Health, Actions and alerts
Sift works out whether each filter needs anything from a person, and shows that one answer everywhere: the health badge on the Filters page, the dashboard’s Needs attention, the alerts bell, and the filter’s own Actions tab. This page explains how that judgement is made, what the Actions tab lists, and when Sift raises and clears an alert.
One judgement, shown everywhere
Section titled “One judgement, shown everywhere”A filter’s health is the list of conditions it is in right now. Each condition has a fixed code and one of two severities:
| Severity | Console badge | Meaning | Examples |
|---|---|---|---|
| Fault | Fault | Something is not doing what the console says it is doing. | Escalations not reaching the agent; the source failing to poll; an alarm on a filter set to keep filtering through it. |
| Attention | Attention | Nothing is broken, but nothing moves until a person acts. | An approval waiting; a fall-back to shadow; an agent returning no verdicts; a setup step that is yours. |
Conditions are ordered by how much they matter, faults first. The first one is the filter’s verdict: the one the Filters page, the dashboard and the bell report. A filter that is both failing to deliver and short of labels is reported as failing to deliver. A filter with no condition is healthy and shows no health badge. A retired filter has no conditions.
Because Sift makes this judgement in one place, the list, the dashboard and the filter’s page cannot disagree. For every code and its meaning, see States, badges and health codes.
Status badges
Section titled “Status badges”A filter’s status is at most two badges, wherever it appears:
- What it is doing. Filtering for an active filter that is
filtering; for any other state, its name (
shadow,proposed,pausedand so on), which says it is not filtering. An active filter that filters nothing shows no state badge, and always carries an Attention badge saying so. - Whether anything is wrong. Fault or Attention, absent when the filter is healthy. This badge links to the filter’s Actions tab.
The Actions tab
Section titled “The Actions tab”Actions is the first tab on a filter’s page. Its count is the conditions and recommendations waiting on a person, coloured by the worst of them. Opening a filter that has any ask lands on this tab; one with none opens on Overview. A retired filter has no Actions tab. The tab lists every ask, of three kinds, in this order:
| Section | Kind | What it holds | Can you dismiss it? |
|---|---|---|---|
| Needs you | Condition | Every condition the filter is in, the verdict first, and each setup step a person owes. | No. It clears when the condition is fixed. |
| Sift recommends | Recommendation | Advice you may take or decline. There is one kind: a bake-off candidate that measured better recall than the deciding recipe, with Promote… and Decline…. | Yes: Decline… with a reason. |
| Sift is doing | Progress | What Sift is doing on the way to filtering, such as collecting labels or training, sometimes with a shortcut such as Train now. While the filter has Setup steps, this shows on the step’s row in Setup instead of in a section of its own. | Not needed. |
When no condition or recommendation is waiting, the tab says “Nothing Sift needs from you.”
Each condition names who acts, and carries the control that acts on it or a link to the page where you do:
| Who chip | Meaning |
|---|---|
| Your move | You can do it now. |
| An operator | It needs an operator or admin. A reviewer sees this chip, and “An operator or administrator does this.” in place of the control. |
| In Kindo | It is done in Kindo, not in Sift (retiring a displaced agent, for one). |
A recommendation’s chip reads Advisory. Nothing in a recommendation acts on its own. Promote… makes the recommended recipe the one that decides, and it filters only after a threshold approval. On a filter that is filtering, the button reads Promote and stop filtering…: the filter drops to shadow until an operating point for the new recipe is approved. Decline… asks Why not, on this evidence (required), and Decline the recommendation records it. A declined recommendation stays hidden until a different recipe is recommended or the holdout it was measured on grows by a tenth.
Changes you choose to make yourself, rather than ones Sift asks for, are in the Settings tab.
Setup steps
Section titled “Setup steps”Below the asks, the Actions tab shows Setup: the steps from a new filter to one that filters, each marked You or Sift (with ” · now” on the step that is yours to do next). Setup steps exist while the filter is in draft, authoring, shadow or proposed. The heading says where the filter is: Setup · step 3 of 7.
| Step | Who |
|---|---|
| Write its spec (only while the filter has none) | You (operator or admin) |
| Connect its agent (only for a filter that sends whole events to a paired agent) | You (operator or admin) |
| Review the gold set | You (judging: anyone signed in; approving: operator or admin) |
| Start watching in shadow | You (operator or admin) |
| Collect {N} labels | Sift |
| Collect 20 actionable events | Sift |
| Train a model | Sift; an operator or admin when the retrain schedule is never, or with Train again after a failed run |
| Approve an operating point | You (operator or admin), with Approve the threshold, or Approve the budget for a filter decided by an hourly budget |
A filter decided by an hourly budget whose last model proposed no budget shows Collect verdicts on ranked hours in place of the label steps.
