carefully design an S4 class or generic so that other package developers with similar needs will be able to re-use your hard work, and so that users of related packages will be able to seamlessly use your data structures. Do not hesitate to ask on the Bioc-devel mailing list for advice.

For any class you define, implement and use a ‘constructor’ for object creation. A constructor is usually plain-old-function (rather than, e.g., a generic with methods). It provides documented and user-friendly arguments, while allowing for developer-friendly implementation. Use the constructor throughout your own code, examples, and vignette.

Implement a show() method to effectively convey information to your users without overwhelming them with detail.

Accessors (simple functions that return components of your object) rather than direct slot access (using @) help to isolate the implementation of your class from its interface. Generally @ should only be used in an accessor, all other code should use the accessor. The accessor does not need to be exported from the class if the user has no need or business accessing parts of your data object. Plain-old-functions (rather than generic + method) are often sufficient for accessors; it’s often useful to employ (consistently) a lightweight name mangling scheme (e.g., starting the accessor method name with a 2 or 3 letter acronym for your package) to avoid name collisions between similarly named functions in other packages.

The following layout is sometimes used to organize classes and methods; other approaches are possible and acceptable.

A Collates: field in the DESCRIPTION file may be necessary to order class and method definitions appropriately during package installation.

Parallel Recommendations

We recommend using BiocParallel. It provides a consistent interface to the user and supports the major parallel computing styles: forks and processes on a single computer, ad hoc clusters, batch schedulers and cloud computing. By default, BiocParallel chooses a parallel back-end appropriate for the OS and is supported across Unix, Mac and Windows. Coding requirements for BiocParallel are:

For more information see the BiocParallel vignette.