SCIP errors when setting start values from previous solutions in iterative solves

[resmaeilbeigi]

When solving a model iteratively (only changing the objective function and adding a constraint based on the previous objective value), SCIP sometimes fails when start values are set from the previous solution.

I have observed two distinct errors in this context:

1. SCIP_INVALIDCALL in SCIPaddSolFree()
   The example below reproduces this issue. It occured after 12 iterations.

2. SCIPsolve(o) yielded SCIP code SCIP_ERROR
   This happened in another test environment when setting start values before each iteration.
   The error followed a series of SCIP internal errors:

presolved problem has 2360 variables (1355 bin, 0 int, 505 impl, 500 cont) and 2069 constraints
    230 constraints of type <varbound>
     14 constraints of type <knapsack>
   1139 constraints of type <setppc>
    431 constraints of type <linear>
    243 constraints of type <logicor>
     12 constraints of type <bounddisjunction>
Presolving Time: 3.34
transformed 3/4 original solutions to the transformed problem space

[lp.c:11621] ERROR: cannot solve LP when loose variable with infinite best bound is present
[lp.c:12068] ERROR: Error <0> in function call
[lp.c:12257] ERROR: Error <0> in function call
[lp.c:12512] ERROR: Error <0> in function call
[solve.c:1512] ERROR: Error <0> in function call
[solve.c:3096] ERROR: Error <0> in function call
[solve.c:4015] ERROR: Error <0> in function call
[solve.c:4313] ERROR: Error <0> in function call
[solve.c:5106] ERROR: Error <0> in function call
[scip_solve.c:2661] ERROR: Error <0> in function call

Both errors happen due to setting start values as in the example below.
Example that reproduces the first error:

using JuMP, SCIP

function main()
  m, n = 100, 150
  model = Model(SCIP.Optimizer)
  set_silent(model)
  @variable(model, x[1:m, 1:n], Bin)
  @variable(model, y[1:max(m, n)], Bin)
  @constraints(
    model,
    begin
      [i in 1:n], sum(x[:, i]) <= y[i]
      [j in 1:m], sum(mod(i, 2) * x[j, i] for i in 1:n) >= sum(y[1:j])
    end
  )
  for k in 1:min(m, n)
    indices = [(i, j) for i in 1:m for j in 1:n if (i + j) % k == 0]
    @constraint(model, sum(x[i, j] for (i, j) in indices) >= 1)
  end
  varValues = nothing
  for k in 1:20
    @info "k = $k"
    obj = @expression(
      model,
      sum(mod(m + n + k, i) * y[i] for i in 1:max(m, n)) +
      sum(mod(m + n + k, i + j) * x[i, j] for i in 1:m, j in 1:n)
    )
    if varValues !== nothing
      for (varIdx, val) in varValues
        set_start_value(VariableRef(model, varIdx), val)
      end
    end
    @objective(model, Min, obj)
    optimize!(model)
    if primal_status(model) == FEASIBLE_POINT
      varValues = Dict{MOI.VariableIndex, Float64}(
        v.index => value(v) for v in all_variables(model)
      )
      obj_val = objective_value(model)
      @constraint(model, obj <= 1.01 * obj_val)
    end
  end
end

main()

Running this code results in the following:

[ Info: k = 1
[ Info: k = 2
[ Info: k = 3
[ Info: k = 4
[ Info: k = 5
[ Info: k = 6
[ Info: k = 7
[ Info: k = 8
[ Info: k = 9
[ Info: k = 10
[ Info: k = 11
[ Info: k = 12
[primal.c:901] ERROR: Cannot add partial solution to storage: limit reached.
[primal.c:1397] ERROR: Error <-8> in function call
[scip_sol.c:2864] ERROR: Error <-8> in function call
ERROR: LoadError: SCIPaddSolFree(o, sol__, stored_) yielded SCIP code SCIP_INVALIDCALL

Environment:

Julia version: 1.11.7
JuMP version: 1.29.1
SCIP.jl version: 0.12.7
SCIP_PaPILO_jll version: 900.200.200+0
Operating system: Ubuntu 20.04.6

[odow]

It looks like we're adding starts each time, instead of replacing them. I'll see if I can fix

[odow]

Oh man. I underestimated how complicated the SCIP API is for storing partial solutions.

[matbesancon]

We can check together after JuMP-dev