A filter whose pipeline has extra stages gets their steps too: for Collapse repeats, Measure collapsing, Collect {N} cluster verdicts and Approve collapsing repeats; for Shape delivery, Pair an agent that takes groups, Measure what grouping saves and Switch grouping on. A filter decided by Collapse repeats alone has no label, training or operating-point steps: Collapse’s own steps lead to filtering. A filter that Sift stopped, or that a person put back in shadow with its approval standing, gets Resume on the current model.
Each step has one of five statuses. On screen a row shows ✓ when it is done, ! when it has stopped, and its number otherwise; the words in the table are what a screen reader hears.
| Status | Mark | Read as | Meaning |
|---|---|---|---|
done | ✓ | Done | Finished. |
action | number | Your move | A person can do it now. |
waiting | number | In progress | Sift is doing it now; its progress, and any shortcut such as Train now, show on the row. |
problem | ! | Stopped | It cannot go on without a person: a build or a training run failed, a label count stopped moving, or the paired agent needs regenerating, its instructions filled in, or its verdicts back. |
upcoming | number | Not yet | It cannot start yet. |
A step that is yours also appears as a card in Needs you; its row in Setup then says “Above, in Needs you.”
Waits that stall become asks. A step Sift is waiting on is not a problem by itself: a quiet day is just a quiet source. But if a label count stands still for 7 days (and no model is training), Sift marks the Collect {N} labels step stopped and gives it a button for an operator, Review where its labels come from, because a source, an agent or a pairing has stopped producing what the step waits for, and nothing Sift does on its own will move it.
Alerts
Section titled “Alerts”The Alerts page, reached from the bell in the header, lists everything that is broken or owed across Sift, worst first, in two groups:
- Faults: “Something is not doing what this console says it is doing. These do not wait for a decision.” Filters whose verdict is a fault, critical deployment warnings, and a Kindo connection that is not answering.
- Waiting on you: “Nothing here is broken. None of it moves until somebody acts.” Filters needing attention, other deployment warnings, and a degraded Kindo connection.
Each heading carries its count, as in Faults (2). Every row is a condition true right now, and none can be dismissed; each clears when it is fixed. The bell’s count is faults plus items waiting. The default-password warning and the Kindo credential warnings are not on this page or in the count: they have their own notices at the top of every screen.
Raised and cleared alerts
Section titled “Raised and cleared alerts”For notifications, Sift turns these conditions into moments: an alert is raised and later cleared. A condition is raised once it has held for 2 minutes, and cleared once it has been absent for 10 minutes, so a condition that flickers briefly is never sent, and one that returns quickly does not close and reopen. Sift checks every 60 seconds. A retired filter raises nothing, and its open alerts clear at once.
A filter raises one alert, for the condition that leads its verdict, with one exception: the conditions of the agent’s load. A held queue losing or filling, a budget bound or recording its overflow, a filter Sift moved to its safer point, and a shadow filter over its agent’s capacity with no budget are each an alert of their own, even behind another condition. See Alerts for the codes.
Not every condition on the Alerts page becomes an alert. Deployment warnings that restate
a filter’s own fault (escalations not delivered or failing, an unfinished or unverified
cutover, overdue group deliveries) are sent as that filter’s alert instead;
the warnings that Sift’s own background processing is unreachable or silent
(broker_unreachable, worker_pool_silent) are never sent, so watch for them on the
Alerts page; and the notice that the decision model is called
directly is not a fault, so it is never sent.
Alerts leave Sift only through webhooks, which an admin sets up under the gear icon (Administration) › Notifications. None are configured by default. Each endpoint subscribes to filter faults, filter attention, deployment conditions, or a mix. A webhook is sent when an alert is raised and when it clears (a clear goes only to the endpoints its raise went to), is signed, and never carries event content, scores or verdicts. See Notifications and webhooks.