Distinguish between Scrimlets and Gimlets

[smklein]

Part of #823

Fixes #1511

Sled-Agent

• Uses the presence/absence of RSS to decide "am I a Scrimlet"
• Sends that info upward to Nexus

Nexus

• Parses the "is this Sled a scrimlet" info out of the sled registration request
  • While I was there, renamed the function, made it PUT not POST
• Stores this info in the DB
  • I plan on using this info in follow-up PRs to decide "should Dendrite be running here".

[bnaecker]

Looks good to me! I don't think it matters much now, but we'll definitely need to handle the sled agent coming up without knowing whether it's running on a Scrimlet, and asking Nexus to update the database once that can be determined.

[smklein]

Looks good to me! I don't think it matters much now, but we'll definitely need to handle the sled agent coming up without knowing whether it's running on a Scrimlet, and asking Nexus to update the database once that can be determined.

Sounds good. Updated the upsert function to update this value accordingly.

[rmustacc]

Just a note that this changes at runtime with respect to hardware. It's not a static thing and the underlying hardware presence will come and go as the Sidecar is power cycled.

[smklein]

That matches my understanding. I do think there is more work that'll need to smooth out the case of "a sled has been removed from one location in the rack, and plugged in elsewhere" - perhaps a call to Nexus, to figure out what services should change - but otherwise, this makes sense.

[rmustacc]

Yeah, while the movement bit seemed captured, I just wanted to make sure it was clear that you might choose to start a bootstrap agent before this is known and wanted to call that out as part of this given what appeared the initial start the agent piece was encoding this knowledge right now.

[bnaecker]

Makes sense, thanks Robert. I believe Sean added functionality to update the database's record including the is_scrimlet value in d172c9c.

One question I have though is about when that value is updated. At RSS-time, it's definitely possible that the sled agent comes up before Sidecar, so we'd need to support updating it from false to true. But it seems kind of "sticky" to me. That is, I'm not sure it makes sense to set this to false, at least not right away, when Sidecar is cycled. If that's done as part of an administrative action or by Nexus, it seems like we'd want to keep the knowledge that this is or was a Scrimlet, and report 503s from dpd or the sled agent. If Sidecar cycles unexpectedly, we probably still want to do that, along with generating an alert of some kind. I don't know if there's any action here, but spelling that out seems useful to me.

[rmustacc]

I suspect an initial policy of stickiness across a given boot will be a pretty useful way to go here. That is, when it's rebooted, it'll be something that is determined fresh and resets stickiness.