Streamline analysis pipeline and logging

[izzet]

This pull request refactors the percentile/threshold bottleneck detection logic out of the core analysis pipeline, simplifies the configuration interface, and introduces structured logging and warning suppression for a cleaner user and developer experience. It also improves the codebase by organizing logging, warning, and checkpointing utilities, and enhances the workflow and documentation to match the updated interface.

Configuration & CLI Interface Simplification

• Removes percentile and threshold arguments from the main analysis pipeline, CLI, and configuration, so bottleneck detection is no longer handled as a core parameter. This change is reflected in dfanalyzer/__init__.py, dfanalyzer/analyzer.py, .github/workflows/ci.yml, and README.md. [1] [2] [3] [4] [5] [6] [7]

Logging & Warning Handling Improvements

• Introduces structured logging using structlog and context managers (console_block, log_block) for clearer, stepwise logging of major analysis phases. Logging is now configured on both the main process and all Dask workers. [1] [2] [3] [4] [5] [6] [7]
• Centralizes and improves warning suppression by moving warning filtering into a utility function and removing direct warnings usage from the main codebase. [1] [2] [3] [4]

Analysis Pipeline & Checkpointing Enhancements

• Refactors the analysis pipeline to add explicit logging blocks for key operations (reading traces, computing metrics, processing views, and checkpointing), making the workflow more transparent and easier to debug. [1] [2] [3] [4]
• Improves checkpointing by collecting Dask tasks for saving flat views and waiting for their completion in a dedicated step, which enhances reliability and performance. [1] [2]

Documentation & Workflow Updates

• Updates the example usage in README.md and the CI workflow to remove references to the deprecated percentile parameter, ensuring consistency with the new interface. [1] [2]

Other Minor Improvements

• Changes the default time_granularity unit from microseconds to seconds for consistency and clarity. [1] [2]

These changes collectively make the codebase more maintainable, improve user experience, and lay the groundwork for more flexible bottleneck detection strategies in the future.

[izzet]

The new version runs ~2.25× faster with much lower runtime variance:

hyperfine '/.../dfanalyzer/.../python -m dfanalyzer analyzer=dftracer analyzer.checkpoint=False analyzer/preset=dlio percentile=0.9 trace_path=/.../data/extracted/dftracer-dlio' '/.../dfanalyzer-fix-demo/.../python -m dfanalyzer analyzer.checkpoint=False analyzer/preset=dlio trace_path=/.../data/extra
cted/dftracer-dlio'
Benchmark 1: /.../dfanalyzer/.../python -m dfanalyzer analyzer=dftracer analyzer.checkpoint=False analyzer/preset=dlio percentile=0.9 trace_path=/.../data/extracted/dftracer-dlio
  Time (mean ± σ):     122.771 s ±  3.610 s    [User: 151.342 s, System: 16.662 s]
  Range (min … max):   117.097 s … 129.307 s    10 runs
 
Benchmark 2: /.../dfanalyzer-fix-demo/.../python -m dfanalyzer analyzer.checkpoint=False analyzer/preset=dlio trace_path=/.../data/extracted/dftracer-dlio
  Time (mean ± σ):     54.466 s ±  1.447 s    [User: 78.171 s, System: 12.346 s]
  Range (min … max):   53.229 s … 58.322 s    10 runs
 
Summary
  '/.../dfanalyzer-fix-demo/.venv/bin/python -m dfanalyzer analyzer.checkpoint=False analyzer/preset=dlio trace_path=/.../data/extracted/dftracer-dlio' ran
    2.25 ± 0.09 times faster than '/.../dfanalyzer/.../python -m dfanalyzer analyzer=dftracer analyzer.checkpoint=False analyzer/preset=dlio percentile=0.9 trace_path=/.../data/extracted/dftracer-dlio'

[rayandrew]

Looks good to me, thank you so much @izzet !