How to use remote cache to fill gaps in local disk cache

[MTomBosch]

In our dev infrastructure we will have a remote cache located on some cloud infrastructure.

The developer A will have a local computer and additionally an enabled disk cache.

In case a CI/CD pipeline was triggered due to some changes of a different developer B the remote cache will be updated to the most current state.

If now developer fetches/pulls the changes into the Git repository and triggers a build it would be really cool if the updated action cache entries that are not available in the disk cache are fetched from the remote cache and then stored in the disk cache so that the remote cache is no more queried for the subsequent builds.

In the end I would like to have following behaviour of the cache:

For a given action:

• Check first if there is a result in the disk cache. If available return it. If not go next
• Check if there is a result in the remote cache. If available return it. If not go next
• Run the action and save the result in the disk cache and optionally (dep. on config) in the remote cache.

This would also mean that the access to the remote cache is minimized as much as possible and favour the disk cache if possible.

Can this somehow be achieved by configuration?

I checked the documentation but could not make up my mind wrt to this question and because of that I started this discussion.

Thank you

[fmeum]

This should already be the case: When a disk cache is used, CAS entries fetched from the remote cache are added to the disk cache at

bazel/src/main/java/com/google/devtools/build/lib/remote/CombinedCache.java

Line 456 in 7ca1159

diskCacheClient.captureFile(tempPath, digest, Store.CAS);

and AC entries at

bazel/src/main/java/com/google/devtools/build/lib/remote/CombinedCache.java

Line 221 in 7ca1159

diskCacheClient.uploadActionResult(actionKey, actionResult),

.

Could you test whether this works in your setup?

[MTomBosch]

I will test it and then report here. I just need some time for the local test setup.

Meanwhile one more question: Is first the disk cache queried for a hit (in order to not contact the remote cache if not really required) or the order by which the disk and remote cache are checked for a cache hit is undeterminstic?

The reason for asking this is the fact that I am expecting (in our setup) the connectivity from the local developers machine to the remote cache to be relatively slow with relatively high latency.

[fmeum]

The disk cache is checked first if used.

[MTomBosch]

I could test it and can confirm that the cached results are transferred from the remote cache to the disk cache in case of hits in the remote cache.