Slides: Using RL Agent to Detect and Remediate ETL Pipeline Failures - Anna Marie Benzon
Source Video
Using RL Agent to Detect and Remediate ETL Pipeline Failures - Anna Marie Benzon
Relationship To World's Fair 2026
These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material.
Related Scheduled Sessions
- No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck.
Extracted Slides
OCR text:
Using RL-based Agent to
Detect and Remediate aan
ETL Pipeline Failures an
=
Ea bad aad
Presenter:
: Anna Marie Benzon
University of the Philippines, Diliman
github.com/ambenzon27/rl-etl-remediation-agent
OCR text:
: PS ~ ED ~ EE - A - - E
. Ca Late or unavailable source data
Th . Schema drift
e -
GN Datetime parsing incompatibilities
Problem *
®) Null-rate spikes
Cloud ETL jobs break because of:
tes ,
4d, Type changes
: =,» . = Unknown runtime errors
7 a Modeled manual MTTR ~ 2.5 working days
OCR text:
Phave 1 Architecture ~ Al Pipeline Health Agent
An End-to-End RL Pipeline
Health Agent for Anomaly
| —E ia Detection and Autonomous
apa Self-Healing of Cloud Data
= reais
SCTE GENERA AES HERERERERARRARETUICETAE
Monitor + Diagnose — Score > Decide —> Safety > Act — Validate
l.Observe logs, schema, and data-quality conditions
2.Diagnose the likely failure family
e 3.Estimate operational risk
4.Select a bounded remediation action
5.Validate whether the action restored a healthy state
OCR text:
Deterministic Anomaly Rules
Schema drift, null spikes, field
removals, type changes
Q-Learning Decision Policy
Retry, coerce schema, rollback,
quarantine, escalate, log
Safety Override e Th :
Critical anomaly + passive action
ver intelligence
OCR text:
: Detection and Diagnosis
Establish the Facts Before Choosing an Action
Schema profiler Structure, types. nesting. null rates
Drift detector Additions. removals. type changes
Data-quality analyzer Completeness, validity, consistency
‘Converts log patterns into failure categories
Deterministic prototype with limited dataset; ML-ready with richer incident history.
OCR text:
RL Selects the Response, Not the Facts
Action:
State:
1.Failure category
2.Risk level 1.Tabular Q-learning
3.Retry count 2.Small, interpretable state space
4.Drift severity 3.Low-memory inference
5.Data-quality condition 4.Inspectable Q-values for every decision
“TECHNICALLY, THIS IS A SINGLE-STEP CONTEXTUAL DECISION PROBLEM IMPLEMENTED WITH TABULAR O-LEARNING.
OCR text:
: Safe Autonomy
The Learned Policy Does Not Have Final Authority
1.Q-learning policy proposes an action
2.Safety layer evaluates critical anomalies .
3.Unsafe passive actions are overridden
4.High-risk and unknown conditions escalate
5.Every decision produces an audit record Ez
a \
4 = Escalation is a correct action, not a failure of autonomy.
OCR text:
| Example Failure
* wjob_names =, synthetic etLjob «,
“triggered_at": “2026-05-21702:05:202", = {A} Real Glue job failure
"error_classification": {
“error_type": “DATETIME FORMAL ERROR”. + {B] Classified by Error Classifier
"confidence": 6.98,
. “root_cause": “Spark 3 datetime format incompatibility”,
“recommended_action”: “APPLY_SCHEMA_COERCION” - [C] Rule engine says: coerce
Sefema/date format
“selected_action": “APPLY_SCHEMA_COERCION”, - [D] Agent selects schema coercion
"“anomaly_override’: false, ~ {E] Safety override did not fire .
remediation": {
“success”: false,
) note": “Manual Glue script update required” - {F] Logged for review
}
Observed condition Decision
1.Glue-style job failure event 1.Policy selects schema coercion
received 2.Safety override does not fire
. 2.Datetime-format incompatibility 3.Automatic coercion is unavailable
detected 4.Incident is logged for manual review
3.Error classified with 0.90 ;
confidence
OCR text:
| Reproducible Evaluation
Designed for Independent Reproduction
ambenzon27/rl-etl-
1.Generalized AWS Lambda-style remediation-agent a
architecture Sn eae
2.Synthetic schemas, records, logs, and Seg, 6
incidents ——EEEEEE
3.No production data or infrastructure reeercrnenceecacngert gated ET ran ser
identifiers Se reese erces erm
4.Four controlled experiments: E1-E4 Snares
5.Robustness check across 30 runs, seeds
42-71
6.Results reported with 95% confidence C> Public benchmark and tests are
‘intervals ‘ ; available in the GitHub
repository. &
OCR text:
| Evaluation Results
Controlled synthetic benchmark MTTR Drops from Days to Minutes in the Benchmark
— 30 seeds, mean # 95% Cl . a
. 2.5 working days
1.Rule-based anomaly detector: a
Precision 1.000 “ 99.85% lower MTTR
2.Recall: 0.800 i
3.F1: 0.889 EF
4.RL successful-case resolution time: sztcnoraohe
approximately 5.2 minutes
5.RL simulated success rate: 74.63 + 1.51% il
6.RL non-escalation rate: 88.63 + 0.89% uaa ””””S:CRC age
OCR text:
| Robustness and Ablation
What produced reliability?
1.RL success matched the , ,
deterministic policy: 0.00 + 0.19 30 synthetic seeds; mean + 95% Cl
percentage-point difference -
2.Deterministic rules beat random -15.03 + 0.66 pp
selection by 15.63 + 1.86 points Safety override jel
3.The safety override reduced vs hone +0.00 + 0.19 pp
non-escalation by 15.03 + 0.66 RL vs rules ®
4 Dawa ‘ded oni tab +15.63 + 1.86 pp
.RL provided an inspectable
learned policy, but did not Rules vs random Fe
outperform rules in this -15 -10 -S5 0 5 10. 15
benchmark Difference (percentage points)
5.Structured decision logic and
external guardrails produced at
most of the reliability ti
OCR text:
What This Prototype Does Not Yet Prove
1.Results are based on synthetic benchmark scenarios
2.The agent is failure-triggered, not a pre-failure predictor
- 3.Production incident diversity may exceed the current state
space
4.Some remediation actions remain simulated or bounded
5.Online learning requires strict operational approval gates
Current This demonstrates feasibility and system
Limitations design, not production completeness.
OCR text:
Small Al Can Still Deliver Operational Value &)
1.Use deterministic logic for observable facts © F. ae 6)
2.Use RL where contextual action selection adds parent ane
value secs eter "Ein, if
3 esol ove
3.Place safety constraints outside the learned policy ii i ae :
4.Treat escalation and validation as core pet apis
capabilities ore Sets
5.Evaluate across repeated seeds, not one favorable (s) ot 2 nh
run
| A practical self-healing pipeline does not
T keaw need a giant model. It needs bounded
Ci ays authority, reproducible evidence, and the
discipline to escalate when it is
uncertain.
OCR text:
it
e Wi
From reactive bar @
e
debugging to Before:
e e \ ¢ Manual log inspection
intelligentrecovery! —~ schematracing
The goal is not to replace * Delayed dashboards _
‘ ¢ Modeled MTTR: 2.5 working days
human judgment.
O af? 3 aa éA
It is to reserveds for the failures as
that truly need ite“Hr i: ¢ Event-triggered diagnosis
Routine ETL failufes. nach erlogy Edie, e RL-guided remediation
explainable, and ecoverabiein' nutes. ¢ Safety-constrained escalation
. ae oy} ¢ Minutes-scale recovery
4
Slide-Derived Subjects To Review
Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.