I seem to have obtained possible evidence of a memory leak. I'm relying on results reported by our diagnostic tools, and of course one caveat is the question of whether the tool is correct.

Here's the test script:

using Grid
A = rand(1000,1000)
x = rand(2:999, 10^6, 2) + rand(10^6, 2) - 0.5
Ai = InterpGrid(A, BCnil, InterpLinear)

function testgrid(Ai, x)
    y = zero(eltype(Ai))
    for i = 1:size(x,1)
        y += Ai[x[i,1],x[i,2]]
    end
    y
end

testgrid(Ai, x[1:5,:])
@time testgrid(Ai, x)

Grid has always been written with the intention of being allocation-free (during interpolation; there is allocation during initialization), but until @loladiro's tuple work returning multiple arguments from functions like interp_coords_1d led to "unnecessary" allocations. I think #4042 changed that; certainly, it led to a 2x speedup in the performance of this test code.

However, I'm still puzzled by the result of the output of @time:

julia> include("testgrid.jl")
elapsed time: 0.35688972 seconds (48006676 bytes allocated)
500208.87506182125

There's a pretty huge amount of memory allocated here, suggesting that it gets allocated on each iteration. Yet the profiler suggests there are no calls to jl_gc_collect. My concern is that the observed speedup is largely due to the absence of gc runs which, from the standpoint of memory consumption, should have been happening.