Each 32 MiB chunk is its own request and opened a new SSH connection and
SFTP session (~200 ms). A worker now keeps its connection to the host for
the next chunk; each request still opens its own SFTP session (~50 ms),
so an SFTP error never reaches the next one. A connection is not reused
after an error, when the host dropped it or once it is 10 minutes old; an
idle one is closed after a minute, and a forked worker leaves the parent's
to the parent. The SFTP start, which paramiko waits for without a deadline,
is ended after 10 s. Live: 256 MiB in 8 chunks 4.5-4.9 s -> 3.7 s.
Each 32 MiB chunk went out as 1024 SFTP writes, each waiting for its
answer: 19 MiB/s on a gigabit LAN, ~1.9 s per chunk with the SSH session.
The writes now go out without waiting and their answers are read as they
come and at the end: ~0.53 s per chunk, the same file (checked by sha256).
paramiko's own pipelined writes (set_pipelined) drop the errors in their
answers, so a full disk would leave a short ISO behind a successful
upload. Each write is registered to its file instead, the way paramiko
does for prefetched reads: every answer, in any order, reaches the file
and an error is raised as soon as one is read.
The upload session was recorded only after the first chunk had been
written, and a retried first chunk was refused when its part file was
missing. The session (with its upload id) is now written before the
part file is created, the part file is created by whichever chunk finds
none, and a missing part file is an error only after the first chunk.
An interrupted first chunk can be sent again without adopting an
unrelated file.
