Hatchet Hatchet Versions Save

New features

• Reader for the new HPCToolkit v4 data format
• Several new capabilities in the Chopper API
• Docs for the Chopper API
• Function to unify multiple graphframes

Other changes

• Drop Python 2.7 support
• Remove cython generated .c files

Support for new file formats:

• Timemory reader
• Perfetto/Omnitrace reader
• HDF5 reader and writer
• TAU reader
• SpotDB reader
• ScoreP reader

New features:

• Query dataframe columns via GraphFrame.show_metric_columns()
• Query nodes within a range using the call path query language
• Extend readers to define their own default metric
• Console tree() output can displays two metrics
• Parallel implementation of filter function
• New compound queries feature for Query language
• Chopper API for simplifying performance analysis

New features

• Timemory reader
• Query dataframe columns with GraphFrame.show_metric_columns()
• Query nodes within a range using the call path query language
• Extend readers to define their own default metric
• Changes to existing APIs
• Tree visualization displays 2 metrics
• Literal output format: add hatchet node IDs
• Parallel implementation of filter function
• Caliper reader: support multiple hierarchies in JSON format
• Adds multiprocessing dependency

New features:

• Interactive tree visualization in Jupyter
• Add mult and division API
• Update hatchet installation steps for cython integration
• Readers: cprofiler, pyinstrument
• Graph output formats: to_literal
• Add profiling APIs to profile Hatchet APIs
• Update basic tutorial for hatchet

Changes to existing APIs

• Remove threshold=, color=, and unicode= from tree API
• Highlighting name disabled by default in terminal tree output is kept in sync with the dataframe
• Internal performance improvements to unify and HPCToolkit reader, enabling analysis of large datasets
• For mathematical operations, insert nan values for missing nodes, show values as nan and inf as necessary in dataframe
• Extend callpath query language to support non-dataframe metrics (e.g., depth, hatchet ID)
• Literal reader: A node can be defined with a "duplicate": True field if it should be the same node (though in a different callpath). A node also needs "frame" field, which is a dict containing the node "name" and "type" (if necessary).

• Add graph syntax query language to filter the graph
• Update HPCToolkit reader to handle sec or usec time units
• Sort nodes in union and tree printout by their frame
• Fix squash edge case where multiple parents are the same
• Add squash parameter to filter function to perform filter and squash in a single call
• Filter function takes an object, which can be a user-supplied function or a query object
• Update tree printout format
• Parameter changes to tree printout API, alias old to new parameters

• Add GraphFrame reindex operator
• Query hatchet module version
• Add depth parameter to tree printout
• Handle MultiIndex for pandas 1.0.0 and newer vs older pandas versions

• Downgrade pandas version for python 3.6 and later (incompatible versions)
• Fix HPCToolkit reader bug for statement nodes
• Add markups to identify missing nodes (from unify)
• Remove sorting nodes by frame in graph union
• Add missing rows for math operations on graphframes, fixes unify
• Add precision parameter to metrics in tree printout
• Restore index for self's dataframe in squash
• tree() to print nodes with negative values higher than threshold
• Fix phase timer to aggregate times for duplicate phases
• Add Graphframe division operator

This is the first major release of Hatchet. Documentation is available at: https://hatchet.readthedocs.io.