This solves a few problems we've been having and simplifies some other things:
The namespace is now independent of the bucket name, which is now particular to each input or output key instead of globally declared.
Directories are now stored directly rather than archiving them and uploading the archive. This simplifies the layout and makes it easy for end users to download the entire tree or browse with the console or gcsfuse.
Log output was coming through in batches of 32, now it displays each line as it arrives.
This solves a few problems we've been having and simplifies some other things: