Shared boxes: not worth the price savings
Having worked at IBM ~15 years ago, I implemented ECC-code generation. The standard then was that any one-bit error was correctable, and any two-bit error was detectable. By default, such errors would result in a machine fault interrupt, which were generally considered non-recoverable (outside a reboot).
Memory was periodically read & written back to ensure that one-bit errors would be corrected.
So, I don't consider rowhammer-class attacks to be serious if ECC memory is in use, except possibly (possibly) as a DOS attack.
On the other hand, on shared hardware, this becomes yet another headache. In addition to the destructive performance loss of these software hurdles against Specter-class attacks, now you need to pay for ECC memory (and the performance loss in keeping it fresh). Go with dedicated boxes, and you can avoid all of this...