Conversation
Replying to
Your benchmark illustrates that tuple deforming is now a big bottleneck. About 105 seconds were spent *preparing* for sort in all 3 cases. I doubt that you're I/O bound there. Parallel CREATE INDEX will help, but we should also directly target the deforming bottleneck.
Replying to
I was I/O bound, my SSD capped at ~280Mo/s. I should try with better storage.
1
Replying to
Interesting. The actual index build at the end doesn't take very long, so I assumed that it wasn't.