Maximum label count? What's "not ok"?

[Timothy Perrett]

Hey all,

After trawling the mailing list and the docs, it seems 10k labels is considered "too many labels", but i have not found any where a guideline for how many labels would be "too many"? Specifically, we were thinking of adding labels in our circuit breaking library such that the breaker state was a metric, and the logical stack it was talking too was a label on that metric... this would give us potentially, say, a thousand labels. Is that viable? The set is pretty bounded, and grows slowly, and tails off over time, so the space of labels for that metric would not grow over time.

Thanks for any advice!

Cheers

Tim 

[Brian Brazil]

On 24 August 2016 at 17:31, Timothy Perrett <tper...@gmail.com> wrote:

Hey all,

After trawling the mailing list and the docs, it seems 10k labels is considered "too many labels", but i have not found any where a guideline for how many labels would be "too many"? Specifically, we were thinking of adding labels in our circuit breaking library such that the breaker state was a metric, and the logical stack it was talking too was a label on that metric... this would give us potentially, say, a thousand labels. Is that viable? The set is pretty bounded, and grows slowly, and tails off over time, so the space of labels for that metric would not grow over time.

If you're just breaking out by logical stack that's probably fine.

Consider what'd happen on 1k targets each talking to 10 stacks. That's not too large a proportion of the 10M or so time series a Prometheus can handle, especially as there will likely be other variables like latency and failures broken out the same way.

Brian

 

Thanks for any advice!

Cheers

Tim 

--
You received this message because you are subscribed to the Google Groups "Prometheus Developers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to prometheus-developers+unsub...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Brian Brazil

www.robustperception.io

[Timothy Perrett]

Thanks Brian, that is useful. We suspected that was the case, but the docs could do with some guideline on this, just to give some kind of frame of reference (i.e. thousands labels x millions of metrics could be a problem).

Cheers

On Wednesday, 24 August 2016 09:46:27 UTC-7, Brian Brazil wrote: