Implement ValueStream?

[thefallentree]

Hi,

I'm surprised to see that prometheus (or should I say, borgmon2) doesn't implement a valuestream function yet (allow transferring timeseries between prometheus instances). Which I thought was crucial to implement a multi-tier aggregation hierarchy,

Any thoughts?

BTW: you guys are awsome creating prometheus!

[matthiasr]

I've been playing around with multi-tier aggregation by exposing handcrafted "metrics" endpoints as console templates, building on this example. You can, instead of matching the name, specify any query (but need to then give it a new name). The effect is that one Prometheus can publish arbitrary (raw, aggregated or otherwise computed) timeseries that are then available to be scraped (albeit not at the standard endpoint) by any other prometheus.

[beorn7]

@sunyc Could you explain what you mean by "valuestream"?

[thefallentree]

@matthiasr which is obviously not very efficient... And there is timestamping issue.

@beorn7 ValueStream is a RPC service that allows an prometheus instance scrape timeseries from another prometheus . Which also retains original timestamp etc.

It has a number of usefulness, namely, one could have a prometheus that valuestream from several peers for aggregation etc etc etc.

[matthiasr]

I'm not too worried about timestamps … when aggregating for example request counts from many individual instances as one timeseries, there is no one timestamp anyway. And if the samples are just passed through 1:1 then why aggregate in the first place?

[beorn7]

The Prometheus scrape protocol already contains a timestamp for that purpose. However, it is currently not possible to scrape multiple samples for one time series in one scrape.

[thefallentree]

@matthiasr because scrape is not synchronized between targets, If there were no timestamp in the output the aggregation would produce wrong values..

[juliusv]

Heya Yucong,

valuestreaming aka hierarchical federation is on our roadmap:
http://prometheus.io/docs/introduction/roadmap/#hierarchical-federation

Right now we only have the "hack" that Matthias mentioned, but soon[tm]
we'll have the real thing. I agree, it should be possible to transfer raw
samples with their original timestamps and uninterpolated values.

Cheers,
Julius

On Wed, Apr 15, 2015 at 5:27 PM, Yucong Sun notifications@github.com
wrote:

@matthiasr https://github.com/matthiasr because scrape is not
synchronized between targets, If there were no timestamp in the output the
aggregation would produce wrong values..

—
Reply to this email directly or view it on GitHub
#638 (comment)
.

[thefallentree]

I would be happy to help (maybe we can also use go-grpc now), believe it or not, i spent quite a bit of of time on it in the past life .

LoL

[beorn7]

I'll close this as a duplicate of #9 (updated that issue with the notion of "streaming").

[juliusv]

@sunyc That'd be great in general. As a smaller starter project it'd probably be good to start with something small to get acquainted with the whole code and contributing business, and then go from there. We're all also hanging out on #prometheus on Freenode for discussions and inspiration :)

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.