1) It's not fair at all to say that Google "continue to ignore the two critical issues of uptime and data store latency". The HR datastore is specifically designed to address the concerns about variable latency of datastore operations, and to prevent both planned and unplanned downtime.
2) In my experience, high CPU cost for datastore operations is generally tied to things like having large numbers of indexes, or doing queries that don't have the right indexes. Each index written involves a bigtable write, if you are doing hundreds of these per entity it can become very expensive. That said, in general it isn't easy enough to know when you are doing something that will be expensive.
3) The bulkloader is definitely painful to use if you aren't familiar with it. In particular, casting schemaless data to a format such as CSV is hard because of the problem that rows can have unexpected keys. An improved "Bulk Datastore Import and Export tool" is on the 6-month roadmap for the product.