Configured to not copy downloaded models from the host machine to the container because then on every build Docker copies them to its "build context" (+6 GB), to the container (another +6 GB), and lastly to the intermediate snapshot (yet another +6 GB), so I would run out of disk space that way :( I found it easier to just redownload models from S3 on every build.
Runs a single Gunicorn worker (it is expected for the scaling to be done by starting more Docker containers, not more Gunicorn workers on a single container)
(Not sure how this might be useful since you're using Dokku, but I created
Dockerfile
for myself so I thought I might as well submit it somewhere.)Dockerfile
that:requirements.txt
globally (because there's not much point in using Virtualenv in an already isolated container)nytlabels
userdownload_models.py
script (so depends on https://github.com/mitmedialab/predict-news-labels/pull/4)Simple usage example is at the top of
Dockerfile
.