feat: add metrics_aggregator service for real-time metric computation

[nv-alicheng]

• MetricsAggregator: subscribes to ZMQ events, computes QPS/latency/TTFT/TPOT
• MetricsTable: columnar storage for efficient percentile calculations
• TokenMetrics: ISL/OSL token-level metrics from PromptData
• MetricsEmitter: periodic metric publishing
• Service entry point with CLI interface
• Comprehensive unit and e2e tests

What does this PR do?

Type of change

• Bug fix
• New feature
• Documentation update
• Refactor/cleanup

Related issues

Testing

• Tests added/updated
• All tests pass locally
• Manual testing completed

Checklist

• Code follows project style
• Pre-commit hooks pass
• Documentation updated (if needed)

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a new metrics_aggregator service designed to provide real-time performance insights for inference requests. By subscribing to event streams, the service calculates critical metrics such as time-to-first-token, end-to-end latency, request duration, and token-level performance (input/output token counts and time-per-output-token). This enhancement allows for detailed monitoring and analysis of the inference endpoint's performance characteristics.

Highlights

• Metrics Aggregation Service: Introduced MetricsAggregatorService to subscribe to ZMQ events and compute real-time metrics like QPS, latency, TTFT, and TPOT.
• Data Storage and Calculation: Implemented MetricsTable for efficient in-memory storage and percentile calculations of per-sample metric data.
• Token-Level Metrics: Added TokenizePool to handle token-level metrics (ISL/OSL/TPOT) using HuggingFace tokenizers in a thread-safe manner.
• Metric Emission: Developed JsonlMetricEmitter for periodic publishing of computed metrics to a JSONL file.
• Command-Line Interface: Provided a command-line interface (CLI) entry point for the new metrics aggregator service.
• Testing: Included comprehensive unit and end-to-end tests to ensure the correctness and reliability of the new service.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces a new metrics_aggregator service for real-time metric computation from ZMQ events. The implementation is well-structured, with clear separation of concerns for event processing, metric emission, and tokenization. The accompanying unit and end-to-end tests are comprehensive and cover many edge cases, ensuring the reliability of the new service.

My review includes a few suggestions to improve maintainability and performance. These include simplifying resource management using context managers, improving typing for better static analysis, and a design suggestion to potentially increase event processing throughput by offloading tokenization to a background task.

[nvzhihanj]

Review Council — Multi-AI Code Review

Found 3 issues across 3 files.

[nvzhihanj]

Review Council — Multi-AI Code Review Council

Reviewed by: Claude
(Codex identified a potential batch-shutdown bug but couldn't produce structured output due to exec-mode sandbox limitations)

Found 3 issues across 3 files:

• 2 high
• 1 medium

Each issue is posted as an inline comment on the relevant file and line.

#	File	Line	Severity	Reviewer(s)	Summary
1	src/.../metrics_aggregator/emitter.py	73	high	Claude + Codex*	emit() missing None guard — will crash if ISL task fires after close() sets _file = None
2	src/.../metrics_aggregator/aggregator.py	222	medium	Claude + Codex*	_schedule_isl fire-and-forget tasks not awaited during _finalize(), racing against emitter close()
3	pyproject.toml	55	high	Claude	sqlalchemy added as core dependency but only used by optional SQL logger — should be in optional-dependencies

*Codex independently identified "a likely batch-shutdown bug in the aggregator" before sandbox restrictions prevented it from producing structured output.

🤖 Generated with Claude Code

[nvzhihanj]

Review Council — Multi-AI Code Review

Found 10 issues (2 high, 6 medium, 2 low).

[nvzhihanj]

Review Council — Multi-AI Code Review

Found 10 issues (2 high, 6 medium, 2 low).

[nvzhihanj]

Review Council — Multi-AI Code Review Council

Reviewed by: Claude (Codex timed out) | Depth: thorough

Found 10 issues across 8 files.

🔴 Must Fix (high)

Issues that will cause incorrect behavior or crashes in production.

#	File	Line	Category	Summary
1	emitter.py	73	bug	JsonlMetricEmitter.emit() missing None guard — crashes after close()
2	aggregator.py	155	concurrency	Fire-and-forget ISL tasks crash on closed emitter; orphaned tasks never cancelled

🟡 Should Fix (medium)

Real issues under specific conditions or design flaws that will compound.

#	File	Line	Category	Summary
3	types.py	84	data-integrity	TextModelOutput.__str__ leaks reasoning into OSL/TPOT metrics
4	test_aggregator.py	30	testing	Missing mode="strict" on all async test markers (project convention)
5	pyproject.toml	55	design	SQLAlchemy added as core dependency but only used by optional SQL writer
6	token_metrics.py	33	concurrency	Thread-local tokenizer cache not keyed by name — wrong tokenizer if reused
7	context.py	82	api-contract	Singleton ignores socket_dir on second init — silent misconfiguration
8	__main__.py	103	bug	call_soon_threadsafe used from same thread — should be call_soon

🔵 Consider (low)

Valid improvements that could be follow-ups.

#	File	Line	Category	Summary
9	metrics_table.py	40	performance	output_chunks list grows unboundedly per in-flight sample
10	types.py	84	bug	Empty string reasoning silently dropped due to truthiness check

Each issue is posted as an inline comment on the relevant file and line.

🤖 Generated with Claude Code