server/{market,swap,db}: free cancels

[buck54321]

server/db:
Add epoch_gap column to cancel order table. When set (not EpochGapNA),
the epoch_gap is the number of epochs between the limit order placement
and the cancel order placement. Update InsertEpoch method to accept an
epoch gap argument (or EpochGapNA).

server/auth:
RecordCancel now accepts an epoch gap argument. Epoch gap is loaded from
the db via ExecutedCancelsForUser with a modified return value.
(*latestOrders).counts now only counts cancel orders if 0 <= epochGap < 2.

server/market:
processOrder now reports the epoch gap to the AuthManager for matched
cancel orders.

[buck54321]

This is the line that matters. This is a rule change. Free cancels after the order sits for one full epoch. If the order is submitted in epoch 10, and the cancel order submitted in epoch 10 or 11, that's a "short" cancel (not a free cancel), and the cancellation can hurt to the user's score.

[chappjc]

Have not looked at the code, but as discussed, this is a very reasonable compromise to outright eliminating the cancellation ratio limit that is overly restrictive. I think a very small freeCancelThreshold on epochGap, which you have proposed here, would not create much more of a challenge for mesh than already exists with the current order history knowledge requirements since it's necessarily very recent history. Good stuff!

However, I still view the cancellation ratio limit as largely unnecessary given the epoch-shuffle matching algorithm, which makes it risky to cancel rapidly for deceptive purposes, combined with the need to have unique coins funding every order and the various order quantity limits, which throttle the ability of a user to place an excessive number of (presumably bogus) orders.

[buck54321]

I think this will improve the character of our order books. Human users should have a fair chance to match with orders that they see. Whereas before, an order could be 1) placed and canceled in the same epoch, or 2) placed near the end of an epoch, and canceled in the next epoch. In either of those cases, there's a chance that I could see that order, place my order immediately, and then that order is canceled before my order comes up for matching. It's like probabilistic spoofing. But if you wait for the free cancel, that's eliminated.

If you truly make a mistake or something and want to cancel quickly, you still have that choice. Cancellations are relatively low impact in terms of penalization. Human traders will never run up against the cancellation ratio. But bots are encouraged to trade in a way that makes the book more stable and usable.

[chappjc]

Well put. I think this is a great balance.

Although I think users always did have a fair chance because as you said its a probabilistic spoofing game, not without risk to the would-be spoofer.

[buck54321]

Although I think users always did have a fair chance

I guess that's true, in a sense. It think the proposed rule change gently incentivizes behavior that contributes to a better trading experience, which was the intended effect of cancellation limits.

In regards to previous conversations regarding market maker order book management, I think the proposed rules are relatively easy to satisfy without a large disruption to existing protocols. All of that careful order management I was talking about is unneeded.

[chappjc]

Looks great. Some minor nits.

[chappjc]

I'm thinking this one should perhaps use a gap of 0. Yes, the user will get the preimage miss strike too, but this is effectively cancelling an order in the same epoch. Redundant with the preimage violation?