Split Until You Cannot

A managed datastore automatically splits a partition when its access pattern becomes uneven, isolating frequently accessed items. A team relies on that behavior and still sees sustained throttling on one item id. Why?

Splitting rebalances the write path only, leaving every read on the original partition until an index rebuild is run manually

A single item lives entirely in one partition, so no split can subdivide it and its throughput stays bounded by one partition's ceiling

Automatic splitting is triggered by stored data size and never by request rate, so a pure access hotspot is not detected at all

The split does occur, but clients keep routing to the old partition until their cached routing metadata expires and refreshes