multi-seller auction vs on-device resources

[jonasz]

Hi all,

In Mode B we're primarily observing single-seller auctions. Long term, we are expecting there to be multiple sellers in each PA auction.

In my understanding, in a multi-seller setting, the amount of computation will scale as number_of_buyers times number_of_sellers.

This will have a big impact on the amount of resources required to convey a full on-device auction. Publishers (or sellers who run the toplevel auction on their behalf) will face a tradeoff: either the latency will grow considerably (hurting UX and ad effectiveness), or they will have to impose stricter limits on the auction, possibly limiting buyers' and sellers' ability to take part in PA auctions.

We'd like to gain confidence that this won't happen. Unfortunately, as noted at the beginning, this is not something we can verify in the ongoing tests.

Please let me know, what are your thoughts on this?

Best regards,
Jonasz

[MattMenke2]

If all sellers work with all buyers, that is true. If each seller works with a distinct set of buyers, or each buyer works with at most some fixed constant number of sellers, it's just the sum of the number of buyers (Edit: Times that constant factor, of course).

[jonasz]

Hi Matt. Of course your comment is mathematically correct, but this is exactly what we're worried about! One worry is that sellers will have to limit the number of buyers they work with, and another is that the number of sellers taking part in an auction will have to be reduced as well. How can we, as a buyer, make sure that we will continue to have access to inventory? How can we make sure this access is diverse and robust?