So Meltano runs singer taps and targets A single tap uses a fd (stdout) piped to a single targets fd (stdin). So even if a tap is running with 20 tables selected, they are all going through the same fd. Singer supports getting messages in any order so long as a schema message precedes a stream. So if you think, BFS (breadth first search), correlate that to the fact a tap could stream all of its streams "simultaneously" but ultimately really just alternating stream messages Or DFS (depth first search) if we follow the comparison where a tap streams a single stream to completion then the next. At the end of the day it's all going through the same proverbial pipe. So it doesn't really matter.

True parallelism (single or multi node) comes from multiple processes And that means taking a single tap-xxx target-yyy and partitioning the selected streams Then executing the partitions simultaneously if supported by target/dest

BATCH records are in the works in singer_sdk now if I'm not mistaken.