Review Postgres CPU specs on Reference Architectures (Q3 2020)
We were asked recently if the current Postgres CPU recommendations on the Reference Architectures were adequate. Indeed reviewing the state of play today there's perhaps not as much headroom today as we'd like (10k):
We unfortunately don't have data immediately to hand for older versions that first influenced the Architectures and if this is in fact a recent product degradation. Testing on older versions may need to be redone to confirm.
Upping the specs unfortunately will impact some customers that are already running with the architectures so we need good justification here to do so. That said the specs today are performing fine on the whole outside of known problem endpoints but the headroom here is not as comfortable now as we'd like.
For reference the current specs we have are:
| Users | Postgres vCPU |
|---|---|
| 1k | 2 |
| 2k | 2 |
| 3k | 2 |
| 5k | 2 |
| 10k | 4 |
| 25k | 8 |
| 50k | 16 |
Investigations:
-
Get current database performance of all reference architectures. -
Do a test run of 10k on higher specced Postgres (8 vCPU) -
Investigate performance data of older versions (mid 12.x, 10k) for comparison. (Investigation thread) -
Investigate if Postgres vacuumdb needs to be performed manually before each performance test and if it impacts CPU usage notably. (Investigation thread) -
Collect and monitor Postgres CPU trends over time with latest nightly packages to see if any patterns form. ([Investigation thread(#293 (comment 379805918)))