Segfault on malformed Convex.jl problem

[ericphanson]

With Convex v0.12.0,

using Convex, ECOS, Test
solver = ECOSSolver()
TOL = 1e-4
x = Variable()
y = Variable()
fix!(x, 1.0)
prob = minimize(y*x, [y >= x, x >= 0.5])
solve!(prob, solver)
@test prob.optval ≈ 1.0 atol = TOL

fix!(x, 2.0)
solve!(prob, solver)
@test prob.optval ≈ 4.0 atol = TOL

free!(x)
fix!(y, 1.0)
solve!(prob, solver)
@test prob.optval ≈ 0.5 atol = TOL

results in a segfault. This is partly Convex.jl's fault, because it creates a problem that doesn't make sense (see jump-dev/Convex.jl#299), but I don't think ECOS should segfault. SCS for example returns NaN.

I'm actually not 100% sure exactly how it's a malformed problem (the why is because x is set to :fixed but has AffineVexity() instead of ConstVexity()). For example,

c, A, b, cones, var_to_ranges, vartypes, conic_constraints = Convex.conic_problem(prob)

gives the abstract problem setup that Convex will then pass to MathProgBase, and I can't see what's wrong with it from there right away.

[odow]

Closing because I can no longer reproduce this:

julia> using Convex, ECOS, Test

julia> solver = ECOS.Optimizer
ECOS.Optimizer

julia> TOL = 1e-4
0.0001

julia> x = Variable()
Variable
size: (1, 1)
sign: real
vexity: affine
id: 181…067

julia> y = Variable()
Variable
size: (1, 1)
sign: real
vexity: affine
id: 120…994

julia> fix!(x, 1.0)
Variable
size: (1, 1)
sign: real
vexity: constant
id: 181…067
value: 1.0

julia> prob = minimize(y*x, [y >= x, x >= 0.5])
minimize
└─ * (affine; real)
   ├─ real variable (id: 120…994)
   └─ real variable (fixed) (id: 181…067)
subject to
├─ >= constraint (affine)
│  ├─ real variable (id: 120…994)
│  └─ real variable (fixed) (id: 181…067)
└─ >= constraint (constant)
   ├─ real variable (fixed) (id: 181…067)
   └─ 0.5

status: `solve!` not called yet

julia> solve!(prob, solver)

ECOS 2.0.5 - (C) embotech GmbH, Zurich Switzerland, 2012-15. Web: www.embotech.com/ECOS

It     pcost       dcost      gap   pres   dres    k/t    mu     step   sigma     IR    |   BT
 0  +1.000e+00  +1.500e+00  +3e+00  3e-01  2e-01  1e+00  1e+00    ---    ---    1  1  - |  -  - 
 1  +9.503e-01  +9.886e-01  +2e-01  2e-02  1e-02  7e-02  1e-01  0.9402  4e-02   1  0  0 |  0  0
 2  +9.994e-01  +9.999e-01  +3e-03  2e-04  1e-04  7e-04  2e-03  0.9884  1e-04   1  0  0 |  0  0
 3  +1.000e+00  +1.000e+00  +3e-05  2e-06  1e-06  8e-06  2e-05  0.9890  1e-04   1  0  0 |  0  0
 4  +1.000e+00  +1.000e+00  +3e-07  2e-08  2e-08  9e-08  2e-07  0.9890  1e-04   1  0  0 |  0  0
 5  +1.000e+00  +1.000e+00  +4e-09  3e-10  2e-10  1e-09  2e-09  0.9890  1e-04   1  0  0 |  0  0

OPTIMAL (within feastol=2.6e-10, reltol=3.6e-09, abstol=3.6e-09).
Runtime: 0.000207 seconds.


julia> @test prob.optval ≈ 1.0 atol = TOL
Test Passed

julia> fix!(x, 2.0)
Variable
size: (1, 1)
sign: real
vexity: constant
id: 181…067
value: 2.0

julia> solve!(prob, solver)

