GitHub Actions Caching Update

I am currently using GitHub Actions to run automated tests for my open source projects. A few weeks ago, I noticed a comment by Brandon Chinn about ever-growing cache when one is not careful with restore-keys. I do not remember which project the comment was on, and I am unable to find it, but the comment linked to another comment that he made on Mark Karpov’s GitHub actions for Haskell CI blog post. I updated the caching configuration for TTC this morning.

The cache action has three parameters:

• path specifies one or more paths to cache
• key specifies how to create the cache key, a string that identifies a specific cache entry
• restore-keys specifies one or more key prefixes that determine which previous cache entry to build on when an exact match is not found

When a job’s cache key matches an existing cache key, that cache entry is loaded. This may allow the job to avoid re-downloading and re-building stuff, which can significantly speed up job execution. It is important to note, however, that the cache entry is not updated. If your cache key is too general, you will end up with an old cache that is not very useful. For this reason, it is recommended to include a hash in the cache key so that significant changes in the code create a new/updated cache entry.

Updating cache is done using restore-keys. When a job’s cache key does not match any existing cache key, restore-keys determines if any existing cache entry should be loaded anyway. In this case, the updated paths are stored under the new cache key upon job success. The issue that Brandon commented about is that general restore-keys can cause the cache to get bigger and bigger, since new dependency versions are added without removing older versions.

The solution that Brandon suggests is to include a timestamp in the cache key. By including the current month, a cache will only grow for a month before being reset. I followed this suggestion, just changing the month format to YYYYmm instead of using the English name.

Cache entries are removed when they have been accessed for over seven days, so old cache entries are removed automatically.

I just updated the configuration for TTC this morning, but I plan to update the configuration for my other projects when I can find the time.