Quality assurance · Production

PerfInsights: Uber's GenAI system detects Go performance antipatterns and cuts optimization engineering time by 93%

The problem

Uber's top 10 Go services cost more than multi-million dollars in compute in March 2024 alone, yet optimizing Go services required deep expertise and days or weeks of manual profiling and analysis—making systematic performance tuning prohibitively expensive for most teams.

First attempt

Initial single-shot LLM-based antipattern detection produced inconsistent and unreliable results—responses varied between runs, included hallucinations, and often generated non-runnable code, with false positives exceeding 80%.

Workflow diagram · grounded in source
1
Production profile collection
trigger
“PerfInsights leverages CPU and memory profiles from production services using Uber's daily fleet-wide profiler during peak traffic periods”
2
Top-30 function selection
validation
“it identifies the top 30 most expensive functions based on flat CPU usage”
3
Static noise filtering
validation
“PerfInsights applies a static filter that excludes open-source dependencies and internal runtime functions. This step trims noise from the candidate set”
4
LLM antipattern detection
ai_action
“PerfInsights passes their full source codes and a list of antipatterns to a large language model (LLM) for analysis. By combining profiling context with pattern awareness, the model can pinpoint inefficient constructs with high precision…”
5
LLM jury validation
validation
“PerfInsights leverages a jury of large language models to validate each detected antipattern. These models independently assess whether an antipattern is present and whether the suggested optimization is valid”
6
LLMCheck rule-based verification
validation
“PerfInsights employs a second layer of verification via LLMCheck by running through several domain-specific rule-based validators to evaluate LLM responses”
7
Output to Optix and developers
output
“validated suggestions flowing directly into Optix, Uber's continuous code optimization tool”
8
Accuracy metrics and drift detection
feedback_loop
“LLMCheck also logs metrics on detection accuracy, tracking failure rates and signaling potential model drift”
Reported outcome

PerfInsights reduced engineering time per performance issue from 14.5 hours to almost 1 hour—a 93.10% time savings—and cut false positives from over 80% to the low teens, with hundreds of diffs already merged into Uber's Go monorepo.

Reported metrics
Go services compute spend (top 10, March 2024)multi-million dollars
Task durationdays now take hours
False positive rate (before)over 80%
False positive rate (after)low teens
Show all 17 reported metrics
Go services compute spend (top 10, March 2024)multi-million dollars
task durationdays now take hours
false positive rate (before)over 80%
false positive rate (after)low teens
validated detections per day average (February)265
validated detections single-day peak500
validated detections per day (June)176
reduction in validated detections over 4 months33.5%
manual engineering hours for sample antipattern setover 1,400 hours
projected manual hours to address annual antipattern volumeapproximately 3,800 hours
engineering time savings per issue93.10%
engineering time per issue (before)14.5 hours
engineering time per issue (after)almost 1 hour
hallucination rate reductionmore than 80%
merged optimization diffshundreds
false positive reduction via prompt engineering80%
cost saving milestone per optimization diff5 digits saving
Reported stack
PerfInsightsLLMLLMCheckOptix
Source
https://www.uber.com/en-GB/blog/perfinsights/?uclick_id=0a73d271-32e7-4b77-9697-a587a4c8d9fe
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

PerfInsights reduced engineering time per performance issue from 14.5 hours to almost 1 hour—a 93.10% time savings—and cut false positives from over 80% to the low teens, with hundreds of diffs already merged into Ube…

What tools did this team use?

PerfInsights, LLM, LLMCheck, Optix.

What results were reported?

Go services compute spend (top 10, March 2024): multi-million dollars; Task duration: days now take hours; False positive rate (before): over 80%; False positive rate (after): low teens (source-reported, not independently verified).

What failed first in this deployment?

Initial single-shot LLM-based antipattern detection produced inconsistent and unreliable results—responses varied between runs, included hallucinations, and often generated non-runnable code, with false positives exce…

How is this quality assurance AI workflow structured?

Production profile collection → Top-30 function selection → Static noise filtering → LLM antipattern detection → LLM jury validation → LLMCheck rule-based verification → Output to Optix and developers → Accuracy metrics and drift detection.

Related quality assurance cases