Each varnishkafka is currently set to write stats periodically to a file like /var/cache/varnishkafka/$instance_name.stats.json. We have a root cron that executes every minute logster, that reads the json and pushes the latest metrics to graphite.
Rather than adding HTTP support to varnishkafka to poll the same set of metrics, we could simply write a python script that reads the json file, picks the last datapoints and then exposes the metrics to Prometheus when polled.