Attribute Clustering requires a sort of the rows before they are stored to disk.
If the rows does not match into PGA, the Sort must be done using TEMP. This can increase the time needed for INSERT dramatically.
To mitigate this effect, I suggest to allow a quick&dirty ATTRIBUTE CLUSTERING: whenever the Sort Area in PGA is totally used, save those rows to disk.
This will lead to sorted sub-chunks on disk, all of the max PGA size.
Even the data will not be that perfect ordered, it can be a reasonable trade-off between un-sorted inserts and slow (dribven by TEMP-Sort) inserts?
This database idea has been accepted and raised as: ENH 27131584 - ENHANCE ATTRIBUTE CLUSTERING TO DO "QUICK & DIRTY" WORK