At the moment for INSERT_SELECT and LDI we turn the row format into csv and pipe into cpimport. This piping only has a 1MB csv row buffer. BLOB lengths are doubled due to hexing so at most you can fit a single 512KB BLOB in a row before things error.

The DML commands don't realise something is wrong until the end of the bulk import. The table remains locked afterwards and needs manual unlocking.

I believe the best way to fix this is to have a row length counter which spots this and errors appropriately as well as having a configurable buffers size for the row.

Alternatively it might be possible to make the row buffer grow with the content.