FYI, posting an intermediate update on the issue.
I disabled index scans to keep existing order, and copied part of the "slow" range into another table (3M rows in 2.2 GB table + 17 GB toast). I was able to reproduce slow readings from this copy. Then I performed CLUSTER of the copy using PK and everything improved significantly. Overall time became 6 times faster with disk read speed (reported by iotop) 30-60MB/s.
I think we can take bad physical data distribution as the main hypothesis of the issue. I was not able to launch seekwatcher though (it does not work out of the box in Ubuntu and I failed to rebuild it) and confirm lots of seeks.
I still don't have enough disk space to solve the problem with original table, I am waiting for this from admin/devops team.
My plan is to partition the original table and CLUSTER every partition on primary key once I have space.
Best regards,
Vlad
