rds_postgres

Synthetic RDS PostgreSQL Suite

This suite benchmarks RDS PostgreSQL root-cause analysis against bundled telemetry fixtures instead of live AWS infrastructure. Each scenario is a static evidence snapshot served through a FixtureGrafanaBackend, which drives the same agentic pipeline (plan → investigate → diagnose) used in production.

Scenario table

Axis 1 — Efficiency (marker: synthetic)

ID	Name	Difficulty	True root cause	Adversarial element	Forbidden
000	healthy	1	healthy	none	resource_exhaustion
001	replication-lag	1	replication_lag	none	—
002	connection-exhaustion	1	connection_exhaustion	none	—
003	storage-full	1	storage_exhaustion	none	—
004	cpu-saturation-bad-query	1	cpu_saturation	none	—
005	failover	1	multi_az_failover_health_check	none	—
006	replication-lag-cpu-redherring	2	replication_lag	CPUUtilization elevated (analytics job)	category: cpu_saturation
007	connection-pressure-noisy-healthy	2	healthy	CPU/connections oscillating near-threshold	category: connection_exhaustion
008	storage-full-missing-metric	3	storage_exhaustion	FreeStorageSpace absent from fixture	—
009	dual-fault-connection-cpu	4	connection_exhaustion	connections + CPU both failing, causally linked	keywords: storage, replication
010	replication-lag-missing-metric	3	replication_lag	ReplicaLag metric absent from fixture	—

Axis 2 — Reasoning (marker: axis2)

Axis 2 scenarios use SelectiveGrafanaBackend. The agent must ask for the right metrics and explicitly rule out alternative hypotheses. Each scenario has ruling_out_keywords that must appear in the agent's output.

ID	Name	Difficulty	True root cause	Red herring / adversarial element	Must rule out
011	cpu-storage-compositional	4	dual_resource_exhaustion	ReplicaLag elevation and connection spike as side effects	connection_exhaustion, replication
012	replication-lag-misleading-events	3	replication_lag	Three historical infra events in event stream (none are root cause)	infrastructure (failover as root cause)
013	storage-recovery-false-alert	3	healthy	FreeStorageSpace spike + WriteLatency brief elevation	resource_exhaustion (already recovered)
014	checkpoint-storm-cpu-saturation	4	checkpoint_io_storm	CPU at 92% — alert fires on CPU but cause is VACUUM checkpoint storm	cpu_saturation (CPU is symptom, not root)

Difficulty levels

Level	Description
1	Single dominant signal — all evidence consistent, root cause identifiable in one step
2	One confounder present — second evidence source needed to rule it out
3	Absent or indirect evidence — key metric missing, or misleading signals require timeline reasoning
4	Compositional fault — two failure modes active; agent must identify both and correctly characterise causally-linked side effects

MECE basis

Uniqueness is on (primary_signal × rate × corroborating_presence × event_presence), not on primary signal alone.

003 and 008 both map to storage_full but have distinct fingerprints:

• 003: FreeStorageSpace present and trending to 0 with elevated WriteIOPS
• 008: FreeStorageSpace absent from the fixture entirely — agent must infer from events + PI write latency

Scoring

Axis 1 (all scenarios — test_suite.py)

Each scenario passes when all of the following are true:

1. The model returns a non-empty root cause
2. The predicted ROOT_CAUSE_CATEGORY matches answer.yml
3. Every required keyword from answer.yml:required_keywords appears in the output
4. The actual category is not in answer.yml:forbidden_categories (level 2+ scenarios)
5. No forbidden keyword from answer.yml:forbidden_keywords appears in the output (level 4 scenario)
6. Every source listed in answer.yml:required_evidence_sources is non-empty in final_state["evidence"] — proves the agent consulted the right evidence, not just keyword-matched the alert title

Additionally, a TrajectoryScore is computed for each scenario with an optimal_trajectory:

• sequencing_ok: all expected action types appear in the agent's executed action log (set membership; parallel execution order is non-deterministic)
• calibration_ok: number of investigation loops ≤ max_investigation_loops
• efficiency_score: mean(sequencing_ok, calibration_ok); 1.0 = full pass

