Open hannes-ucsc opened 1 year ago
We should do the mirroring inside AWS, obviously, but that would require an EC2 instance. We have an EC2 instance in each AWS account but it is provisioned by the gitlab
component which depends on the shared
component. So a classic chicken-vs-egg problem. Additionally, the GitLab instance does not have enough permission to deploy the gitlab
or shared
components, either, for good reason.
We could start allowing operators with slow uplinks to maintain their own bastion EC2 instance but then that instance would reside inside the authorization boundary and would have to be subject to strict compliance requirements.
There really is no good answer. My uplink was fast enough to do the initial mirroring of all images in all deployments. It required a bit of planning ahead about timing.
My recommendation is to
1) ensure that the STS token is fresh before deploying shared (#5302 will add documentation to recommend that as a general best practice for operators)
2) Have a separate Azul checkout to run operator tasks from so that the operator can continue to work on other things
Alternatively, the operator could drive to campus or WRP, hop on eduroam and perform the task from there.
Spike to estimate monthly cost for Comcast plan upgrades & test image mirroring to ECR from Eduroam at WRP.
… now that it includes the mirroring of images.
The GitLab image is huge (> 1GiB) so mirroring it to ECR will take an hour or more on a consumer internet connection (we all work from home).