commit 3bce568591e5a1e705fb65b75b88280b915efb96
author Luca Milanesio <luca.milanesio@gmail.com> Thu Nov 03 19:41:01 2022 +0000
committer Luca Milanesio <luca.milanesio@gmail.com> Mon Nov 07 17:12:32 2022 +0000
tree 393ea09bbd4e434cf7f86726627c60654e41f1cf
parent cbf312f854e38f963ab0dc2fb3f2e62c6e9e3975

Introduce a PassthroughLoadingCache for disabled caches

Gerrit documentation states that when the memoryLimit of a cache
is zero then it is disabled. However, the code told a different
story because there was always a Guava/Caffeine cache created
with a potential impact on performance due to the inherent locking
of the underlying ConcurrentHashMap.

The ConcurrentHashMap would always lock additions and removals on
the same key, and the only way to avoid locking is to avoid
the instantiation of the ConcurrentHashMap.

Why should someone want to ever disable a cache?

Disabling the cache typically isn't a good idea, however, there
could be situations where:
1. The rate of the access of the cached value for K is low and the
   load time is high
2. The rate of the eviction of the cached value for K is high

The Issue 16379 is one example of the above condition, however,
there are more instances of the same problem.

The eviction done at 2. for K would then cause all the incoming threads
to wait until the load at 1. for K is done.

Introduce a simple LoadingCache implementation called
PassthroughLoadingCache that is *really* just delegating the
calls to the underlying loader without involving any caching
or locking.

This change also makes the memoryLimit behaviour consistent with the
diskLimit set to zero means disabling the implementation of the cache.

Bug: Issue 16379
Release-Notes: Increase overall performance of disabled caches by avoiding any locking due to Guava/Caffeine caches with zero weight.
Change-Id: I5906a1f250014d96aaadcda0f64a37a78030d047