Partition::clear

[marvin-j97]

Just run a "Nuke compaction" (followed by dropping all blob files if blob tree) - clear() is not transactional!!!

[anidotnet]

Does this clear() method intends to delete all items from partition? Because, I am looking for such operation and I could not find one for a Partition.

[marvin-j97]

Yes. Generally I find clearing a table to be an anti-pattern, because it's faster to just create a new partition. But with clear() it's possible to clear it in O(1) and reuse the partition name, so it's just more convenient, but not transactional.

[anidotnet]

Thanks for your quick response. Could you please give an estimate by when this api would be available in release?

[marvin-j97]

Not right now, it should be pretty simple, there just needs to be some coordination added to make sure there are no outstanding compactions going on when clearing (there needs to be a write lock & the clear() needs to make sure it gets all the disk segments) - which would also allow fjall-rs/lsm-tree#105.

[anidotnet]

So in it's current form how to clear a partition efficiently instead of iterating and removing each item?

Provided, in my use case I am holding a Partition in a struct and calling clear from inside the struct's method. The struct is a map like data structure and I am not planning to have any mut &self method for the struct.

Holding a keyspace object instead of a partition in the struct and opening partition for each map operations is not feasible due to the write lock in open_partition method which will significantly harm bulk operations.

[marvin-j97]

So in it's current form how to clear a partition efficiently instead of iterating and removing each item?

Currently none, there are two ways really, either 1) a range tombstone, which can be transactional, or 2) this issue, which can not be transactional because it will simply nuke all the data.

Holding a keyspace object instead of a partition in the struct and opening partition for each map operations is not feasible due to the write lock in open_partition method which will significantly harm bulk operations.

It's an RwLock, so for the most time, the lock should not actually lock.