In order to continuously monitor the performance of ggplot2 the following piece of code is used to generate a profile and inspect it:

library(ggplot2)
library(profvis)

p <- ggplot(mtcars, aes(x = mpg, y = disp)) + 
  geom_point() + 
  facet_grid(gear ~ cyl)

profile <- profvis(for (i in seq_len(100)) ggplotGrob(p))

profile

In general, a minimal plot is used so that profiles are focused on low-level, general code, rather than implementations of specific geoms. This might be expanded at the point where improving performance of specific geoms becomes a focus. Further, the profile focuses on the steps up until a final gtable have been constructed. Any performance problems in rendering is likely due to grid and the device, more than ggplot2.

Profiles for old version are kept for reference and can be accessed at the github repository. Care should be taken in not comparing profiles across versions, as changes to code outside of ggplot2 can have profound effect on the results. Thus, the intend of profiling is to identify bottlenecks in the implementation that are ripe for improvement, more then to quantify improvements to performance over time.

Performance focused changes across versions

To keep track of changes focused on improving the performance of gtable they are summarised below:

v3.3.3

  • Avoid costly factor construction in scale_apply This issue only really appeared when plotting huge datasets (>1e6 rows). In order to train and map the position aesthetics rows has to be matched based on their panel. This require splitting the row indexes by panel which included a factor() call. The factor constructor is very slow for large vectors and can be simplified considerably for this specific case. Further, the split can be avoided completely when there is only one panel

v3.1.0

  • Caching of calls to grid::descentDetails() The absolute biggest offender was the construction of titles. In recent versions this has included calls to grid::descentDetails() to ensure that they are aligned across plots, but this is quite heavy. These calls are now cached so they only have to be calculated once per font setting.
  • Use a performant data.frame constructor throughout the codebase The data.frame() function carries a lot of overhead in order to sanitize and check the input. This is generally not needed if you are sure about the input and will just lead to slower code. The data.frame() call is now only used when dealing with output from other packages where the extra safety is a benefit.
  • Use a performant alternative to utils::modifyList modifyList() is a nice convenience function but carries a lot of overhead. It was mainly used in the plot element constructions where it slowed down the application of theme settings. A more performant version has been added and used throughout.
  • Speed up position transformation The transform_position helper was unreasonably slow due to the slowness of getting and assigning columns in data.frame. The input is now treated as a list during transformation.