Support to hide metrics which hasn't update for a long time

[acelyc111]

[acelyc111]

Hi @gjasny , could you please take a look at this PR? thanks

[fluffy]

Hi @acelyc111 - I was just reading your code in this PR and it looks like exactly what I need. Is this code working well for you. I'm considering just pulling your PR into my fork. Thanks.

[acelyc111]

Hi @acelyc111 - I was just reading your code in this PR and it looks like exactly what I need. Is this code working well for you. I'm considering just pulling your PR into my fork. Thanks.

@fluffy Yes it works well, feel free to try it.

[gjasny]

Hello,

I see some problems with this pull request:

1. Setting last_update_ts_ is not synchronized between the updating thread and the collecting thread. A 32 bit system setting a 64 bit time_t might read funny values.

2. std::time is not monotonic and might step backwards.

3. The metric types have been designed for low-overhead when setting/updating. Looking at the clock every time adds a huge overhead. The included benchmark shows some numbers (on my M1 Mac Mini)

Before:
BM_Counter_Increment                              10.3 ns         10.3 ns     55964630

After:
BM_Counter_Increment                              45.3 ns         45.3 ns     15390579

Maybe a smarter algorithm could shave-off most of the overhead while also providing thread safety.

@jupp0r Do you have any comments?

Thanks, Gregor