Evaluating off-device attribution proposals at larger scales

[saxena-shobhit]

Off-device attribution is an interesting direction to explore for advertising attribution. It is important that the proposals in this space meet the scalability requirements of the real world traffic loads, however. The current scale at which these proposals (such as the Interoperable Private Attribution) have been evaluated so far seems to be at the lower end of real world traffic loads. We want to better understand the scalability characteristics of these off-device attribution proposals.

We propose that the off-device attribution proposals be evaluated at the following ad-event volume scales: 10M, 100M, 1B, 10B, 100B. This would help in understanding the performance for advertisers of various sizes, supporting queries which span across multiple advertisers and supporting queries across wider date ranges. For the evaluation, ad-event and unattributed conversion trigger volumes should be considered to be at an identical order of magnitude (i.e. at the high end, tests of 100B ad events and 100B conversions).

[benjaminsavage]

Thank you @saxena-shobhit for filing this issue!

Wow… 100 billion is a large number!!! Thanks for inspiring us to shoot for such a Big Hairy, Audacious Goal (BHAG).

I would be happy to discuss ideas for how we might scale IPA to these volumes on Github issues, or in our biweekly call (patcg-individual-drafts/ipa#47).

We recently mentioned horizontal scaling as one of our most important “Open Questions” in our talk at “Real World Crypto” (https://m.youtube.com/watch?v=Q3glyMsaWIE&t=3340s). We would welcome any academic collaborators who are interested in this question!

I recently filed an issue with a sketch of an idea for how to address this challenge: patcg-individual-drafts/ipa#49, and I would love to hear other early sketches for alternative ideas!

I suspect scaling will be a major focus of ours in the coming months. Let’s see if we can reach these ambitious targets!