`DataLoaders(..., parallel=true)` hanging

In the following MWE I successively create an out-of-memory data source of 20 MNIST images using FileDataset. I can the wrap the source as MLUtils.DataLoader with the default parallel=false option and collect the result. However, if I specify parallel=true then the collect hangs.

Pkg.activate("data", shared=true)
import MLDatasets: MNIST
using MLDatasets
using ScientificTypes
using MLUtils
using FileIO

ENV["DATADEPS_ALWAYS_ACCEPT"] = true
images, labels = MNIST.(split=:train)[:];

N = 20
images = coerce(images, GrayImage)[1:N];

# save some MNIST images as tiff files:
const dir = tempname()
for i  in eachindex(images)
    filename = joinpath(dir, "$i.tiff")
    FileIO.save(filename, images[i])
end

# create out-of-memory image source:
X = MLDatasets.FileDataset(dir)

sequential = DataLoader(X, batchsize=2, collate=true)
collect(sequential) # executes as expected

parallel = DataLoader(X, batchsize=2, collate=true, parallel=true);
collect(parallel); # hangs

Here's my setup:

(@data) pkg> status
Status `~/.julia/environments/data/Project.toml`
  [5789e2e9] FileIO v1.16.0
  [82e4d734] ImageIO v0.6.6
  [eb30cadb] MLDatasets v0.7.6
  [f1d291b0] MLUtils v0.3.1
  [321657f4] ScientificTypes v3.0.2

julia> versioninfo()
Julia Version 1.8.3
Commit 0434deb161e (2022-11-14 20:14 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin21.4.0)
  CPU: 12 × Intel(R) Core(TM) i7-8850H CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-13.0.1 (ORCJIT, skylake)
  Threads: 5 on 12 virtual cores
Environment:
  JULIA_LTS_PATH = /Applications/Julia-1.6.app/Contents/Resources/julia/bin/julia
  JULIA_PATH = /Applications/Julia-1.8.app/Contents/Resources/julia/bin/julia
  JULIA_EGLOT_PATH = /Applications/Julia-1.6.app/Contents/Resources/julia/bin/julia
  JULIA_NUM_THREADS = 5
  JULIA_NIGHTLY_PATH = /Applications/Julia-1.8.app/Contents/Resources/julia/bin/julia

JuliaML / MLUtils.jl

`DataLoaders(..., parallel=true)` hanging #132