Intermittent bootstrap failure on AWS Nitro instances

Summary

On AWS Nitro-based instances with an encrypted ephemeral EBS disk, the bosh-agent intermittently fails to bootstrap with timed out pinging VM or timed out sending get_task to instance.

The agent attempts to partition the root disk instead of the ephemeral disk, which causes the bootstrap to fail.

Environment

• bosh v282.1.2
• bosh-aws-cpi v107.0.1
• Instance type: t3a.nano (Nitro, no local instance storage)
• Cloud config:

  cloud_properties:
    ephemeral_disk:
      encrypted: true
      size: 20000
    instance_type: t3a.nano
  
  

Observed behaviour

• Intermittent — not consistently reproducible, timing-dependent
• Agent attempts to partition the root disk
• VM becomes unresponsive, deployment times out

Analysis

On Nitro instances, all disks appear as NVMe devices. The agent resolves /dev/sdb (the ephemeral disk name in settings) to an NVMe device via /dev/disk/by-id/ symlinks using IDDevicePathResolver. It is unclear whether encrypted EBS disks produce a different symlink name in /dev/disk/by-id/ than what the agent's glob pattern expects, or whether the disk ID passed from the CPI does not match the symlink.

Related

• cloudfoundry/bosh-agent#407 — fixes the same class of problem for NVMe instance
  storage discovery
• cloudfoundry/bosh-aws-cpi-release#196 — companion CPI-side change, may also be relevant here

Co-authored with Claude Code

Start with IDDevicePathResolver and investigate how encrypted ephemeral EBS disks are matched through /dev/disk/by-id/ on Nitro instances. Read cloudfoundry/bosh-agent#407 and bosh-aws-cpi-release#196 for the related agent and CPI changes. Done means bootstrap consistently selects the ephemeral disk instead of partitioning the root disk and avoids the reported timeouts.