Alerts#

Alerts tell people about a few important events — a run failed, a machine stopped answering, GPUs are unhealthy, a workspace reached its quota — by e-mail, in a Slack channel, or at a webhook of yours. An organisation sets up channels (where alerts go) and rules (which events go to which channel). To send every event to a SIEM instead, use an event stream.

Before you begin#

You are an owner or admin of the organisation. Only they see and change alerts.
Alerts are derived from events. Every cluster Astralyx provides — Astraeus Cloud or a cluster dedicated to your organisation — sends its events to your organisation; nothing to configure.

How alerts work#

flowchart LR
  E[Cluster event] --> P["Astralyx control plane (SaaS)"]
  P -->|classified| R{Rules that want it}
  R --> D[Delivery queue]
  D -->|e-mail / Slack / webhook| C[Channel]

Each event that reaches the platform is classified as one of the alerts below, or as none.
Every enabled rule of the organisation that wants that alert, for the event's workspace or for every workspace, gets one delivery for that event.
The platform sends deliveries in order of their due time, and retries failed ones (see Delivery and retries).

Alerts reference#

The names are used in rules and sent in payloads; they keep the API's words (job is a run, task a worker).

Alert	Console description	Sent when
`job_failed`	A run failed	A run enters `Failed`.
`job_completed`	A run completed	A run enters `Completed`.
`task_failed`	A worker failed (each one of a run's)	A worker enters `Failed`: one alert per worker.
`quota_reached`	A run waits because its workspace reached its quota	A run starts waiting on its workspace's quota: once when it starts waiting, not again while it keeps waiting.
`machine_down`	A machine stopped answering	A machine enters `Down`.
`machine_up`	A machine answers again	A machine is `Up` again after its heartbeat returns.
`gpu_degraded`	A machine's GPUs are not healthy (a fault, no driver)	A machine's `GPUsHealthy` or `GPUSubsystemReady` condition turns `False`.
`machine_condition`	A machine's disks, memory or RAID are in trouble	`MemoryPressure`, `DiskPressure` or `CPUPressure` turns `True`, or `StorageRedundant` turns `False`.
`deployment_down`	A deployment failed, or has served nothing for 5 minutes while not scaled to zero	A deployment enters `Failed`; or it left `Ready` for `Pending` or `Starting` and stayed there 5 minutes (checked every minute; not when it scaled to zero on purpose).

Recoveries other than machine_up (a condition back to normal, a run restarting) are not alerts.

Machine events belong to the organisation, not to a workspace: only rules for every workspace, and the machines receive machine_down, machine_up, gpu_degraded and machine_condition.

Add a channel#

Kind	Settings	Notes
E-mail (`email`)	Addresses: 1 to 20, comma-separated	Sent from the platform's mail sender.
Slack (`slack`)	URL: a Slack incoming webhook, `https://hooks.slack.com/…`	Create one in Slack under Apps → Incoming Webhooks for the channel.
Webhook (`webhook`)	URL: `http://` or `https://`	Signed; see Webhook payload.

Slack and webhook URLs are credentials: they are stored encrypted and shown afterwards only as their scheme and host (https://hooks.slack.com/…). The platform reaches them only on public addresses (400 UNREACHABLE_ADDRESS otherwise).

ConsoleAPI

Open Organisation → Alerts.
Under Channels, select Add channel.
Choose the Kind, enter a Name (for example on-call), and the Addresses or URL.
Select Add. For a webhook, the next dialog shows The webhook's signing key once: copy it now.
Select Send a test in the channel's row. The test appears under Sent within seconds.

The Alerts page with two channels, three rules and recent deliveries

$ curl -sS -X POST "$ASTRA_URL/api/v1/orgs/acme/alerts/channels" \
    -H "Authorization: Bearer $ASTRA_TOKEN" -H "Content-Type: application/json" \
    -d '{"kind": "webhook", "name": "pager", "url": "https://hooks.example.com/astraeus"}'
{"id":"0192…","kind":"webhook","name":"pager","signing_key":"9c1f…e04a"}

signing_key is in this answer only, and null for e-mail and Slack. For e-mail, send "addresses": "[email protected], [email protected]" instead of url.

Error	Cause
`400 INVALID_CHANNEL`	Unknown kind; no address or more than 20; an address without `@` or over 254 characters; a URL that is not `http(s)`; a Slack URL not on `hooks.slack.com`.
`400 UNREACHABLE_ADDRESS`	The URL's host is not a public address.

Add a rule#

A rule sends one or more alerts, from one workspace or from all of them and the machines, to one channel. A channel needs to exist first.

ConsoleAPI

In Organisation → Alerts, under Rules, select Add rule.
Under When, tick the alerts. job_failed, machine_down and gpu_degraded are ticked by default.
Under Where from, choose a workspace, or every workspace, and the machines.
Under To, choose the channel.
Select Add.

