When caching is too fine-grained, an excessive number of cache
requests can result in HTTP 429 errors due to rate limiting.
By caching all artifacts of a particular type in a single entry
we hope to mitigate this, at the expense of some reduction in
cache space optimization.
This change also adds caching for all dependency jars, as well as
instrumented jars in the 'caches/jars-X' directory.
Unfortunately, doing this overloads the GitHub actions cache infrastructure
leading to failures and unpredictable results.
A later solution may re-implement artifact sharing for dependency jars
as well as jars within the `caches/jars-9` directory. But for now these
will be duplicated across each Gradle User Home cache entry.
Similar to wrapper distributions, these large files are common
to many Gradle User Home cache entries. Storing them separately removes
this redundancy from the Gradle User Home cache.
In the current model, each cached Gradle User Home could contain
a copy of one or more downloaded wrapper distributions. This results
in large cache entries which could easily lead to premature eviction.
With this change, wrapper dists are cached separately from the rest
of the Gradle User Home directory. The artifact file is replaced by
a marker file which allows the action to restore the artifact from
cache when the Gradle user Home cache is restored.
Attempt to capture as much context as possible about the job run
to generate a unique cache key. Unfortunately much of the matrix context
is not available to the action implementation.
This will eliminate cache entries from previous workflow runs, allowing
us to test cache functionality in isolation. If the `CACHE_KEY_SEED` environment
variable is not set, this will have no impact.
The Gradle daemon is not useful for ephemeral builds, and the process
can hold file locks which interfere with cache entry generation.
In the case where multiple Gradle invocations occur in the same job,
we could provide a way for users to override this behaviour, taking care
of stopping any daemon process at the end of the job.
- Do not restore cache when GUH exists
- Include RUNNER_OS in the cache key
- Do not save cache on exact hit
- Only save cache in the final post action
- Log before saving cache
