Custom settings not available in node pools

[ankon]

I'm seeing this error for a template rendered with 0.9.4-rc.3, and a node pool with a change that requires the custom settings:

Error: Failed to initialize cluster driver: failed to load node pool #0: failed to render worker cloud config: template: userdata/cloud-config-worker:389:65: executing "userdata/cloud-config-worker" at <.CustomSettings.work...>: can't evaluate field CustomSettings in type *config.ComputedConfig

Change in the user data:

    - name: swap.service
      command: start
      content: |
        [Unit]
        Description=Turn on swap
        [Service]
        Type=oneshot
        Environment="SWAP_PATH=/var/vm" "SWAP_FILE=swapfile1"
        ExecStartPre=-/usr/bin/rm -rf ${SWAP_PATH}
        ExecStartPre=/usr/bin/mkdir -p ${SWAP_PATH}
        ExecStartPre=/usr/bin/touch ${SWAP_PATH}/${SWAP_FILE}
        ExecStartPre=/bin/bash -c "fallocate -l {{.CustomSettings.workerSwapSize}} ${SWAP_PATH}/${SWAP_FILE}"
        ExecStartPre=/usr/bin/chmod 600 ${SWAP_PATH}/${SWAP_FILE}
        ExecStartPre=/usr/sbin/mkswap ${SWAP_PATH}/${SWAP_FILE}
        ExecStartPre=/usr/sbin/sysctl vm.swappiness=10
        ExecStart=/sbin/swapon ${SWAP_PATH}/${SWAP_FILE}
        ExecStop=/sbin/swapoff ${SWAP_PATH}/${SWAP_FILE}
        ExecStopPost=-/usr/bin/rm -rf ${SWAP_PATH}
        RemainAfterExit=true
        [Install]
        WantedBy=multi-user.target

[redbaron]

not to say that .CustomSettings shouldn't be fixed, but why do you need swap? there is even a flag to enable startup with swap, but description suggests, that starting from version 1.6 kubelet going to refuse to start if it detects that swap is enabled:

 --experimental-fail-swap-on                               Makes the Kubelet fail to start if swap is enabled on the node. This is a temporary option to maintain legacy behavior, failing due to swap enabled will happen by default in v1.6.

[ankon]

Interesting. Didn't know that, so that's certainly to be considered.

Why do I need swap: I'm not 100% sure I do, but given the apps I'm running in the cluster I know that they might have quite a bit of "unused" memory (Java ...).

[redbaron]

resources.request.memory is used to schedule, but doesn't limit resource usage. so it can be set smaller than actual usage, which itself is limited by resources.limit.memory.

Actually Java memory is one of the most predictable. from previous experience it was safe to set container limit to Heap size (defined by -Xmx) + 256 MB