You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently our pipeline uses poller mode for scheduling. This leads to some of the following problems.
An unintended blocking implementation may result in all queries not being scheduled in a timely manner.
poller may not be scheduled in time.
The pipeline poller consumes an entire CPU core, resulting in a waste of resources.
Describe the solution you'd like
The main purpose of this PR (#49972) is to remove pollers through the observer notify mechanism, and we will continue this work and make better improvements.
It can save one CPU core in the scenarios such as point query/TPCDS.
In the new implementation, we don't need a poller thread at all, or even a block driver queue. And there is no need to introduce additional locks during scheduling.
Feature request
Currently our pipeline uses poller mode for scheduling. This leads to some of the following problems.
Describe the solution you'd like
The main purpose of this PR (#49972) is to remove pollers through the observer notify mechanism, and we will continue this work and make better improvements.
It can save one CPU core in the scenarios such as point query/TPCDS.
In the new implementation, we don't need a poller thread at all, or even a block driver queue. And there is no need to introduce additional locks during scheduling.
The text was updated successfully, but these errors were encountered: