Module health 

The alert rules the credit registration dashboard renders.

An alert carries identifiers and numbers, never prose: the study registry’s own error text is written for an integrator and is not translated, so the frontend renders one key per alert id with these values interpolated. The thresholds travel with the alerts, not hardcoded twice.

Structs

CreditRegistrationAlert

CreditRegistrationHealth

Enums

CreditRegistrationAlertId

CreditRegistrationAlertSeverity

Constants

CREDENTIAL_REJECTION_WINDOW 🔒

Within this much of the past, one rejected credential is enough.

IDLE_QUEUE_DEPTH 🔒

Queued work that makes a day without a single completion mean something.

LATENCY_REGRESSION_FACTOR 🔒

LATENCY_REGRESSION_FLOOR 🔒

Under this the registry is quick enough that a doubling says nothing.

LATENCY_WINDOW 🔒

LINKING_MAIL_HOURLY_CAP 🔒

Linking mails one hour may hand over before the volume itself is the problem.

LINKING_MAIL_RATE_WINDOW 🔒

LINKING_MAIL_WINDOW 🔒

MISREGISTRATION_CRITICAL_COUNT 🔒

A reversal is always worth saying; this many at once is an incident.

NEVER_ENTERED_MIN_AGE 🔒

How long a completion may sit outside the ledger before materialize is the suspect rather than the clock.

NEVER_ENTERED_SAMPLE_LIMIT 🔒

Bounds the anti-join behind that rule; a bigger backlog reports as this many.

PERMANENT_FAILURE_COUNT 🔒

PERMANENT_FAILURE_RATE_PERCENT 🔒

PHASE_CONSECUTIVE_FAILURE_LIMIT 🔒

Failures in a row before a phase counts as broken rather than unlucky.

PHASE_HEARTBEAT_INTERVAL_MULTIPLIER 🔒

A phase is late once this many of its own intervals have passed without a heartbeat.

PHASE_SUCCESS_CALL_MARGIN 🔒

PHASE_SUCCESS_INTERVAL_MULTIPLIER 🔒

A phase that owns a nonempty queue and has not succeeded within this many of its own intervals is running without getting anywhere, which no failure count catches. Never less than its slowest possible iteration plus PHASE_SUCCESS_CALL_MARGIN, or one slow call would look like a wedge.

SERVICE_OUTAGE_FAILURE_SHARE_PERCENT 🔒

SERVICE_OUTAGE_MIN_ITEMS 🔒

Below this many items the share below is one bad batch, not a signal.

SERVICE_OUTAGE_WINDOW 🔒

STUCK_CRITICAL_COUNT 🔒

Above this many stuck rows the backlog stops being something to look at tomorrow.

STUCK_THRESHOLDS 🔒

SUOTAR_PENDING_WAIT 🔒

The longest submissionPending asks verify to wait before polling again.

TERMINAL_WINDOW 🔒

The window every “in the last day” rule shares.

UNREACHABLE_CONSECUTIVE_FAILURES 🔒

Below this the run is a bad minute rather than an outage.

UNREACHABLE_WINDOW 🔒

Long enough to hold three failed calls at the longest request timeout.

Functions

course_configuration_alert 🔒

Modules the last configuration check found broken. Never checked is not counted: the Courses tab renders unknown and broken differently, and so must this.

depth_of 🔒

evaluate

Runs every rule and ranks what it found.

is_heartbeat_late 🔒

A phase counts as late once more than PHASE_HEARTBEAT_INTERVAL_MULTIPLIER of its own interval has passed since its last heartbeat. A paused phase is never late: it is not expected to be heartbeating at all.

is_phase_failing 🔒

Whether a phase counts as failing: too many failures in a row, or a nonempty queue with no success for too long, or an iteration hung past that same bound. Never while paused. The one definition behind both the PhaseFailing alert and the Workers tab’s failing flag.

latency_regression_alert 🔒

How long the study registry is taking to confirm, this week against last. count is this week’s p95 in seconds and total last week’s, so the banner can name both.

linking_mail_alert 🔒

Linking mails we could not hand over at all. The recipient domain rides along because an undeliverable host is the usual cause.

linking_mail_rate_alert 🔒

The volume guard: how many people we mailed in the last hour against what an hour should hold. Counts addresses, which is what the per-person caps govern.

never_entered_alert 🔒

Completions old enough that materialize has had every chance and still has no ledger row for them. Sampled rather than counted, so the anti-join stops early on a large backlog.

phase_alerts 🔒

What the phase table says about itself: phases that stopped reporting, and phases that report but get nowhere.

roster_course_code_alert 🔒

Course codes whose roster fails even when listed on their own, so enrolment discovery is backing them off. The worst one rides along, since the usual cause is that one code’s configuration.

service_outage_alert 🔒

The share of recent items that failed on Suotar or Sisu being unavailable. Our only proxy for Sisu’s uptime, which is why it is a rule of its own rather than part of the request-level one above.

state_name 🔒

The state’s own wire name, taken from its serialisation so the two cannot drift.

stuck_alert 🔒

Rows the pipeline should have moved on by now. Terminal states are outside this by construction rather than by a filter that could be forgotten.

stuck_thresholds

study_registry_conflict_alert 🔒

Accounts the study registry reported a number for that another live link kept us from linking.

terminal_outcome_alerts 🔒

The three rules read off the last day’s terminal outcomes: failures piling up, reversals, and a pipeline that finished nothing while holding work.