Segmentation fault when writing non-linear model to LP

[jonathanberthias]

Describe the bug

I have a test that started segfaulting with v6 (SCIP 10), but the issue only occurs when running with pytest. The test writes the problem in LP format.

To Reproduce

I had to spend a bit of time to come up with the smallest example, so here we go.
Create a file repro_test.py with these contents:

from pyscipopt import Model


def test_abs():
    model = Model()
    x = model.addVar()
    model.addCons(abs(x) >= 1)
    model.setObjective(x, "minimize")
    model.writeProblem("abs_example.lp")  # segfaults


if __name__ == "__main__":
    test_abs()

Then run pytest on it (here using uvx to have a fresh environment):

uvx --with pyscipopt==6.0.0 --with pytest pytest repro_test.py

This causes a segmentation fault:

repro_test.py Fatal Python error: Segmentation fault

Current thread 0x000079f2a721f740 [python] (most recent call first):
  File "~/dev/cvxpy-translation/repro_test.py", line 10 in test_abs

With a debug build of SCIP (scipopt/scip@9e13056), I see:

python: ~/dev/scip/src/scip/expr.c:4267: SCIPexprAreQuadraticExprsVariables: Assertion `expr->quaddata != NULL' failed.
Fatal Python error: Aborted

Expected behavior

The code should work without segfaulting. It does work when any of these change:

1. Without calling writeProblem
2. Without abs
3. Running python instead of pytest on the file
4. Using PySCIPOpt 5.7.1

This is probably caused by pytest modifying the environment in some way. The version of pytest I use does not change the outcome though, only the version of SCIP seems to matter.

System

• OS: Ubuntu
• Version: 6.0.0
• SCIP version: 10.0.0
• How did you install pyscipopt? PyPI and from source

[Joao-Dionisio]

Ah this is interesting @jonathanberthias

I will try to take a look, but I wonder if this can be on the SCIP side

[jonathanberthias]

I will try to take a look, but I wonder if this can be on the SCIP side

Yes I think you're right, I initially thought it was on the Python side due to the difference between Pytest and standard Python but I just realized I was mistaken and I get the same error with Python.
For context, Pytest uses the faulthandler module to make segfaults loud, but without it the error is silent. And even though it segfaults, a .lp file gets created, but its contents are wrong (the abs constraint is missing).

So the reproducer now is:

import faulthandler
from pyscipopt import Model

faulthandler.enable()

model = Model()
model.addCons(abs(model.addVar()) >= 1)
model.writeProblem("abs_example.lp")

This is the C stack dump on Python 3.14:

  Binary file "pyscipopt.libs/libscip-f6cc0c84.so.10.0", at SCIPwriteLp+0x332b [0x7b714949bbcb]
  Binary file "pyscipopt.libs/libscip-f6cc0c84.so.10.0", at +0x4948d1 [0x7b71494948d1]
  Binary file "pyscipopt.libs/libscip-f6cc0c84.so.10.0", at +0x67c859 [0x7b714967c859]
  Binary file "pyscipopt.libs/libscip-f6cc0c84.so.10.0", at +0x728c30 [0x7b7149728c30]
  Binary file "pyscipopt.libs/libscip-f6cc0c84.so.10.0", at +0x7291f0 [0x7b71497291f0]
  Binary file "pyscipopt.libs/libscip-f6cc0c84.so.10.0", at SCIPwriteOrigProblem+0x21 [0x7b7149729451]
  Binary file "pyscipopt/scip.cpython-314-x86_64-linux-gnu.so", at +0x29c126 [0x7b714aa9c126]

[Joao-Dionisio]

Oh I didn't know faulthandler, will be pretty useful! Usually, I use gdb, which can give you the backtrace down to the C-SCIP level. I'll try to take a look, thank you @jonathanberthias !

[Joao-Dionisio]

Ah... I should have paid more attention. You are trying to write a problem with an absolute term with the lp file format, which it can only write quadratically constrained problems. I always use .cip, but that's a SCIP-specific format.

[jonathanberthias]

Ah... I should have paid more attention. You are trying to write a problem with an absolute term with the lp file format, which it can only write quadratically constrained problems. I always use .cip, but that's a SCIP-specific format.

Indeed I was abusing this a little, but at least it didn't error in v5.7.1 (SCIP 9), it simply commented the constraint:

\ SCIP STATISTICS
\   Problem name     : model
\   Variables        : 1 (0 binary, 0 integer, 0 implicit integer, 1 continuous)
\   Constraints      : 1
Minimize
 Obj: 0 x1
Subject to
\   [nonlinear] <c1>: abs((<x1>)) >= 1;
Bounds
End

For my usage, I don't need to read the generated file back, so having this was sufficient. And in any case, a segmentation fault is always a problem.

I guess I'll move to CIP format, though I don't find it as readable as LP.

[Joao-Dionisio]

@DominikKamp do you know if something changed in the LP reader?

[DominikKamp]

Multiple things changed, but I can not reproduce this with SCIP neither for scipopt/scip@9e13056 nor for latest v100-bugfix.

[jonathanberthias]

but I can not reproduce this with SCIP neither for scipopt/scip@9e13056 nor for latest v100-bugfix.

Hi @DominikKamp, did you try to reproduce with C-SCIP or through PySCIPOpt?

[DominikKamp]

C-SCIP where writing the commented abs function in the LP format works for me, so PySCIPOpt might construct the expressions in a different way to trigger this issue (maybe it is relevant that I only get \ [nonlinear] <c1>: abs(<x1>) >= 1; without duplicate parentheses).