EstimateMedianVal returns higher fee for higher confTarget

sdaftuar commented at 8:16 pm on December 19, 2020: member

Running on recent master, commit f1dbf92ff0475a01d20170ea422c1d086acbbc57

If I run:

0test/functional/feature_fee_estimation.py --randomseed 108574360997204915

I get:

 02020-12-19T20:14:06.175000Z TestFramework (INFO): Initializing test directory /tmp/bitcoin_func_test_bcwexp78
 12020-12-19T20:14:07.738000Z TestFramework (INFO): This test is time consuming, please be patient
 22020-12-19T20:14:07.738000Z TestFramework (INFO): Splitting inputs so we can generate tx's
 32020-12-19T20:14:10.870000Z TestFramework (INFO): Finished splitting
 42020-12-19T20:14:12.104000Z TestFramework (INFO): Will output estimates for 1/2/3/6/15/25 blocks
 52020-12-19T20:14:12.104000Z TestFramework (INFO): Creating transactions and mining them with a block size that can't keep up
 62020-12-19T20:14:15.879000Z TestFramework (INFO): Creating transactions and mining them at a block size that is just big enough
 72020-12-19T20:14:21.104000Z TestFramework (ERROR): Assertion failed
 8Traceback (most recent call last):
 9  File "/home/sdaftuar/ccl-bitcoin/2020-12-latest-master/test/functional/test_framework/test_framework.py", line 126, in main
10    self.run_test()
11  File "test/functional/feature_fee_estimation.py", line 256, in run_test
12    check_estimates(self.nodes[1], self.fees_per_kb)
13  File "test/functional/feature_fee_estimation.py", line 140, in check_estimates
14    check_smart_estimates(node, fees_seen)
15  File "test/functional/feature_fee_estimation.py", line 130, in check_smart_estimates
16    % (feerate, last_feerate))
17AssertionError: Estimated fee (0.000565) larger than last fee (0.000479) for lower number of confirms

sdaftuar added the label Bug on Dec 19, 2020

maflcko commented at 6:09 am on December 20, 2020: member

Thanks for finding the random seed that triggers this reliably

maflcko added the label Tests on Dec 20, 2020

bitcoin deleted a comment on Dec 20, 2020

maflcko removed the label Tests on Dec 21, 2020

maflcko added the label TX fees and policy on Dec 21, 2020

maflcko commented at 8:33 am on December 21, 2020: member

With --tracerpc shows that this is really an issue with estimatesmartfee, not the test:

0...
1-4469-> estimatesmartfee [1]
2<-4469- [0.001963] {"feerate": "0.00067579", "blocks": 2}
3-4470-> estimatesmartfee [2]
4<-4470- [0.002432] {"feerate": "0.00067579", "blocks": 2}
5-4471-> estimatesmartfee [3]
6<-4471- [0.002226] {"feerate": "0.00047895", "blocks": 3}
7-4472-> estimatesmartfee [4]
8<-4472- [0.002212] {"feerate": "0.00056521", "blocks": 4}
9...

maflcko commented at 9:31 am on December 21, 2020: member

The incorrect return values (for target 3 and 4) are calculated in this line (short horizon): https://github.com/bitcoin/bitcoin/blob/e9efb64a0772777f2ce88c491101550b931a21d6/src/policy/fees.cpp#L736

the values are:

confTarget=3 successThreshold=0.85 estimate=47894.7 nBestSeenHeight=230
confTarget=4 successThreshold=0.85 estimate=56521 nBestSeenHeight=230

With corresponding debug-log line:

FeeEst: 3 > 85% decay 0.96200: feerate: 47894.7 from (40774.3 - 66417.1) 85.17% 84.9/(99.7 0 mem 0.0 out) Fail: (38832.7 - 40774.3) 74.91% 26.8/(35.7 0 mem 0.0 out)
FeeEst: 4 > 85% decay 0.96200: feerate: 56521 from (47201.4 - 66417.1) 85.47% 75.9/(88.8 0 mem 0.0 out) Fail: (38832.7 - 47201.4) 82.16% 38.3/(46.6 0 mem 0.0 out)

maflcko renamed this:
~~QA: Occasional failure in feature_fee_estimation.py~~
EstimateMedianVal returns higher fee for higher confTarget
on Dec 21, 2020

maflcko commented at 11:14 am on December 21, 2020: member

Changed title

maflcko commented at 3:08 pm on January 27, 2021: member

Related #11800?

practicalswift commented at 9:06 am on February 12, 2021: contributor

Unfortunately this larger than life^H^H^Hast fee issue in feature_fee_estimation.py keeps popping up.

From https://cirrus-ci.com/task/4756253495787520:

0AssertionError: Estimated fee (0.000400) larger than last fee (0.000338) for lower number of confirms

ajtowns commented at 1:16 pm on February 12, 2021: contributor

Fix in #21161 I think. I think what’s going on is that the bucket ranges weren’t consistent between confirmation targets, because newBucketRange = true is only set when the threshold is met, and buckets are extended otherwise. Not quite sure what the arrangement needs to be for that to result in a higher target giving a higher fee though.

morcos commented at 3:55 pm on February 12, 2021: member

Related #11800?

I think its an unrelated different bug

ismaelsadeeq commented at 1:24 pm on July 6, 2023: member

I try recreating this issue on master HEAD bc4f6b13feb29146b7e10e86f93dc7f6fb6937f2
with the same command test/functional/feature_fee_estimation.py --randomseed 108574360997204915

The tests passed.

maflcko commented at 1:37 pm on July 6, 2023: member

@ismaelsadeeq Can you try with the commit specified in OP as well? If it fails, can you try bisecting?

mzumsande commented at 8:37 pm on July 6, 2023: contributor

@ismaelsadeeq Can you try with the commit specified in OP as well? If it fails, can you try bisecting?

Maybe not necessary: There are newer crashes reported in #23165 (which seems to be the same problem). The seed 7288975409214300634 from https://api.cirrus-ci.com/v1/task/6110951150190592/logs/ci.log leads to a crash on current master c325f0fbae2d6472a78be733d499783f8b69d6b8 (the test passes with #21161 rebased on master).

glozow closed this on Nov 2, 2023

EstimateMedianVal returns higher fee for higher confTarget #20725