rfc: dynamic routing candidates

[alanshaw]

📖 Preview

Summary

Proposal for an alternative to the presupposed "static" affinity routing where ALL writes to a given space are routed to a specific storage node.

[bajtos]

LGTM!

[parkan]

I am generally in favor of minimizing central state tracking as much as possible but I'm a bit anxious that this could lead to e.g. dropping multiparts on the floor without an easy way for anyone to find them after clients think everything is ACK'd = pushes dealing with that to Ingot/Guppy actual S3 client impls? if that's acceptable 👍 but I feel like it doesn't quite match S3 semantics

(ref #3 Async-ACK under https://github.com/fil-one/RFC/blob/main/rfcs/2026-05-filone-forge-deployment-proposal.md#the-upload-path)

[alanshaw]

I'm a bit anxious that this could lead to e.g. dropping multiparts on the floor without an easy way for anyone to find them after clients think everything is ACK'd = pushes dealing with that to Ingot/Guppy actual S3 client impls?

I'm not sure I follow...

[parkan]

I'm a bit anxious that this could lead to e.g. dropping multiparts on the floor without an easy way for anyone to find them after clients think everything is ACK'd = pushes dealing with that to Ingot/Guppy actual S3 client impls?

I'm not sure I follow...

consider this situation:

• client does a multi-part upload with dynamic affinity, ACK'd at S3 boundary
• node crashes/restarts partway through processing the multipart pieces (i.e. before accept on all chunks)
• if the client (or something) doesn't keep add task state until everything is actually accepted not just ACK'd -> dangling blob(s) that needs reconciliation and doesn't have a receipt that links it to a provider
• also what happens if the affinity set changes upon retry?

I might be misunderstanding the receipt retention design? (but the way I'm reading it another problem is that all receipts need to be retained for entire blob lifecycle as the only authoritative source of truth, otherwise removes would break -- the receipt chain + indexer design also triggering some IPNI PTSD but that's neither here nor there)