Blackfire Continuous Profiler
Low-overhead continuous profiling for app performance optimization.
VizTracer
Low-overhead Python tracer with interactive Perfetto visualizations
VizTracer records function entry/exit, threading, async, and PyTorch events with minimal impact, then visualizes traces in a Perfetto‑based UI that handles gigabytes of data smoothly.
Overview
Overview
VizTracer is a lightweight tracing tool for Python that captures detailed function entry and exit information, source code locations, and runtime events such as threading, multiprocessing, async coroutines, and PyTorch GPU activity. It requires little to no code changes—simply run your script with viztracer or use the provided context manager—and produces a JSON trace that can be opened in the built‑in vizviewer UI.
How it works
The trace data is rendered by a Perfetto‑powered front‑end, offering smooth navigation, zooming (AWSD), flame‑graph generation, and the ability to handle GB‑scale files. Advanced features include customizable filters, extra variable logging, and custom events, while optional orjson support speeds up JSON serialization. VizTracer runs on Linux, macOS, and Windows, and integrates with remote attach, making it suitable for a wide range of development and debugging workflows.
Highlights
Detailed timeline with source code and function arguments
Zero‑code‑change usage via CLI or context manager
Supports threading, multiprocessing, async, and PyTorch profiling
Perfetto UI renders gigabyte‑scale traces smoothly
Pros
- Minimal runtime overhead for typical workloads
- Cross‑platform support (Linux, macOS, Windows)
- Rich, interactive visualization without external dependencies
- Flexible filtering and custom event insertion
Considerations
- Overhead can increase for highly recursive functions
- Very large traces may require the external processor flag
- Full feature set (e.g., sys.monitoring) needs Python 3.12+
- Visualization runs a local HTTP server, which may be restricted
Managed products teams compare with
When teams consider VizTracer, these hosted platforms usually appear on the same shortlist.
Looking for a hosted option? These are the services engineering teams benchmark against before choosing open source.
Fit guide
Great for
- Developers needing quick insight into performance bottlenecks
- Teams debugging multi‑threaded or async Python applications
- Data scientists profiling PyTorch training loops
- Engineers who prefer CLI‑driven tracing without extra packages
Not ideal when
- Scenarios demanding sub‑microsecond profiling precision
- Environments where running a local web server is prohibited
- Projects with extremely deep recursion where overhead spikes
- Users requiring out‑of‑the‑box HTML reports without a viewer
How teams use it
Debugging a Flask web service
Identify slow request handlers and middleware latency
Profiling a PyTorch training loop
Visualize GPU kernel timings and Python‑level function calls
Analyzing an async data pipeline
See coroutine scheduling, I/O waits, and event ordering
Tracing a multi‑process ETL job
Correlate subprocess activities across workers in a single view
Tech snapshot
Python65%
C27%
C++8%
Makefile1%
Frequently asked questions
How do I install VizTracer?
Run `pip install viztracer` to install the package from PyPI.
Do I need to modify my code to start tracing?
No. You can invoke `viztracer` from the command line or use the `VizTracer` context manager; most features work without code changes.
Can VizTracer handle large trace files?
Yes, the Perfetto UI can render gigabyte‑scale traces, and the `--use_external_processor` flag helps with very large files.
Is there support for profiling PyTorch GPU events?
Enable with `--log_torch` or `log_torch=True`; VizTracer integrates with `torch.profiler` to capture native calls and GPU timings.
What platforms are supported?
VizTracer runs on Linux, macOS, and Windows.
Project at a glance
Active- Stars
- 7,516
- Watchers
- 7,516
- Forks
- 471
LicenseApache-2.0
Repo age5 years old
Last commit4 days ago
Primary languagePython
Last synced 2 days ago