A transit AI team evaluates ECMP load distribution for a collective that is sensitive to packet reordering. Which judgment is appropriate?

A transit AI team evaluates ECMP load distribution for a collective that is sensitive to reordering. The answer is to prefer load-distribution methods that minimize harmful reordering—per-flow or per-prefix hashing keeps packets for a single collective flow on the same path, so in-order delivery is preserved while traffic is still distributed across spines. That balance protects the collective. Preferring per-packet spraying across spines with tuning that limits harmful reordering for that collective while improving path diversity still sends consecutive packets of the same flow over different spines, so latency differences can still reorder packets and stall this order-sensitive collective. Preferring flowlet switching with short inactivity gaps that preserve packet order for that collective while spreading load across more spine paths is contradictory because short gaps make flowlet splits more likely, so later packets can still take a different spine and reintroduce the reordering this collective cannot tolerate. Preferring mid-collective rehashing of flows with order-aware safeguards that curb harmful reordering for that collective while balancing spine use still changes a flow's spine in flight, which is exactly the kind of mid-collective path change that risks packet reordering this collective cannot tolerate. Exam caveat: the hash function must be stable across the collective duration so the same flow uses the same spine throughout; mid-collective rehashing of flows with order-aware safeguards that curb harmful reordering for that collective while balancing spine use still introduces the in-flight path change and reordering risk the question warns against. Operational check: enable per-flow or per-prefix hashing on ECMP, run a reorder-sensitive collective test, and show packets for a single flow stay on the same spine while different flows spread across available spines.