Comment on page
Reduce switch memtable latency
The MemTableRepFactory class (each memtable representation inherits it) can be set as a “PrepareMemtableCreation” which will create a switch background thread that is responsible to construct a new memtable object to be used when switch occurs. A new memtable will be created when we fetch this memtable, so we'll always have one memtable available for use.
In case we don't have a ready memtable (this can happen if the switch memtable happened before the previous trigger was triggered) we create the memtable immediately .
To be able support this ability , memtable representation needs to divide the memtable construction to pre/post phases . This allows us to get the correct memtable’s parameters as defined in the mutable CF configuration.
The MemTableRepFactory class (each memtable representation inherits it) can be set as a “PrepareMemtableCreation” which will create a switch background thread that is responsible to construct a new memtable object to be used when switch occurs. A new memtable will be created when we fetch this memtable, so we'll always have one memtable available for use.
In case we don't have a ready memtable (this can happen if the switch memtable happened before the previous trigger was triggered) we create the memtable immediately .
To be able support this ability , memtable representation needs to divide the memtable construction to pre/post phases . This allows us to get the correct memtable’s parameters as defined in the mutable CF configuration.
1 - A switch memtable (size of the write buffer, threshold of the write buffer manager, wall size etc.) does not have a memory construction penalty and will allow for fluent acceptance of new writes.
2 - Simplified code
1 - Each CF will have two mutable memtables (one active and one on hold), so we multiply the memtable construct memory usage.
Last modified 11mo ago