Compared with Chadwick

chadwickpy is a function-by-function translation of Chadwick's C source. The rule is "translate the C, never guess from output", so the output matches wherever the C behaves in a defined way (see where it differs).

	Chadwick (C)	chadwickpy
Install	build from source with a C compiler, or a system package	pip install chadwickpy
Needs	compiler, autotools	Python 3.11+ (zero dependencies)
Output	the reference	identical on every season with game data, 1897-2025, for all six tools, for inputs where the C has defined behaviour
Speed (single file)	cwevent on one team-season: ~0.07 s	~0.66 s (pure Python)
Speed (full season)	cwevent 2,430 games: ~2.25 s (1 core)	~2.34 s (auto-parallelized across cores)
Throughput	~83,000 plays/s	~80,000 plays/s
Commands	cwevent, cwgame, cwdaily, cwsub, cwcomment, cwbox	the same names and options (plus -j for concurrency control)
Use from Python	via other wrappers	import chadwickpy

Benchmarks & Performance

The benchmark below compares original Chadwick C (0.10.0, measured before 0.11.0 was released) against chadwickpy on the complete 2024 MLB Retrosheet season (30 team event files, 2,430 games, 188,937 plays):

Configuration	Total Time	Plays / Second	Output Match
Chadwick (C) (single core)	2.25 s	~83,900 plays/s	Reference (26.8 MB)
chadwickpy (sequential, -j 1)	20.04 s	~9,400 plays/s	100% byte-for-byte identical
chadwickpy (default auto-parallel)	2.34 s	~80,700 plays/s	100% byte-for-byte identical

One core, every tool

Measured on one core (-j 1) against the C tools, whole 2025 season, default output, on a busy 40-core machine, so treat the ratios as approximate:

Tool	C	chadwickpy -j 1	Slower by
cwevent	2.4 s	27.2 s	about 11x
cwgame	1.3 s	40.3 s	about 31x
cwdaily	2.3 s	38.3 s	about 17x
cwsub	0.6 s	14.4 s	about 24x
cwcomment	0.5 s	13.6 s	about 27x
cwbox	1.6 s	34.8 s	about 22x

Across the seasons measured the ratio ran from about 5x to 38x. The C tools can also be run on several cores (one process per team file): on the same machine that took cwevent about 0.5 s for a season, so the "matches C" result above is against C on a single core.

How to reproduce this benchmark

1. Download and unzip the 2024 Retrosheet season files into a folder:

   curl -O https://www.retrosheet.org/events/2024seve.zip
   unzip -q 2024seve.zip -d retro2024 && cd retro2024
   

2. Run with standard Chadwick C:

   time cwevent -y 2024 -n 2024*.EV* > c_events.csv
   

3. Run with chadwickpy (auto-parallel is default):

   time cwevent -y 2024 -n 2024*.EV* > py_events.csv
   

4. Check that both CSVs are 100% identical byte-for-byte:

   cmp c_events.csv py_events.csv
   

When to choose which

• Use chadwickpy to get going quickly, in notebooks, on Windows, in CI, or anywhere you cannot compile software. Full seasons take just ~2.3 seconds on modern multi-core machines out of the box with zero native compilation.
• Use the C tools if you process many seasons repeatedly and the speed matters more than convenience.

Using several cores

Output is identical at every core count (checked from 1 to 40 cores, with Python's fork, spawn and forkserver worker start methods, and inside a Docker container limited to 2 CPUs). Rough times for 8 files of the 2023 season on a busy 40-core host:

Cores	1	2	4	8	16
cwevent, all fields (seconds; C on one core: 2.2)	22.5	12.1	6.6	6.3	3.7

It never starts more workers than there are files. See using several cores.

Where it intentionally differs

• Where the C crashes or reads memory it should not, chadwickpy stops with a clear error or uses a defined value. Examples: a game with no date, a month of 0 or 13, more than 50 line-score innings, and cwbox text on a game that has no plays (all of Retrosheet's box-score-only seasons 1901-1907 crash the C cwbox text output; cwbox -X works in both). The reasons are recorded in the project's design decisions.
• SportsML output (cwbox -S) is deprecated. Chadwick's own -S crashes on nearly every game, so there is nothing to compare it with. The pb attribute of cwbox -X depends on uninitialised memory in the C and is not compared.
• -j / --jobs and the CHADWICK_JOBS variable exist only here (the C tools run one process).
• On Windows, wildcards such as 2010*.EV* are expanded as Chadwick's Windows build does. Unlike that build, a folder in the pattern is kept (sub\*.EVA).
• The reference is the Chadwick 0.11.0 release (tag v0.11.0, 2026-09-10). Earlier chadwickpy versions (up to 0.3.0) followed the development commit c685ab5 that preceded it; see Chadwick versions.
• Inputs that make the C behave in undefined ways are not compared.

If both are installed

The command names are identical. Whichever directory comes first on PATH runs. To always get this package's version, use chadwickpy cwevent ... or python -m chadwickpy cwevent ....

Chadwick versions

chadwickpy follows the official Chadwick release exactly, including its command line. Chadwick 0.11.0 changed the quiet switch from -q to -Q (the C tools now reject -q), added -D dir (where the team and roster files are), changed how cwgame writes dates (-dsf, -dsp, -dnf, -dnp) and added fields (cwgame field 84, cwsub COUNT_TX), and tightened the play parser (for example ? is no longer an unknown fielder, and pickoffs may follow a caught-stealing). chadwickpy 0.4.0 does the same. If you upgrade from 0.3.x, replace -q with -Q in your scripts. A weekly check builds the newest Chadwick and runs the whole test suite against it, so a future change is noticed.