bstein/ananke
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: replace legacy hecate README with ananke runbook

Browse Source
This commit is contained in:
Brad Stein 2026-04-07 12:40:45 -03:00
parent 26d934b675
commit d19862285a
1 changed files with 68 additions and 122 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

190
README.md
View File

@ -1,151 +1,97 @@
# Hecate # ananke
Hecate is the host-level bootstrap and power-protection service for Titan. `ananke` is the host-side power + bootstrap orchestrator for Titan.
It runs on `titan-db` and handles: It runs outside Kubernetes (systemd on host), so it can:
- Staged **startup** (including Flux/Gitea bootstrap deadlock fallback) - shut the cluster down gracefully before battery/runtime redlines
- Graceful **shutdown** - bring the cluster back after power returns
- UPS-driven automatic shutdown decisions based on discharge/runtime - recover common Flux/Kustomize startup deadlocks
- Multi-UPS operation via multiple Hecate instances (for example `titan-db` + `tethys`) - validate service health from the outside before declaring startup done
- Full hardware poweroff sequencing after graceful Kubernetes shutdown
## Why host-level ## Why `ananke`
A service inside Kubernetes cannot start a cluster that is fully down. I wanted a name that fits Titan/mythology, but also describes what this service actually does.
Hecate runs outside the cluster under systemd, so it can always orchestrate bring-up.
## Commands In Greek myth, **Ananke** is inevitability/necessity. That matches this tool: when power events happen, graceful sequencing is not optional.
- `hecate startup --config /etc/hecate/hecate.yaml --execute --force-flux-branch main` UPS names in this cluster are also part of the story:
- `hecate shutdown --config /etc/hecate/hecate.yaml --execute` - `Statera`: powers `titan-23`, `titan-24`, `titan-jh`
- `hecate daemon --config /etc/hecate/hecate.yaml` - `Pyrphoros`: powers all other nodes
- `hecate status --config /etc/hecate/hecate.yaml`
Key startup guards: ## Breakglass reminder
- Startup is blocked on hosts configured as `coordination.role: peer` (unless `--allow-peer-startup` is used intentionally).
- `--auto-peer-failover` makes peer hosts hand off startup to the coordinator first, then run local startup only if the coordinator is unreachable.
- Startup is blocked while UPS is on battery by default (unless `--allow-on-battery` or `coordination.allow_startup_on_battery: true` is set).
- Startup is blocked when a shutdown intent is active (`/var/lib/hecate/intent.json`).
- Stale shutdown intents are auto-cleared after `coordination.startup_guard_max_age_seconds`, so old outage residue cannot permanently deadlock startup.
- Startup checks configured `coordination.peer_hosts` intents to avoid peer/coordinator split-brain startup races.
- Startup waits for time sync in `strict` or `quorum` mode (`startup.time_sync_mode`, `startup.time_sync_quorum`).
- Startup can block until storage is healthy (`startup.require_storage_ready` + critical PVC checks).
- Startup can block until external probes pass (`startup.require_post_start_probes` + `startup.post_start_probes`).
- Startup refreshes and can use a cached bootstrap manifest set under `/var/lib/hecate/bootstrap-cache` when local fallback paths fail.
- Vault unseal now falls back to a local cached key file (`startup.vault_unseal_key_file`) if `vault-init` cannot be read yet.
- Optional off-site break-glass retrieval can be configured with `startup.vault_unseal_breakglass_command` (for example, an SSH `cat` command to a remote key escrow host).
## Manual install on titan-db Vault unseal breakglass is wired for remote retrieval (magic mirror host). If local key retrieval fails, Ananke can use the configured breakglass command.
## What “startup complete” means now
Ananke does **not** stop at “Flux says Ready”. Startup only completes when all configured gates pass:
- Flux source drift guard passes (`expected_flux_source_url` + branch expectation)
- Flux kustomizations are healthy
- controller convergence is healthy (deployments/statefulsets/daemonsets)
- external service checklist passes (for example Gitea + Grafana health endpoints)
- stability soak window passes (no regressions, no CrashLoop/ImagePull failures)
If any gate fails, startup is blocked with a concrete reason.
