coin selection code does very badly in some cases

[dooglus]

Take a look at transaction ce9c9afee7b3dd1deebdc49ceb1560d2771acb4cd0a4725042beecde726a3463.

He was trying to send 0.0011 BTC, and was charged 0.0445 in fees. That's a 4045% charge. According to the user, there were "a lot of better fitting inputs".

The coin selection code doesn't attempt to minimise fees. Perhaps it should.

The transaction spent the following outputs:

571 single satoshis, plus these 15 amounts which sum to 0.05559429:

0.00007550 0.00009679 0.00011887 0.00014177 0.00019014
0.00042293 0.00045708 0.00049249 0.00056730 0.00134063
0.00178897 0.00187366 0.00196149 0.01070019 0.03536648

If just the last of those 15 was used, the fee would have been either 0.0005 or 0.

[murchandamus]

I have suggested an improvement at #4906 and would appreciate feedback.

[sipa]

I believe #9404 may also have improved upon this.

[maflcko]

@sipa I don't think so, but I might be wrong. The issue is that ApproximateBestSubset right now will try to create an exact match, as was the case when this issue was filed. The bug you are referencing solved a fee calculation bug.

[Sjors]

Recent discussion on coin selection #12605

[maflcko]

Let's discuss further steps in #12605