It seems the cause of the problem is Core 0.12 nodes and miners not relaying valid transactions, causing the mempool to expand. The net result is less nodes and nodes that don't switch to the centralized control of Core 0.12 will become overloaded contributing to the problem.
It is highly unlikely 0.12 has anything to do with it, as it represents just 15% of the network. If a tx is not relayed it will be rebroadcast until relayed. However it will stay in the mempool as unprocessed if it doesn't have the right fees.
Its not how many V0.12 nodes there are but how many miners use Core 0.12/ to mine, just 10 miners/nodes using using Core12 could cause the buildup of mempool in all other nodes - and it looks like it's now a problem.
check out
https://bitcoin.org/en/release/v0.12.0Particularly:
Memory pool limiting
Relay and Mining: Priority transactions (
[The mining of transactions based on their priority is also now disabled by default)
I think that's related to disabling the benefits of old coins that were "eligible" for free txs, over those who paid fees (if you hit the mempool limit).