## Command quick sheet
From `titan-db` (coordinator):
```bash ```bash
git clone git@gitea-admin:bstein/hecate.git sudo /usr/local/bin/ananke status --config /etc/ananke/ananke.yaml
cd hecate sudo /usr/local/bin/ananke startup --config /etc/ananke/ananke.yaml --execute --force-flux-branch main
sudo HECATE_ENABLE_BOOTSTRAP=1 ./scripts/install.sh sudo /usr/local/bin/ananke shutdown --config /etc/ananke/ananke.yaml --execute --reason graceful-maintenance --mode cluster-only
sudoedit /etc/hecate/hecate.yaml sudo /usr/local/bin/ananke shutdown --config /etc/ananke/ananke.yaml --execute --reason emergency-power --mode poweroff --skip-drain --skip-etcd-snapshot
sudo systemctl restart hecate.service
``` ```
The installer is idempotent: From `titan-24` (`tethys` peer):
- Re-runs safely on every update
- Preserves existing `/etc/hecate/hecate.yaml`
- Automatically migrates legacy defaults (for example `default_budget_seconds: 300` and `runtime_safety_factor: 1.10`)
- Ensures required dependencies are installed (`kubectl`, `nut-*`, `ssh`, `go`, etc.)
- Installs/refreshes systemd units and enables boot-time self-update
- Applies declarative NUT + udev UPS configuration by default (can be tuned via env vars)
Installer knobs (optional):
- `HECATE_ENABLE_BOOTSTRAP=1` enables `hecate-bootstrap.service` on this host.
- `HECATE_ENABLE_BOOTSTRAP=0` disables it; default `auto` enables bootstrap by default.
- `HECATE_MANAGE_NUT=0` skips writing NUT/udev files.
- `HECATE_FORCE_CONFIG_TEMPLATE=coordinator|peer|example` overwrites `/etc/hecate/hecate.yaml` from a known template during install.
- `HECATE_NUT_UPS_NAME` (default inferred from `/etc/hecate/hecate.yaml` target, fallback `pyrphoros`)
- `HECATE_NUT_VENDOR_ID` / `HECATE_NUT_PRODUCT_ID` (defaults `0764` / `0601`)
- `HECATE_NUT_MONITOR_USER` / `HECATE_NUT_MONITOR_PASSWORD` (defaults `monuser` / `hecateupsmon`)
Bootstrap now (without reboot):
```bash ```bash
sudo systemctl start hecate-bootstrap.service sudo /usr/local/bin/ananke shutdown --config /etc/ananke/ananke.yaml --execute --reason graceful-maintenance --mode cluster-only
``` ```
## Preconditions on titan-db Systemd:
- `kubectl` installed and configured (`kubeconfig` path in config) ```bash
- SSH reachability to all cluster nodes sudo systemctl status ananke.service
- Remote sudo rights to run: sudo systemctl start ananke-bootstrap.service
- `systemctl start/stop k3s` sudo systemctl start ananke-update.service
- `systemctl start/stop k3s-agent` ```
- UPS telemetry available via NUT (`upsc`)
Optional SSH jump/bastion: ## Shutdown modes (explicit)
- Set `ssh_jump_host` (and optional `ssh_jump_user`) to route node SSH through a jump host like `titan-jh`; Hecate now falls back to direct SSH automatically if jump routing is unavailable.
- Set `ssh_port`, `ssh_config_file`, `ssh_identity_file`, and `ssh_node_hosts` so root-run systemd actions can actually reach node SSH daemons during cold-start recovery.
- Use `ssh_node_users` for per-node username overrides (for example `titan-24: tethys`).
- Use `ssh_managed_nodes` to limit host-level SSH start/stop actions to nodes Hecate can actually authenticate to.
## Multi-UPS topology `ananke shutdown` now supports explicit mode selection:
- `--mode config`: use config default (`shutdown.poweroff_enabled`)
- `--mode cluster-only`: stop cluster services only (no host poweroff)
- `--mode poweroff`: include host poweroff path
Recommended: This removes ambiguity during drills.
- `titan-db` runs Hecate as the shutdown coordinator with UPS `Pyrphoros` (`pyrphoros@localhost`).
- `tethys` runs Hecate as a peer with UPS `Statera` (`statera@localhost`) and forwards shutdown triggers to `titan-db`.
- The bootstrap unit now runs on both roles; peer role uses auto-failover handoff to coordinator before local fallback startup.
- If forwarding fails, fallback local shutdown can remain enabled.
- Use `coordination.role: coordinator` on `titan-db` and `coordination.role: peer` on `tethys`.
