Some metrics report wrong values

[ctyar]

Describe the bug

Some metrics including cache.miss, cache.set, and cache.get_or_set sometimes report incorrect values.

To Reproduce

If you call GetOrSet() with the same key 3 times, the first call results in a miss and a set event, which is correct.
The second call will result in a miss, a set (although the cache doesn't actually call to get value from user code), and a hit.
The third call will result in a hit event, which is correct.
I've written a test that shows the scenario here:
#434

Based on my debugging it seems the issue is caused by this condition

FusionCache/src/ZiggyCreatures.FusionCache/FusionCache_Async.cs

Line 914 in e53e2c5

if (ClearExpireTimestamp < 0 || (HasDistributedCache && HasBackplane == false))

and ClearExpireTimestamp being -1 in the second call, which results in a secondary call to GetOrSetAsync.

Expected behavior

If you call GetOrSet() with the same key 3 times I expect to see 1 miss, 1 set, 2 hits and 4 events in total.

Versions

I've encountered this issue on:

• Latest version from main branch
• .NET 9.0
• Windows 10

[jodydonetti]

Hi @ctyar , thanks for the heds up: I'll take a look as soon as possible.

Meanwhile, please note that Tagging is implemented via underlying additional cache entries, see here for how it works.

[jodydonetti]

Hi @ctyar , I doublechecked and the behaviour is in fact technically correct.

Let me explain: as mentioned above it has to do with Clear support (more here), which itself is based on Tagging (more here). In particular it is checking for Clear(true) which means "expire all".

More specifically, since Clear is based on 2 special tags ("__fc:t:!" and "__fc:t:*"), FusionCache needs to check them when it finds an entry in the cache to see if one of the 2 Clear() has been called since when the entry has been saved.

But since you are seeing only 1 extra MISS/SET couple, why it is checking only 1 of them? It has to do with an optimization caled Raw Clear (more here).

So, the steps are the following:

• 1st GetOrSet "foo":
  • MISS "foo"
  • SET "foo"
• 2nd GetOrSet "foo":
  • "foo" found ... is it still valid?
  • GetOrSet "__fc:t:*":
    • MISS "__fc:t:*"
    • SET "__fc:t:*"
    • all good
  • HIT "foo"
• 3nd GetOrSet "foo":
  • "foo" found ... is it still valid?
  • all good (no need to check again because of internal optimizations)
  • HIT "foo"

Makes sense?

Hope this helps, let me know.

[jodydonetti]

PS: thanks to this issue I'm thinking about an extra optimization I can make, will update you later on it.

[ctyar]

@jodydonetti Thank you for taking the time to explain, I now have a better understanding of how this happens.

I found out about this while optimizing our cache hits. Since we weren't using tags, it seemed irrelevant at first. We’ll likely continue to use the metrics but won’t rely on the exact values.

It might be worth considering separate metrics for tags, though I’m not entirely sure.

If no changes are planned, feel free to close the issue.

Thanks again for maintaining such a great project!

[jodydonetti]

It might be worth considering separate metrics for tags, though I’m not entirely sure.

Thanks to the input: I'm not sure it would be easy to separate them in this way, instead I was thinking more of having an option to disable observability for tags & co.

If no changes are planned, feel free to close the issue.

No, I'll keep this open because I just made the extra optimization I come up eith, that sparked from this issue.
Btw now it's not doing anymore the extra check for __fc:t:* in some cases like yours 🤟

Thanks again for maintaining such a great project!

Thanks for using it, I'm glad you're liking it 😊

[jodydonetti]

Hi all, v2.2.0-preview-1 is out 🥳

[jodydonetti]

Hi @ctyar have you been maybe able to give the preview a try? The initial extra HIT/MISS should be gone now.

[ctyar]

I've just tested the v2.2.0-preview-1, and I can confirm the extra hit and miss are gone
Thanks a lot

[jodydonetti]

Awesome, thanks @ctyar !

[jodydonetti]

Hi all, v2.2.0 is out 🥳
And for the corresponding LTS-only release, there's v2.0.2.