diff --git a/reproduce_container_bug.md b/reproduce_container_bug.md
new file mode 100644
index 0000000..2707119
--- /dev/null
+++ b/reproduce_container_bug.md
@@ -0,0 +1,30 @@
+Create the container on the login node:
+
+```bash
+export WS_DIR=$(ws_find workspace_dir) # adjust this
+cd $WS_DIR
+wget https://fex.hlrs.de/fop/FYaJqyzw/ray.tar # download the container archive
+export CONTAINER_NAME=ray
+export CONTAINER_TAG=latest
+export UDOCKER_DIR="$WS_DIR/.udocker/" # to store the image layers
+udocker images -l # this will create a repo the first time you use it
+udocker rmi $CONTAINER_NAME:$CONTAINER_TAG # results in error since the image does not exist
+udocker load -i $WS_DIR/$CONTAINER_NAME.tar $CONTAINER_NAME
+rm /$WS_DIR/$CONTAINER_NAME.tar # you no longer need the tar archive
+```
+
+Allocate a CPU node:
+
+```bash
+module load bigdata/udocker/1.3.4
+export WS_DIR=$(ws_find benchmarks)
+udocker run --volume $WS_DIR:/workspace --volume /run/user/$PBS_JOBID/tmp:/tmp $CONTAINER_NAME
+```
+
+You should see a Python shell.
+
+```python
+import ray
+# ray.init(num_cpus=4) # Works with a small number of CPUs
+ray.init() # But, it can't use all the available CPUs
+```