Untick On in a rule's row to pause it; Remove deletes it.

$ curl -sS -X POST "$ASTRA_URL/api/v1/orgs/acme/alerts/rules" \
    -H "Authorization: Bearer $ASTRA_TOKEN" -H "Content-Type: application/json" \
    -d '{"channel_id": "0192…", "workspace": null, "events": ["job_failed", "machine_down", "gpu_degraded"]}'
{"id":"0192…"}

Change it with PATCH /orgs/{org}/alerts/rules/{id} and {"enabled": false} and/or {"events": [...]}; delete it with DELETE /orgs/{org}/alerts/rules/{id}.

Error	Cause
`400 INVALID_RULE`	No events, or an unknown alert name.
`404 CHANNEL_NOT_FOUND`	The channel is not the organisation's.
`404 WORKSPACE_NOT_FOUND`	No workspace with that slug.

Removing a channel removes its rules too.

Messages#

E-mail#

Subject [Astraeus] <title>, for example [Astraeus] Run train failed: exit code 1. The body gives the title, the workspace and cluster (or the cluster, for machines), the organisation, the time, and a link to the run, worker or cluster in the console.

Slack#

One message per alert: the title in bold, where it happened, and an Open in Astraeus link.

Webhook payload#

Each delivery is a POST with a JSON body:

{
  "alert": "job_failed",
  "title": "Run train failed: exit code 1",
  "org": "acme",
  "cluster": "lab-a",
  "workspace": "vision",
  "kind": "job",
  "name": "train",
  "state": "Failed",
  "reason": "exit code 1",
  "at": "2026-10-01T09:12:44.103Z",
  "delivery": 4812,
  "url": "https://console.astralyx.cloud/o/acme/w/vision/jobs/lab-a/train"
}

Field	Description
`alert`	The alert name; `test` for a test.
`title`	A one-line summary.
`org`, `cluster`	Slugs.
`workspace`	The workspace's slug, or `null` for a machine's event.
`kind`	`job`, `task`, `node` or `deployment` (`test` for a test).
`name`	The run, worker, machine or deployment, as the workspace names it.
`state`, `reason`	The state entered and why.
`at`	When it happened (RFC 3339).
`delivery`	The delivery's ID: the same on every retry of it.
`url`	The console page it is about.

Headers:

Header	Value
`Content-Type`	`application/json`
`User-Agent`	`Astraeus-Alerts`
`X-Astraeus-Delivery`	The delivery ID. Use it to drop duplicates: a delivery may arrive more than once.
`X-Astraeus-Event`	The alert name.
`X-Astraeus-Signature`	`sha256=` and the hex HMAC-SHA256 of the raw body.

Verify the signature#

The HMAC key is the signing key exactly as shown (its 64 characters as ASCII bytes; do not hex-decode it). Compute the HMAC over the raw request body, before any JSON parsing, and compare in constant time.

verify.py

import hashlib
import hmac

def verified(body: bytes, header: str, key: str) -> bool:
    expected = "sha256=" + hmac.new(key.encode(), body, hashlib.sha256).hexdigest()
    return hmac.compare_digest(expected, header)

From a shell, for a saved body:

$ printf 'sha256=%s\n' "$(openssl dgst -sha256 -hmac "$SIGNING_KEY" -hex < body.json | sed 's/.*= //')"
sha256=5b0c7e…

Delivery and retries#

Deliveries are sent within seconds of the event reaching the platform.
A webhook or Slack delivery succeeds on any 2xx answer within 15 s.
A failed delivery is retried after 1, 2, 4, 8, 16 and 32 minutes, then after 1 hour; after 8 attempts it is marked failed.
Sent in the console lists the last 50 deliveries with their state (pending, delivered, failed), attempts and last error.
Delivery is at least once. Deduplicate on X-Astraeus-Delivery.

Test a channel#

Send a test (or POST /orgs/{org}/alerts/channels/{id}/test, answered 202 {"delivery": <id>}) queues a message titled "Test alert: this channel works", delivered like any other alert, with alert and kind set to test. Check its state under Sent.

Troubleshooting#

Symptom	Cause	Fix
No alerts at all from a cluster	No enabled rule wants these alerts, or the rule is for another workspace.	Check the rules under Organisation → Alerts; check the cluster's events under Organisation → Clusters & machines → the cluster → Events.
Machine alerts never arrive	The rule is for one workspace.	Use every workspace, and the machines.
A delivery is `failed` with `answered 401`	The webhook refused the request.	Check the receiving end, then Send a test.
A delivery is `failed` with `… is not reachable from here`	The URL resolves to a private address.	Use a public endpoint.
Signature mismatch	The HMAC was computed over parsed JSON, or with a hex-decoded key.	Use the raw body and the key's ASCII bytes.
E-mails never arrive	The message was filtered as spam, or an address is wrong.	Check the delivery's state under Sent, then the recipients' spam folder.