Completely drop the DEFAULT_SOLVER logic (#250)

- Dropped get_default_solver and set_default_solver
- Reorganized tests to not depend on get_default_solver (the giant at-testset for loop can be changed).
- Only API change is that if you didn't define a problem with a solver/model then `solve!(p)` throws an argument
  (e.g., `solve!(p, solver)` is preferred now. NOTE: The existing `solve!(p, solver)` just assigns the problem model anyways
  with `problem.model = MathProgBase.ConicModel(s)`

src/problems.jl

@@ -25,11 +25,12 @@ mutable struct Problem
     constraints::Array{Constraint}
     status::Symbol
     optval::Float64OrNothing
-    model::MathProgBase.AbstractConicModel
+    model::Union{MathProgBase.AbstractConicModel, Nothing}
     solution::Solution
 
     function Problem(head::Symbol, objective::AbstractExpr,
-                     model::MathProgBase.AbstractConicModel, constraints::Array=Constraint[])
+                     model::Union{MathProgBase.AbstractConicModel, Nothing}=nothing,
+                     constraints::Array=Constraint[])
         if sign(objective)== Convex.ComplexSign()
             error("Objective can not be a complex expression")
         else
@@ -40,8 +41,9 @@ end
 
 # constructor if model is not specified
 function Problem(head::Symbol, objective::AbstractExpr, constraints::Array=Constraint[],
-                 solver::MathProgBase.AbstractMathProgSolver=DEFAULT_SOLVER)
-    Problem(head, objective, MathProgBase.ConicModel(solver), constraints)
+                 solver::Union{MathProgBase.AbstractMathProgSolver, Nothing}=nothing)
+    model = solver !== nothing ? MathProgBase.ConicModel(solver) : solver
+    Problem(head, objective, model, constraints)
 end
 
 # If the problem constructed is of the form Ax=b where A is m x n

src/solution.jl

@@ -18,6 +18,14 @@ function solve!(problem::Problem;
                 check_vexity=true,
                 verbose=true)
 
+    if problem.model === nothing
+        throw(ArgumentError(
+            "The provided problem hasn't been initialized with a conic model.
+            You can resolve this by passing in `AbstractMathProgSolver` such as:
+            solve!(problem, ECOSSolver())"
+        ))
+    end
+
     if check_vexity
         vex = vexity(problem)
     end

src/solver_info.jl

@@ -1,46 +1,7 @@
 using Pkg
 import MathProgBase
 export can_solve_mip, can_solve_socp, can_solve_sdp, can_solve_exp
-export set_default_solver, get_default_solver, isinstalled
-
-function set_default_solver(solver::MathProgBase.AbstractMathProgSolver)
-    Base.depwarn(
-        "Default solvers are deprecated and will not be provided in a future release.
-        You can still manually load and use solvers such as ECOS by running:
-        using ECOS
-        p = minimize(...);
-        solve!(p, ECOSSolver())",
-        :set_default_solver
-    )
-
-    global DEFAULT_SOLVER
-    DEFAULT_SOLVER = solver
-end
-
-function get_default_solver()
-    Base.depwarn(
-        "Default solvers are deprecated and will not be provided in a future release.
-        You can still manually load and use installed solvers such as ECOS by running:
-        using ECOS
-        p = minimize(...);
-        solve!(p, ECOSSolver())",
-        :get_default_solver
-    )
-
-    if DEFAULT_SOLVER == nothing
-        error("The default solver is set to `nothing`
-              You must have at least one solver installed to use Convex.
-              You can install a solver such as SCS by running:
-              Pkg.add(\"SCS\").
-              You will have to restart Julia after that.")
-    end
-
-    return DEFAULT_SOLVER
-end
-
-# TODO: I have not listed solvers such as CPLEX etc because I have not tested Convex with them
-solvers = [("ECOS", "ECOSSolver"), ("SCS", "SCSSolver"), ("Gurobi", "GurobiSolver"), ("Mosek", "MosekSolver"),
-           ("GLPKMathProgInterface", "GLPKSolverMIP")]
+export isinstalled
 
 function isinstalled(pkg)
     for path in Base.DEPOT_PATH
@@ -52,24 +13,10 @@ function isinstalled(pkg)
     return false
 end
 
-for (dir, solver) in solvers
-    if isinstalled(dir) && DEFAULT_SOLVER == nothing
-        eval(Meta.parse("using $dir"))
-        eval(Meta.parse("set_default_solver($solver())"))
-    end
-end
-
+# TODO: I have not listed solvers such as CPLEX etc because I have not tested Convex with them
+solvers = [("ECOS", "ECOSSolver"), ("SCS", "SCSSolver"), ("Gurobi", "GurobiSolver"), ("Mosek", "MosekSolver"),
+           ("GLPKMathProgInterface", "GLPKSolverMIP")]
 
-if get_default_solver() == nothing
-    packages = join([dir for (dir, solver) in solvers], " | ")
-    @warn "***********************************************************************************************
-          You don't have any of
-          "*packages*" installed.
-          You must have at least one of these solvers. You can install a solver such as SCS by running:
-          Pkg.add(\"SCS\")
-          You will have to restart Julia after that.
-          ***********************************************************************************************"
-end
 
 function can_solve_mip(solver)
     name = typeof(solver).name.name

test/runtests.jl

@@ -1,10 +1,20 @@
 using Convex
+using Convex: DotMultiplyAtom
 using Test
 using ECOS
 using SCS
 using GLPKMathProgInterface
 using Random
 
+import LinearAlgebra.eigen
+import LinearAlgebra.I
+import LinearAlgebra.opnorm
+import Random.shuffle
+import Statistics.mean
+
+TOL = 1e-3
+eye(n) = Matrix(1.0I, n, n)
+
 # Seed random number stream to improve test reliability
 Random.seed!(2)
 
@@ -24,9 +34,13 @@ if isinstalled("Mosek")
     push!(solvers, MosekSolver(LOG=0))
 end
 
-for solver in solvers
-    println("Running tests with $(solver):")
-    set_default_solver(solver)
-    println(get_default_solver())
-    include("runtests_single_solver.jl")
+@testset "Convex" begin
+    include("test_utilities.jl")
+    include("test_affine.jl")
+    include("test_lp.jl")
+    include("test_socp.jl")
+    include("test_sdp.jl")
+    include("test_exp.jl")
+    include("test_sdp_and_exp.jl")
+    include("test_mip.jl")
 end