it seems that scaling down addons can fail at times. this can happen
if the kernel is busy swapping things etc. so, we do not block the
platform startup for this operation.
Nov 13 05:20:00 localhost dockerd[28831]: time="2018-11-13T05:20:00.365056059Z" level=error msg="Handler for POST /v1.37/containers/mongodb/update returned error: Cannot update container 6532d4a923ce9f10303f2e2aa7f03c35383864f44f3db6abd5c58da3c1a7702f: docker-runc did not terminate sucessfully: failed to write 419430400 to memory.memsw.limit_in_bytes: write /sys/fs/cgroup/memory/docker/6532d4a923ce9f10303f2e2aa7f03c35383864f44f3db6abd5c58da3c1a7702f/memory.memsw.limit_in_bytes: device or resource busy\n: unknown"
Nov 13 05:20:00 localhost node[5486]: box:shell updatemongodb (stderr): Error response from daemon: Cannot update container 6532d4a923ce9f10303f2e2aa7f03c35383864f44f3db6abd5c58da3c1a7702f: docker-runc did not terminate sucessfully: failed to write 419430400 to memory.memsw.limit_in_bytes: write /sys/fs/cgroup/memory/docker/6532d4a923ce9f10303f2e2aa7f03c35383864f44f3db6abd5c58da3c1a7702f/memory.memsw.limit_in_bytes: device or resource busy
Moved the graphite logic to new graphs.js
The settings code now does change notification itself. Over time,
it makes sense to just having settings code do this for everything
and not have this change listener logic. This lets us:
* Maybe the settings can only return based on final handler result
* All dependant modules otherwise have to "init"ed to listen on startup
* Easier to test those handlers without having to actually change the
setting (since they will now be in "exports" naturally)
Also, maybe someday with this abstraction we can allow apps to have their
own isolated databases etc
