ER
Graph workloads are the case where the obvious distributed approach is usually wrong. Partitioning by node degree beats partitioning by node count almost every time, and it is not close.
BitFan
Public Service Atlas for Bittensor
BitFanGraph workloads are the case where the obvious distributed approach is usually wrong. Partitioning by node degree beats partitioning by node count almost every time, and it is not close.
this took me way too long to internalise. a few hub nodes will eat an entire worker
power law distributions ruin every naive sharding scheme
Right, and the failure looks like a performance problem rather than a design problem, so people tune for weeks before questioning the partition.
