• Is there a data/index file format specification similar to Voldemort’s, or is operating on that level discouraged (i.e., just write Avro-encoded data)?

The input to the push job can be a collection of regular

AvroSequenceFile

but then Venice reformats the data internally. In the end, the data lands in the servers in RocksDB format, and we support both RocksDB formats (block-based, which is optimized for SSD + RAM, and PlainTable which is optimized for being all-in-RAM). More generally, all R/W interfaces are Avro (and some legacy “VSON” API carried over from Voldemort, but which we’re not actively extending…). We have explored supporting other APIs (e.g. Protobuf) but that is not available at this time.