## Config ## Config file
See `configs/hecate.example.yaml`. Primary path:
- `/etc/ananke/ananke.yaml`
Break-glass unseal fallback knobs: Core settings to keep accurate:
- `startup.vault_unseal_breakglass_command`: optional shell command that prints the unseal key to stdout. - `expected_flux_branch`
- `startup.vault_unseal_breakglass_timeout_seconds`: timeout for the command (default `15`). - `expected_flux_source_url`
- `startup.shutdown_cooldown_seconds`: cooldown window after shutdown completion before startup proceeds (default `45`). - `startup.service_checklist`
- `startup.service_checklist_stability_seconds`
- `startup.ignore_unavailable_nodes` (for planned temporary node outages)
- `coordination.role`, `coordination.peer_hosts`
UPS auto-shutdown trigger uses: ## Install / update
- runtime threshold = `runtime_safety_factor * estimated_shutdown_budget`
- default safety factor `1.25`
- debounce across multiple polls to avoid noise
- emergency trigger budget defaults to `shutdown.emergency_budget_seconds` and is learned from historical UPS-triggered shutdown runs once enough samples exist
- UPS-triggered shutdown executes the emergency fast path by default (`shutdown.emergency_skip_drain: true`, `shutdown.emergency_skip_etcd_snapshot: true`)
Estimated shutdown budgets are derived from historical successful shutdown runs (`/var/lib/hecate/runs.json`) with config fallbacks: ```bash
- `estimated_shutdown_budget_seconds`: full/manual shutdown path sudo ./scripts/install.sh
- `estimated_emergency_shutdown_budget_seconds`: UPS/emergency path ```
Power metrics: Installer behavior:
- Hecate exposes Prometheus metrics on `:9560/metrics` by default. - builds and installs `/usr/local/bin/ananke`
- This is intended for a dedicated Grafana power dashboard and a high-level overview row. - installs `ananke*.service` units
- migrates and enforces current `ananke` config/state paths
## Notes ## Notes
- Default behavior for `startup` and `shutdown` is dry-run unless `--execute` is set. - Apply changes through Git/Flux manifests; avoid manual in-cluster edits for durable changes.
- Hecate tracks intent in `/var/lib/hecate/intent.json` (`normal`, `startup_in_progress`, `shutting_down`, `shutdown_complete`) to avoid startup/shutdown fighting each other. - For controlled shutdown/startup drills, treat any manual intervention as a bug and fold the logic back into Ananke.
- Startup now waits out the recent-shutdown cooldown window instead of failing immediately when shutdown completed moments ago.
- In multi-instance setups, set `coordination.peer_hosts` on each host (for example `titan-db` <-> `titan-24`) so startup guards account for remote intent too.
- `hecate-bootstrap.service` is enabled to run at host boot and perform staged startup automatically.
- `HECATE_ENABLE_BOOTSTRAP=1` forces bootstrap on, `HECATE_ENABLE_BOOTSTRAP=0` forces it off, and `auto` enables by default.
- `hecate-update.timer` runs on boot and periodically to pull latest `main` and reinstall Hecate declaratively.
- Peer startup fallback now checks coordinator intent/bootstrap activity before allowing local startup.
- Automatic etcd recovery can run during startup if API never becomes reachable (`startup.auto_etcd_restore_on_api_failure`).
## Etcd Recovery
- Manual: `hecate etcd-restore --config /etc/hecate/hecate.yaml --execute`
- Optional snapshot override: `--snapshot /var/lib/rancher/k3s/server/db/snapshots/<name>`
- Startup can automatically invoke the same restore path after API timeout using:
- `startup.auto_etcd_restore_on_api_failure: true`
- `startup.etcd_restore_control_plane: <control-plane-node>`
- If control planes are configured with `--datastore-endpoint` (external DB), Hecate will skip etcd restore and retry control-plane startup instead.
- Etcd restore now verifies snapshot existence, minimum size, listing presence, and SHA-256 before reset starts.
## Disruptive startup drills
Hecate includes scripted disruptive drills that intentionally break critical services and verify startup recovery paths:
- `scripts/hecate-drills.sh list`
- `scripts/hecate-drills.sh run flux-gitea-deadlock --execute`
- `scripts/hecate-drills.sh run foundation-recovery --execute`
- `scripts/hecate-drills.sh run reconciliation-resume --execute`
- `scripts/hecate-drills.sh run controlled-cycle --execute` (uses `HECATE_DRILL_SHUTDOWN_CONFIG`, defaults to `/tmp/hecate-drill-no-poweroff.yaml`)
These drills are intentionally **not** part of regular `go test ./...`.