Why Would Anyone Do Refunds?
Sometimes, a payment can’t proceed as planned. This usually happens when nodes go offline, the channel capacity decreases, nodes change their behavior (e.g. fees) or nodes misbehave for other reasons.
When that happens, the initiator could just start a new payment over a different route until it succeeds. But the funds used in the first transfer will stay locked until the failed payment times out, thereby reducing the networks liquidity.
Refunds return the funds to undo the payment to a certain degree. If the funds are returned all the way to the initiator, he can remove the lock by revealing the secret. Refunds to mediators allow mediators to retry the mediation along a different path.
At the moment, refunds are always propagated all the way back to the initiator. The initiator does not do anything on receiving the refund and the payment times out. This is the worst possible apporoach because:
Refunds don’t provide any benefit
Refunds lock up additional funds until the payment times out
Refunds significantly increase complexity of the mediation code
The light client has an incomplete and untested refund implementation, and Andre prefers to “research better approaches to the issue than to revive the Refund”.
Without significant improvements in our refund handling, the benefits are dubious. The upside of removing refunds is quite clear: reduced complexity in implementation, testing, documentation and the overall concept.
Fix Existing Refunds
The existing implementation can certainly be improved, but there are open questions for nearly every potential improvement. Some of the questions are:
Should the refunds go back along the same route (might lack capacity) or choose a new route (who will pay the PFS)?
How likely are refunds to succeed?
Is the increased liquidity due to early lock release bigger than the amount of tokens locked by the refunds during the refund process (and until the timeout if the refund fails)?
Will refunds be visible to end users? How?
Should refunds incur mediation fees?
Should mediation fees for the forward payment be refunded?
Should the initiator reveal the secret for a fully refunded payment even when he will have to pay mediation fees?
Try to Improve on Existing Refunds
Even if the current refunds are useless, they might serve as a foudation for a better but similar refund mechanism. Things to consider:
Can we remove the liquidity requirements for refunds by going back along the original route? The refund might be able to free the locked liquidity on each step instead of locking up more tokens, even before the secret is revealed
The mediation code could be simplified by limiting the allowed cases (e.g. only refund back all the way to the initiator, no new routes tried by mediators)
Use Different Refund Approach
There are other ways to do refunds or cancel payments, which might be worth further investigation:
One suggestions by Augusto: https://github.com/raiden-network/raiden-contracts/issues/1216 (could not be made into a viable suggestion, yet)
Lightning has some kind of refund or cancellation mechanism
Final Decision by the Team
After discussion, only two options were left that could be implemented without huge effort:
Do refunds back to the initiator, including all fees. Revealing the secret will free up the locked funds, without moving any tokens. This way, everyone has an incentive to participate.
The seconds approach still has the problem of locking up additional funds and potentially failing. The benefits are too small to justify the increased complexity. So the recommendation is to remove the refunds.
This discussion was based on the assumption of strict routing, i.e. strictly following the route planned by the initiator. If we allow retrying different routes at any node, the results might be different.