Uploaded July 2026 | Updated September 2026, 3 weeks ago
Traditional monitoring reported the system healthy: latency down, errors at zero. A mandatory processing step had been silently skipped, and nothing caught it except the graph. Ritvik Pandya's team at JP Morgan models each API request as a short lived execution graph, a DAG of the middleware steps it passes through, learned from telemetry at over 1,600 requests per second. Compare what actually ran against that learned graph and a skipped, reordered, or injected step stops hiding behind healthy averages.
The same graph localizes performance problems to the exact node instead of the whole endpoint. In production it flagged a 41x deviation at a single node that service level monitoring never saw, cutting root cause from hours to under 30 seconds. The talk separates a one off anomaly from real drift, a slow shift that needs a new baseline, and sorts drift into structural, volume, and behavioral, using per node baselines and KL divergence rather than one threshold for every request. The payoff is a cheap tier one check that only escalates when the graph says something actually changed.
Speaker info:
- linkedin.com/in/ritvik-pandya
Timestamps:
0:00 - Execution graphs for anomaly and drift detection
1:07 - What a short lived execution graph is
3:28 - Tiered checks and per client baselines
5:23 - The method: baseline, deviation, localize, act
6:16 - Localizing a slow node, and how the system is trained
7:33 - Anomaly versus drift
8:55 - The three kinds of drift: structural, volume, covariate
12:46 - The pipeline: from telemetry to gradual rollout
13:54 - Hot path versus recon, and worked examples
15:21 - Tuning it: delayed events, sampling, cold starts
17:09 - Results and lessons
Traditional monitoring reported the system healthy: latency down, errors at zero. A mandatory processing step had been silently skipped, and nothing caught it except the graph. Ritvik Pandya's team at JP Morgan models each API request as a short lived execution graph, a DAG of the middleware steps it passes through, learned from telemetry at over 1,600 requests per second. Compare what actually ran against that learned graph and a skipped, reordered, or injected step stops hiding behind healthy averages.
The same graph localizes performance problems to the exact node instead of the whole endpoint. In production it flagged a 41x deviation at a single node that service level monitoring never saw, cutting root cause from hours to under 30 seconds. The talk separates a one off anomaly from real drift, a slow shift that needs a new baseline, and sorts drift into structural, volume, and behavioral, using per node baselines and KL divergence rather than one threshold for every request. The payoff is a cheap tier one check that only escalates when the graph says something actually changed.
Speaker info:
- linkedin.com/in/ritvik-pandya
Timestamps:
0:00 - Execution graphs for anomaly and drift detection
1:07 - What a short lived execution graph is
3:28 - Tiered checks and per client baselines
5:23 - The method: baseline, deviation, localize, act
6:16 - Localizing a slow node, and how the system is trained
7:33 - Anomaly versus drift
8:55 - The three kinds of drift: structural, volume, covariate
12:46 - The pipeline: from telemetry to gradual rollout
13:54 - Hot path versus recon, and worked examples
15:21 - Tuning it: delayed events, sampling, cold starts
17:09 - Results and lessons