ECOS 2.0.5 - (C) embotech GmbH, Zurich Switzerland, 2012-15. Web: www.embotech.com/ECOS

It     pcost       dcost      gap   pres   dres    k/t    mu     step   sigma     IR    |   BT
 0  +4.000e+00  +5.094e+00  +5e+00  1e-01  2e-01  1e+00  2e+00    ---    ---    1  1  - |  -  - 
 1  +3.955e+00  +4.080e+00  +5e-02  4e-03  6e-03  1e-01  1e-01  0.9725  3e-02   1  0  0 |  0  0
 2  +4.000e+00  +4.001e+00  +6e-04  5e-05  6e-05  1e-03  1e-03  0.9890  1e-04   1  0  0 |  0  0
 3  +4.000e+00  +4.000e+00  +7e-06  5e-07  7e-07  1e-05  1e-05  0.9890  1e-04   1  0  0 |  0  0
 4  +4.000e+00  +4.000e+00  +8e-08  6e-09  8e-09  2e-07  1e-07  0.9890  1e-04   1  0  0 |  0  0
 5  +4.000e+00  +4.000e+00  +9e-10  7e-11  9e-11  2e-09  2e-09  0.9890  1e-04   1  0  0 |  0  0

OPTIMAL (within feastol=8.7e-11, reltol=2.3e-10, abstol=9.0e-10).
Runtime: 0.000064 seconds.


julia> @test prob.optval ≈ 4.0 atol = TOL
Test Passed

julia> free!(x)
Variable
size: (1, 1)
sign: real
vexity: affine
id: 181…067
value: 2.0

julia> fix!(y, 1.0)
Variable
size: (1, 1)
sign: real
vexity: constant
id: 120…994
value: 1.0

julia> solve!(prob, solver)

ECOS 2.0.5 - (C) embotech GmbH, Zurich Switzerland, 2012-15. Web: www.embotech.com/ECOS

It     pcost       dcost      gap   pres   dres    k/t    mu     step   sigma     IR    |   BT
 0  +7.500e-01  -3.500e-08  +8e-01  6e-03  1e-08  1e+00  6e-01    ---    ---    1  1  - |  -  - 
 1  +6.151e-01  +4.951e-01  +1e-01  9e-04  7e-09  2e-01  9e-02  0.8529  1e-02   1  0  0 |  0  0
 2  +5.016e-01  +4.982e-01  +3e-03  2e-05  4e-09  3e-03  2e-03  0.9890  1e-02   1  0  0 |  0  0
 3  +5.000e-01  +5.000e-01  +4e-05  2e-07  9e-11  3e-05  2e-05  0.9890  1e-04   1  0  0 |  0  0
 4  +5.000e-01  +5.000e-01  +4e-07  3e-09  2e-12  3e-07  3e-07  0.9890  1e-04   1  0  0 |  0  0
 5  +5.000e-01  +5.000e-01  +5e-09  3e-11  2e-14  4e-09  3e-09  0.9890  1e-04   1  0  0 |  0  0

OPTIMAL (within feastol=2.9e-11, reltol=9.5e-09, abstol=4.8e-09).
Runtime: 0.000071 seconds.


julia> @test prob.optval ≈ 0.5 atol = TOL
Test Passed

(convex) pkg> st
      Status `/private/tmp/convex/Project.toml`
  [f65535da] Convex v0.14.18
  [e2685f51] ECOS v0.13.0
  [8dfed614] Test

[ericphanson]

To reproduce, one probably needs Convex v0.12.0. With jump-dev/Convex.jl#299 I fixed the bug that caused us to emit malformed problems. But maybe it was just a stackoverflow or something like that, which would be undestandable and seems to result in segfaults on MacOS (which I was probably using).

[odow]

That would mean using an older version of ECOS as well and I just re-wrote it :). We can dig deeper if someone ever triggers this again.

[ericphanson]

Ok, sounds good 🙂