how much disk space is required to switch from SizeTiered to level compaction strategy in Scylla db?

Question

I have a 20 node scylla db cluster and each node is at 70% disk space usage. I want to switch to leveled compaction strategy from Size tiered strategy. Can I do that with this much disk space left? How much disk space at max will be required?

Answer 1

As was mentioned, you don't have enough disk space to switch compaction strategies. I can think of 2 or possibly 3 ways to work around this.

1. add more disk space - if you are in a cloud environment, it probably won't be a problem to grow your data disks and expand the file system.
2. add more nodes - however, this will take a while because the data will have to be streamed to all the new nodes and then you'll need to run nodetool cleanup on the cluster to reclaim the unused space.
3. change the compaction strategy on one table at a time. Start with small tables that can fit in your 30% space. this will also take a long time and there is the risk that you'll run out of space if you choose the wrong table to ALTER.

Answer 2

Generally with Size Tiered Compaction Strategy or Leveled Compaction Strategy, you want to keep about 50% of disk space free.

We have a Scylla Enterprise feature, Incremental Compaction Strategy, that is much more suited to run comfortably at 70% disk utilization.

Changing compaction strategies, and re-writing all of your current tables will require twice as much disk space as you are presently using. (140% of what you are using at present rate.) So it looks like you might need to temporarily expand your cluster in order to run a compaction properly.

From what I understand, that would first take time to rebalance the nodes evenly. Then you should run a full compaction using your existing strategy, and change to your new one. Then, if you had saved room through the compaction, you could then decommission any unneeded node(s).

If you have more specific questions, I'd invite you to join our Slack to ask our community members directly.