Comment by JoshTriplett

There's no fundamental reason for this. Much as with per-CPU memory allocation, shards could pre-reserve chunks of the next serial space, and then only coordinate when they need a new chunk.