When is it required to initiate an intentional partitioning of the Tangle?
If two parties are disconnected from the internet (no connection to main Tangle) temporarily and still want to send Iotas to each other, they could do that in an offline subtangle. Further explanation can be found here:
- How is consensus achieved in offline transactions?
- How do subtangles work?
- How to start a new offline subtangle?
- What happens to a Subtangle if it attaches to an invalid transaction?
In the future, it would be overkill, if every node would process all transactions. If we have millions of transactions, it would even be physically difficult, probably impossible for every node to to process all transactions. These clusters would almost act like different cryptocurrencies, but they would follow they almost the same (or exactly the same) rules and could be merged or at least synchronized if needed.
This sounds like an inconvenient solution and I'm not denying that it is, but you can't argue with physics. And to be honest, the coffeemaker in China doesn't have to know that the drone in South Africa just got paid for delivering a package.
Since it's not necessary to do economic clustering now, everything about the topic is very experimental. Maybe we find a different solution to the problems that is not at all similar to my description. But this is how it could work.
Also: Article on EC by CFB
There are probably more reasons that no one thought of yet. Point is: Partition tolerance is a useful and necessary feature of the Tangle.
How is the Tangle able to survive long-sustained partitioning of the network without the reversal of a large number of transactions?
As long as the top and bottom transactions don't conflict e.g. try to spend the same funds twice, this is possible.
Since a block can only reference one previous one, even if there are no conflicting transactions at the top and bottom, the transactions can never be merged again and the top transactions have to be redone. This is a typical blockchain fork where the longest chain wins.