Address ephemeral memory eviction for BinderHubs on JS2 #6923
Description
See #6498 for more context.
On JS2, user nodes are long lived because they can't scale down to zero. This means that the normal cycle of re-creating nodes doesn't occur, and as such, we do not have good control on unbounded growth of the build cache for Docker. We should tackle this so that JS2 hubs do not experience periodic outages.
Definition of done
- JS2 deployments like Pythia can run hubs for a long time without pods crashing
- The fix is well documented