IntersectMBO/cardano-ledger

Consistent test failure in cardano-ledger-alonzo-test

neilmayhew opened this issue · 4 comments

neilmayhew commented

The failure is in Shelley.Rules.ClassifyTraces, in Chain and Ledger traces cover the relevant cases. (This is a nightly-only test.)

The messages are similar to this:

Only 7.5% at least 5 epochs in a trace, 20% of the time, but expected 20.0%

The percentage varies, but the problem is always in the same test, with all versions of ghc.

The failure can be reproduced locally using

NIGHTLY=true cabal test cardano-ledger-alonzo-test --test-options='-p "/Chain and Ledger/"'
neilmayhew commented

The problem started with the index-state bump for Hackage that was in 0f28081157bef3ae5d7fcd742be13984bf749026 Make it build with ghc 9.10. Bisecting the index dates, I was able to isolate it to 2024-06-08 and 2024-06-09. The only changes to our deps were hashable and the associated bumps to aeson and strict.

neilmayhew commented

Downgrading hashable with a constraint fixes the problem on a branch starting from the ghc 9.10 commit 0f28081157bef3ae5d7fcd742be13984bf749026. However, the same constraint doesn't fix master.

neilmayhew commented

This fixes master:

constraints:
  , hashable == 1.4.4.0
allow-older:
  , aeson:hashable
  , strict:hashable

The only difference in the build plan is then the version of hashable (1.4.4.0 vs 1.4.7.0).

neilmayhew commented

hashable 1.4.5.0 causes the tests to fail. The changelog has two entries:

  • Drop support for GHCs prior 8.6.5
  • Use xxhash for hashing bytestrings and bytearrays