When the controller rejects a request it cannot decode, the error returned to the client now explains why the payload was invalid instead of only saying it was invalid
The VM controller no longer crashes when several constellation, batch or VPN-orchestration requests reach the same worker VM at the same time
Opening a new file transfer or terminal stream on a busy connection no longer pauses the processing of incoming data on that connection
The VM controller no longer crashes when a connection fails while it is still sending queued replies or stream data
When a connection drops while the controller is still sending to it, pending operations now fail right away instead of hanging. Shutdown no longer waits up to two seconds or gets stuck on streams that were still open
A worker or batch node VM shutting down while the controller is sending it a request no longer crashes the controller. The request now fails with an error instead
The VM controller now pings its connections every 30 seconds and drops a peer that stops responding for three minutes, instead of hanging forever. A peer that stops reading no longer blocks every request on that connection, because each message write now times out after two minutes
The VM controller now signs its connection handshake with FIPS mode on, instead of briefly turning FIPS mode off for the whole VM, and it no longer turns FIPS mode on for images that have it turned off
SLURM batch jobs can now request up to 100 TiB of memory per node. Previously, jobs requesting more than about 1 TiB were rejected
