Duplicated timeseries in CollectorRegistry with Multiprocess Gunicorn

I know this is a subject that comes up somewhat frequently but for the love of me I can't figure out what I'm doing wrong.

1. I have a service in Amazon ECS thats running a single task with multiple workers (actually the problem happens in my other service that just has one worker also).

2. I've created the directory and set the PROMETHEUS_MULTIPROC_DIR in the Dockerfile:

RUN mkdir -p /tmp/prom-metrics
ENV PROMETHEUS_MULTIPROC_DIR /tmp/prom-metrics

3. I'm using the sample code in the README to create the registry in the /metrics request and return it:

registry = CollectorRegistry()
if getenv('PROMETHEUS_MULTIPROC_DIR'):
  multiprocess.MultiProcessCollector(registry)
data = generate_latest(registry)
status = '200 OK'
response_headers = [
    ('Content-type', CONTENT_TYPE_LATEST),
    ('Content-Length', str(len(data))),
]
return Response(data, status, response_headers)

4. I've created the gunicorn.conf.py file with the sample from the README and passed it into my gunicorn startup script via -c:

from prometheus_client import multiprocess

def child_exit(server, worker):
    multiprocess.mark_process_dead(worker.pid)

In my two services, gunicorn starts them as follows:

# app 1 with workers
gunicorn -c /app/utils/gunicorn.conf.py -b :5000 -t 3600 --keep-alive 60 --threads 8 --workers 3 app:app

# app 2 without workers
gunicorn -c /app/utils/gunicorn.conf.py -b :5000 -t 3600 --keep-alive 60 --threads 8 app:app

The service boots successfully and accepts some metrics which are definitely collected in multiprocess mode, seeing as the HELP line simply displays Multiprocess metric.

This works for a few calls but eventually I get the dreaded Duplicated timeseries in CollectorRegistry error and no additional metrics are populated.

What might I be doing wrong?

Start with the README's multiprocess metrics example and the referenced gunicorn.conf.py configuration, then reproduce the failure using the Dockerfile settings and Gunicorn commands shown. Done means identifying the cause of the duplicated timeseries error and documenting or correcting the configuration so /metrics continues returning metrics.