[sunshowers]

This required many other replaces since we needed to pull in the appropriate key type impls. (This is a deliberate decision to make clients and servers agree on type semantics.)

There are two places where duplicates are newly detected on the client:

1. In list_all_probe_create_params_for_sled_batched.
2. In group_attached_subnets_by_sled_and_vmm.

What happens if there were duplicates? Previously, with both 1 and 2:

• The client would generate and send a Vec with the duplicates.
• The server would attempt to deserialize as an IdOrdMap.
• IdOrdMap rejects duplicates, so the server would return a 400.

With this change, duplicates are now detected on the client before they're sent to the server.

• For list_all_probe_create_params_for_sled_batched, this doesn't change anything -- we previously fail closed and now continue to fail closed. I didn't investigate beyond that point.

• For group_attached_subnets_by_sled_and_vmm, there is a change, because instead of failing with a 400, we now push duplicates to the SledSubnetDetails::errors list and otherwise successfully create a config to send to the server. I did spend some time trying to figure out whether uniqueness was guaranteed (because if it is, then this isn't a functional change in practice). My understanding is that:

  • This commit does not introduce any functional changes in practice.
  • But this relies on an application-level invariant in Nexus, not a database-level invariant.
  • Support for attaching VPC Subnets to network interfaces #9580 will likely affect this and should be kept in mind.

  I've tried capturing my understanding in code comments -- happy to take suggestions on improvements.

[sunshowers]

@bnaecker added you since you filed #9580 and probably have the most context about attached subnets.

[bnaecker]

Can you file an issue for this? We should have such a constraint.

[sunshowers]

#11568. thanks!