I am sorry to open one more thread about this topic, but I though I understood roughly how distributed deletes work and while testing it, it appeared that I don't.
I have deleted a lot of counter columns and "normal" rows and columns.
I am using a 1.2.6 single node cluster. I reduced the gc_grace for every CF (from 864000 to 60) and started a major compaction (I am using STC only).
This process removed all ghost (as far as I can see with OpsCenter) from some CFs and didn't remove any of the tombstoned rows on other CFs.
Specifically, major compaction removed the data from a CF and not from an other one defined in the exact same way and using the same code to delete the columns.
What am I missing ? Is there a random possibility not getting ride of range ghost after a major compaction ?