Major Description
When a scan inits a tablet iterator, the tablet iterator inits iterator for each rowset that is deemed relevant. When there's a lot of rowsets (usually because of some missing feature like KUDU-1400 or a pathological partition schema or configuration), this can take a long time, leading to scan timeouts like
WARNINGS: Unable to open scanner: Timed out: Scan RPC to 10.1.11.187:7050 timed out after 169.988s (SENT)
For non-fault-tolerant scans, it seems like we should be able to init the rowset when we first go to retrieve rows from, and thereby amortize all the seeks to open rowsets across many ScanRequest RPC round trips.
For fault-tolerant scans, things are more complicated, but still it should be possible to be lazier about initting rowset iterators.