Short summary: vSphere starts using an uplink as soon as its physical layer becomes operational, which might happen during ToR switch startup phase, or before a ToR switch port enters forwarding state.
Many devices verify higher-layer availability before they start to use a point-to-point physical network link. These mechanisms could operate at layer-2 (example: LACP or UDLD) or above it – from DHCP-based address allocation to blocking - listening - learning - forwarding state transition of STP and explicit adjacencies formed by most modern routing protocols. Using a link just because your hardware detects a carrier signal is often too risky, particularly for a device that claims to be a switch with tens of servers behind it.
Obviously different rules apply to x86-based networking (or at least some simplistic variants of it) – you’re free to use a link for user traffic as soon as you can start sending packets, and when users start complaining, they get a recommendation to use manual failback and change switch configuration. One must wonder when the be conservative in what you do principle got lost.
I’m picking on VMware, but I’m positive other operating systems aren’t much better. If they are, please write a comment.
VMware finally implemented a reasonable mechanism in ESXi 5.0 – a configurable link-up delay. Too bad its default value is too low, it’s hidden somewhere in the Advanced Settings tab, and the only documentation is an arcane knowledge base article. Having an active CFM probe between uplinks belonging to the same port groups (not constant beaconing, just a link-ready-for-forwarding test) would be even better, but it’s probably too much to hope for … or not?