Closed vishnu2kmohan closed 1 year ago
Can we not modify the docker config to log all containers to journald? It seems like configuring every individual container leaves it open to potentially missing one somewhere.
Can we not modify the docker config to log all containers to journald? It seems like configuring every individual container leaves it open to potentially missing one somewhere.
I'm reluctant to set the log-driver globally at the Docker Engine level because all container logs (which can easily trigger rate limits and/or drop logs for very large multi-GPU and/or multi-node distributed training trials - esp. if debug mode is enabled) will be shipped to journald which has historically had trouble keeping up with massive log volumes.
I did a little digging on performance tuning systemd-journald, and apparently Pottering just doesn't care about logs. So yeah, it's not a great place for high volume, and can't easily be made to be. It might realistically not be the right place for anything that we actually want some near-guarantee of delivery. :)
This is fine.gif
Ensure that Docker containers relating to the HPE Machine Learning Development System log to
journald
so that their logs are automatically aggregated to the HPCM Admin node under/var/log/HOSTS/<HPCM_node_name>
.