Downsampling, a la Thanos

[gouthamve]

The downsampling in Thanos is really interesting: https://youtu.be/l8syWgJ98sk?t=1392

@fabxc Do you think it makes sense to upstream this and also make Prometheus use this?

/cc @brian-brazil

[fabxc]

Given long-term storage is still a clear non-goal for Prometheus itself, I think downsampling doesn't add much value. Adding it just to TSDB is something possibly worth thinking about. But given Thanos can implement it as an add-on without issues, I'm not sure how much value there is in pulling it into the core. At least while it's still experimental, I'd not make pre-mature efforts here.

…

On Wed, Apr 4, 2018 at 12:13 PM Goutham Veeramachaneni < ***@***.***> wrote: The downsampling in Thanos is really interesting: https://youtu.be/l8syWgJ98sk?t=1392 @fabxc <https://github.com/fabxc> Do you think it makes sense to upstream this and also make Prometheus use this? /cc @brian-brazil <https://github.com/brian-brazil> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#313>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEuA8kZuLKkFzDZIJGRqWc9kZnYfISRZks5tlJzAgaJpZM4TGhvQ> .

[gouthamve]

Hmm, but reduction in CPU and query time improvements? At what scale do you think the benefits start to show?

[SuperQ]

I feel like it would be useful to have downsampling in the tsdb directly. But I also agree with @fabxc that it's a little soon for this integration.

[fabxc]

Benefits start to show once your retention becomes so long that you'd rather consider it long-term storage. In which case Prometheus itself is not the answer.

…

On Wed, Apr 4, 2018 at 12:34 PM Ben Kochie ***@***.***> wrote: I feel like it would be useful to have downsampling in the tsdb directly. But I also agree with @fabxc <https://github.com/fabxc> that it's a little soon for this integration. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#313 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEuA8g4bWqfeCMBiCZPsqK08l8j-PpK9ks5tlKGrgaJpZM4TGhvQ> .

[brian-brazil]

I agree with @fabxc that this doesn't make sense for Prometheus and thus the TSDB.

[AlekSi]

Duplicate of #56.

[fabxc]

I'd propose closing this. There are different ways to approach downsampling – none of them is universally better than others. I see no way for providing a canonical solution here.

[gouthamve]

Agreed.

[dmitriy-lukyanchikov]

Hello guys.
@fabxc @brian-brazil does it make sence to implement optional downsampling to Prometheus? We experience some issues with Thanos, and we want to implement downsampling to Prometheus itself?