refreshAfterWrite not allowed to nullify non-expired keys?

[SHxKM]

Requirement:

1. Attempt to refresh an entry t1 after it has been added
2. Expire the entry t2 after it has been added, unless it has been "refreshed", in which case its expiry should renew

refreshAfterWrite seems to be the ideal way to do it, however, I found something about it that works perfectly for my use-case but was surprising to me: if the refresh callback returns null, it will not remove the entry (again, perfect for my use-case, but kinda unintuitive. Here's a test case:

Update: it seems like the behavior is different in real environments as opposed to unit-tests, meaning that if refreshAfterWrite is due, and the refresh function returns null, then the value is nullified and evicted from cache.

@Test
void testRefreshAfterWriteWhenNullIsFetched() throws InterruptedException {
    Cache<String, String>  testCache = Caffeine.newBuilder()
            .refreshAfterWrite(Duration.ofSeconds(1))
            .expireAfterWrite(Duration.ofSeconds(10))
            .build(this::fetchFromRemote);

    testCache.put("foo", "bar");
    assertThat(testCache.getIfPresent("foo")).isEqualTo("bar");

    Thread.sleep(3000);
    assertThat(testCache.getIfPresent("foo")).isEqualTo("bar");

    Thread.sleep(9000);
    assertThat(testCache.getIfPresent("foo")).isNull();
}

private @Nullable String fetchFromRemote(String s) {
      return null;
}

Is this because according to caffeine, refresh is only for "hydrating" the value, and not ever for removing it?

[ben-manes]

Sorry, I'm unsure what the question is. A refresh to a null value should remove the entry. The various loads try to follow their Map compute counterpart, here a refresh is an asynchronous computeIfPresent. If you implement the CacheLoader explicitly then you can override reload or asyncReload to leverage the old value. The asyncReload is the callback and you can delay the future, e.g. to batch reloads or simply never complete one. If the entry is updated, removed, or evicted then the in-flight reload will be abandoned as the cache tries to have linearizable semantics for reasoning against (or close enough to be a good mental model). I'm not sure if any of that helped but let me know either way.

[SHxKM]

Sorry, I'm unsure what the question is. A refresh to a null value should remove the entry. The various loads try to follow their Map compute counterpart, here a refresh is an asynchronous computeIfPresent.

That what the question indeed, in my unit tests I saw the value was not nullified, but I guess that's because the background thread didn't get a chance to run. Thanks @ben-manes!