For a cached code chunk, its output and objects will be automatically loaded from the previous run, as if the chunk were executed again. Caching is often helpful when loading results is much faster than computing the results. However, there is no free lunch. Depending on your use case, you may need to learn more about how caching (especially ) works, so you can take full advantage of it without confusing yourself as to why sometimes knitr invalidates your cache too often and sometimes there is not enough invalidation.

The most appropriate use case of caching is to save and reload R objects that take too long to compute in a code chunk, and the code does not have any side effects, such as changing global R options via options() (such changes will not be cached). If a code chunk has side effects, we recommend that you do not cache it.

You have to let knitr know if the data file has been changed. One way to do it is to add another chunk option cache.extra = file.mtime('my-precious.csv') or more rigorously, cache.extra = tools::md5sum('my-precious.csv'). The former means if the modification time of the file has been changed, we need to invalidate the cache. The latter means if the content of the file has been modified, we update the cache. Note that is not a built-in knitr chunk option. You can use any other name for this option, as long as it does not conflict with built-in option names.

Similarly, you can associate the cache with other information such as the R version (cache.extra = getRversion()), the date (cache.extra = Sys.Date()), or your operating system (cache.extra = Sys.info()[['sysname']]), so the cache can be properly invalidated when these conditions change.

If you are not happy with knitr’s design for caching, you can certainly cache objects by yourself. Below is a quick example:

if (file.exists("results.rds")) {
} else {
  saveRDS(res, "results.rds")

In this case, the only (and also simple) way to invalidate the cache is to delete the file results.rds. If you like this simple caching mechanism, you may use the function introduced in Section 14.9.