Axis 2 (adversarial scenarios — test_suite_axis2.py, marker: axis2)

Axis 2 runs through SelectiveGrafanaBackend, which:

• Returns only the metric series matching the agent's metric_name query (case-insensitive substring)
• Records every metric name the agent requested in queried_metrics (audit log)

On top of all Axis 1 checks, Axis 2 asserts a ReasoningScore:

• ruling_out_ok: every token in answer.yml:ruling_out_keywords appears in agent output (proves the agent addressed and dismissed alternative hypotheses)
• queries_ok: every entry in answer.yml:required_queries was requested by the agent via query_timeseries — not yet testable: the current action registry hardcodes metric_name="pipeline_runs_total" for all query_grafana_metrics calls; required_queries is reserved for when the agent supports per-metric querying
• reasoning_score: mean(ruling_out_ok, queries_ok); 1.0 = full pass

Gap metric

The gap between Axis 1 and Axis 2 pass rates is the primary health indicator for adversarial robustness. A large gap means the agent can find answers when handed all data but cannot reason when it must choose what to look at. Track it per difficulty level:

python -m tests.synthetic.rds_postgres.run_suite --mock-grafana --axis2

Each scenario folder contains

• scenario.yml: scenario metadata (engine, difficulty, adversarial_signals, depends_on)
• alert.json: synthetic alert payload
• aws_cloudwatch_metrics.json: CloudWatch metric evidence (may omit metrics to simulate collection gaps)
• aws_rds_events.json: RDS event stream for the incident window
• aws_performance_insights.json: top SQL and wait-event evidence
• answer.yml: expected category, required keywords, optional forbidden constraints, required evidence sources

Running

Via the interactive CLI (recommended):

opensre tests synthetic

Run the whole Axis 1 suite directly:

python -m tests.synthetic.rds_postgres.run_suite --mock-grafana

Run Axis 2 adversarial scenarios via pytest:

pytest -m axis2 tests/synthetic/rds_postgres/test_suite_axis2.py -v

Run a single scenario:

python -m tests.synthetic.rds_postgres.run_suite --scenario 006-replication-lag-cpu-redherring --mock-grafana

Print JSON results:

python -m tests.synthetic.rds_postgres.run_suite --mock-grafana --json

CI tier strategy

• Axis 1, Levels 1–2 (scenarios 000–007): run on every commit (@synthetic)
• Axis 1, Levels 3–4 (scenarios 008–010): deferred to nightly — require indirect inference
• Axis 2 (scenarios 011–013, @axis2): run nightly alongside Axis 1 levels 3–4; gap is reported each run

Known gaps

• Temporal ordering: all scenarios deliver evidence as a static snapshot. Production delivers evidence incrementally (alert fires → query metrics → query events → …). Testing temporal ordering requires architectural changes to the fixture backend and is out of scope.
• Level 4 coverage: two compositional fault scenarios (009, 011). A fuller curriculum would include 3–4 dual-fault combinations across different failure mode pairs.
• Slack/markdown renderer for multi-fault: the renderer displays a single root_cause string. Compositional faults may eventually need a root_causes: list field in the schema.
• Axis 2 required_queries not yet enforced: The agent action registry hardcodes metric_name="pipeline_runs_total" for all query_grafana_metrics calls. SelectiveGrafanaBackend records all queried metric names as an audit log but does not filter results. When the agent is updated to pass dynamic CloudWatch metric names (e.g. metric_name="CPUUtilization"), re-enable filtering in SelectiveGrafanaBackend.query_timeseries and set required_queries in answer.yml.

Dependency: healthy_rca_state

Scenario 007 depends on HEALTHY_SHORT_CIRCUIT=true (the default) and the healthy category being wired into the LLM prompt. If you run with HEALTHY_SHORT_CIRCUIT=false, scenario 007 will fall through to the LLM path, which should still classify as healthy — but the test is most deterministic with the short-circuit enabled.