62 lines
2.1 KiB
Markdown
62 lines
2.1 KiB
Markdown
---
|
|
title: Bringing Cloud-Native Agility to Bare-Metal Kubernetes with Cluster API and Metal³
|
|
weight: 5
|
|
tags:
|
|
- capi
|
|
- baremetal
|
|
---
|
|
|
|
<!-- {{% button href="https://youtu.be/rkteV6Mzjfs" style="warning" icon="video" %}}Watch talk on YouTube{{% /button %}} -->
|
|
<!-- {{% button href="https://docs.google.com/presentation/d/1nEK0CVC_yQgIDqwsdh-PRihB6dc9RyT-" style="tip" icon="person-chalkboard" %}}Slides{{% /button %}} -->
|
|
|
|
|
|
Premise: We are in the **cloud native** era, not the **cloud* era -> We can do stuff onprem
|
|
|
|
## Baseline
|
|
|
|
- Baremetal is cool b/c: No overhead, no hypervisor license, direct scheduling access
|
|
- But: Low provisioning, inflexible, needs drivers and so on
|
|
- Cluster Provisioning can be done in many different ways, even kube-native with CAPI (yay)
|
|
- CAPI is cool and can be used with many different infra providers
|
|
|
|
> I wont repeat the idea behind CAPI here (which were the next 5-10min of the talk)
|
|
|
|
## What about baremetal
|
|
|
|
> Metal³ (MEtal-kubed) to the rescue
|
|
|
|
```mermaid
|
|
graph LR
|
|
CAPI-->BareMetalOperator-->Ironic-->BMC("BMC (IPMI, etc)")
|
|
```
|
|
|
|
### Ironic
|
|
|
|
- API: Expose actions
|
|
- Conductor: Talk to MBC
|
|
- Agent: Ramdisk for disk cleaning and all other initial local stuff
|
|
- Inspector: Collect HWInfo
|
|
- DB: Store state
|
|
|
|
### Baremetal operator
|
|
|
|
- Runs in management cluster and watched BaremetalHost CR
|
|
- Can be used with or without CAPI
|
|
- BaremetalHost controls stuff like: Firware, Power State, Raid, Boot mode, ...
|
|
- HardwareData CRD contains information about the Host after it has been checked by Ironic Inspector
|
|
|
|
TODO: Steal state flow chart from slides
|
|
|
|
### Provisioning
|
|
|
|
- Baseline: Host is in the Available state
|
|
- Action: Assign a OS image with userdata to a host and specify the storage device (defaults to /dev/sda)
|
|
- Reaction: Metal³ chooses a fitting baremetalhost to consume
|
|
- Result: System boots into os image and initialized via ignition (userdata)
|
|
|
|
## Observations/Other
|
|
|
|
- Metal³ is pretty slow (due to hardware and ironic)
|
|
- You need spare servers for rolling update -> You can reuse a existing server during upgrade via Recreate Update
|
|
- We can controll some harware operations via node annotations like shutdown/reboot/maintainance
|