We try to add scatter for a table with nine indexes, which makes PD CPU increase to 2000%.
Because it will add n+1 scatter-range-scheduler when we add scatter to a table with n indexes and every scatter-range-scheduler
contains balance-leader-scheduler and balance-region-scheduler in the range cluster which specified the start key and end key.
Describe the feature you'd like
we need a more flexible and more unified scatter-region-scheduler.
Describe alternatives you've considered
Flexible means we can choose any index or table. And it can run steadily.
Unified means we can add all scatter to the balance-leader-scheduler and balance-region-scheduler, which doesn't cost external threads and can reduce conflict with the balance-leader-scheduler and balance-region-scheduler. Of course, we also need to create operators batch to avoid being too slow.
lhy1024
commented
1 year ago
Feature Request
Describe your feature request-related problem
We try to add scatter for a table with nine indexes, which makes PD CPU increase to 2000%.
Because it will add n+1 scatter-range-scheduler when we add scatter to a table with n indexes and every scatter-range-scheduler contains balance-leader-scheduler and balance-region-scheduler in the range cluster which specified the start key and end key.
Describe the feature you'd like
we need a more flexible and more unified scatter-region-scheduler.
Describe alternatives you've considered
Flexible means we can choose any index or table. And it can run steadily.
Unified means we can add all scatter to the balance-leader-scheduler and balance-region-scheduler, which doesn't cost external threads and can reduce conflict with the balance-leader-scheduler and balance-region-scheduler. Of course, we also need to create operators batch to avoid being too slow.
Teachability, Documentation, Adoption, Migration Strategy