On 1/7/19 6:34 AM, Tsunakawa, Takayuki wrote:
1. Doesn't provide precise data
Sampling could miss intermittent short waits, e.g., buffer content lock waits
during checkpoints. This might make it difficult or impossible to solve
transient performance problems, such as infrequent 100 millisecond response
times while the normal response time is a few milliseconds.
The proposed wait event collection doesn't miss anything.
2. Overuses resources
We may be able to shorten the sampling interval to 10 ms or even 1 ms to detect
short periods of problems. However, the sampled data of active sessions become
voluminous in memory and storage. It would take longer to analyze those
samples. Also, the background sampling process prevents the CPU core from
becoming idle to save power, which bgwriter and walwriter tries to avoid by
hibernation.
The proposed wait event collection just records what actually happened. No
waste. Would it use many resources if waits happen frequently? That leads to
our motivation to reduce waits.
FIY, wait events have been added in PoWA by using pg_wait_sampling
extension :
https://rjuju.github.io/postgresql/2018/07/09/wait-events-support-for-powa.html
pg_wait_sampling sample the wait events in shared memory and PoWA store
them.
