At one point, the tangle might grow so large or fast that a single iri node can't be aware of all transactions happening. How to have consensus about address balances without iri nodes trusting each other? How to create snapshots if there's no single node that sees the entire tangle?
A single IOTA node can already handle around 1000 transactions per second, this is quite a bit but probably not enough in the future. Once this becomes an issue sharding can be a solution. This means that only a part of the tangle is handled by cluster of nodes that for example only handle transactions starting with the letter 'A'; Another cluster will be responsible for transactions starting with the letter 'B', etc. I think the IOTA team still has some time to solve this issue completely before it becomes an issue.