[saponifi3d]

BaseDetectorHandler._evaluate now runs every detector through one lifecycle. evaluate decides which groups trigger. For each triggered evaluation that has no result yet, the platform calls create_occurrence and builds the IssueOccurrence, evidence, fingerprint, and event data. Before this, each handler did that work itself.

flowchart LR
    P[DataPacket] --> X[extract_value]
    X -->|empty DetectorGroupValues| Skip[skip]
    X --> E["evaluate(data_packet, values)"]
    E -->|triggered, result=None| C[create_occurrence]
    C --> O[platform: IssueOccurrence + evidence + fingerprint + event data]
    E -->|already has a result, e.g. StatusChangeMessage| Out[results]
    O --> Out

Loading

A new detector that decides with the detector's data conditions now needs just two methods:

class ExampleDetectorHandler(DetectorHandler[ExamplePacket, float]):
    def extract_value(self, data_packet: DataPacket[ExamplePacket]) -> float:
        return data_packet.packet.value

    def create_occurrence(
        self, evaluation: DetectorEvaluation, data_packet: DataPacket[ExamplePacket]
    ) -> tuple[DetectorOccurrence, EventData]:
        return DetectorOccurrence(issue_title=..., subtitle=..., type=ExampleGroupType, level="error", culprit=""), {}

Thresholds and priorities come from the DataConditions on the detector's trigger group. To change ids or fingerprints, override get_event_id, get_occurrence_id, or get_issue_fingerprint. A handler that doesn't use a condition group subclasses BaseDetectorHandler and implements evaluate.

• create_occurrence(evaluation, data_packet): priority, group key, and trigger group evaluation now come from the DetectorEvaluation.
• StatefulDetectorHandler.evaluate adds dedupe, threshold counters, and state on top of super().evaluate() (this resolves its TODO). Its event/occurrence id behavior is now expressed through the get_event_id and get_occurrence_id hooks.
• Uptime overrides evaluate instead of _evaluate, so no handler overrides _evaluate anymore.
• GroupedDetectorEvaluationResult is renamed to DetectorEvaluations. trigger_group_evaluation is optional for handlers that don't use a condition group.

Issue platform payloads are unchanged. I recorded every _evaluate result (IssueOccurrence, StatusChangeMessage, event_data, and the evaluation artifact logged to EAP) from the product detector test suites on master and on this branch. After masking per-run values (DB ids, UUIDs, timestamps, random fixture names), both runs produced the same output:

Detector suite	Tests	Evaluations	Identical
Metric issues (incidents/)	34	49	34
Uptime (uptime/)	52	11	52
Preprod size analysis (preprod/size_analysis/)	31	23	31
Processing errors (processing_errors/)	9	8	9
Data packets (workflow_engine/processors/)	1	1	1

Five ProcessUpdateComparisonAlertTest cases were left out because they also give different results between runs on master. Product test files are untouched.

The one contract change: a triggered evaluation returned without a result used to pass through with no result. Now the platform builds an occurrence for it. Only the mock handler in processors/test_detector.py returned that shape, so the two tests that relied on it now assert the occurrence.

[RudraPatel2003]

I do not think from a type standpoint it is super clear that result=None means that this method will handle create_occurrence for you

I think it would be nice if we could have a more explicit typing to make it super clear from the perspective of building a detector "In THIS case I want to create issues, in THIS OTHER case I do not"