They also have different scaling tradeoffs as they scale really well (the graph can be arbitrarily wide as tx's can run in parallel) when most transactions are causally unrelated but hit bottlenecks with every "killer app" that gains traction (graph narrows as more and more transactions become causally related).

I'm not sure what domains see this problem crop up but I've seen it in quant finance strategy execution. Say you have 1k strats that you want to run in parallel but risk needs to bound the bank's per-equity positions globally. When the strats work on different subsets, DAG approaches aren't a problem, but when most of the strats trade APPL/IBM your once very wide (parallel execution) graph narrows significantly (becomes more sequential) as the strats need to check with risk w.r.t. AAPL/IBM sequentially.

NB: for this reason, causal approaches are pretty rare to see today (though it depends on the domain as high latency contexts don't really care).

I know that there might be spam problems if one manages to precompute the local PoW hashes (solved by pruning the chain), but I would love if you explained more this security problems. Are they consensus-related as some people in reddit suggested? If this is the case, it might be good for Nano to get someone like @aphyr to audit the code.