v3.44.0

New

• Each tenant can have its own logo, uploaded by the tenant administrator under Admin Settings. It appears wherever the logo is shown.
• The cookie notice is shown in the new interface too. It used to appear only in the classic interface, so someone who used only the new one was never asked.

Fixed

• Uploads to remote storage could be sent to deleted nodes and hang. Deleted nodes are no longer chosen for storage, and a node that answers with an error is skipped in favour of the next one.
• An upload in a single piece to remote storage never went through, and failed with “None of the nodes is reachable”, which pointed at the wrong cause.
• The scheduler could stall a few seconds after a restart and stop doing anything while still running. A brief database outage made it more likely.
• Port and tunnel handling is more robust, on the nodes and centrally. Concurrent workers no longer compete for the same ports and close each other’s live sessions, a node no longer restarts in a loop when its tunnel state cannot be read, shutdown takes seconds, and one silent upstream no longer blocks every other client.
• Superadmin: impersonating a tenant sent you back to the login page, and the button to leave a tenant was never shown, so there was no way back. Both are fixed.

Changed

• Fewer database connections. Connections are shared within each process instead of every component opening its own pool, which matters most on a shared database server.
• Faster scheduler failover on Kubernetes. With two replicas, the standby takes over in about a minute when a node fails, instead of the five to six minutes it takes Kubernetes to move a single pod. The lock behind it now causes a tenth of the API traffic it used to.