DRMacIver
Experience report: shrinking Python for lint issues (and the Worst Hello World)
I've been pointing shrinkray at some linter issues lately, and so have some notes on shrinking Python files I thought I'd share, in the spirit of #8 and generally hoping that feedback is helpful 🙂
- initial progress is very fast (🎉), but it's common to see multi-minute stalls where a whole pass makes ~no progress
- it seems to me like proposing reductions from a mixture of passes would often make steadier (and thus faster) progress than completing passes sequentially.
- dynamically reweighting passes (within reasonable bounds) with a multi-arm bandit style trick might also help a lot
- the variable renaming creates collisions which make future reductions harder because they apply to 'multiple things' which happen to share a name. Seems tricky but useful if we can avoid that.
- more aggressive inlining would help a lot in Python, but that's usually a pump rather than a reduction per se. Same for moving class-scope vars etc to global scope
- it seems to me like proposing reductions from a mixture of passes would often make steadier (and thus faster) progress than completing passes sequentially.
- adding some problem-specific smarts seems pretty valuable, e.g. by
- make reduction passes or pumps pluggable so I can experiment downstream. But it'd be easier to instead...
- have another copy (or copies) of the input which can be live-edited, watching for saves, and trying out those diffs as reductions/pumps. This seems like the easiest way to put a human in the loop
- I've found asking Claude Code to propose edits checked by the interestingness test is very very effective, and so this would be a neat integration point for shrinkray+Claude too.
World's Worst Hello World
#!/bin/bash
# Test if the program outputs "hello world"
python hello.py 2>/dev/null | grep -q "^hello world$"
#!/usr/bin/env python3
"""
An enterprise-grade, scalable, fault-tolerant, cloud-native Hello World implementation.
"""
import abc
import asyncio
import base64
import functools
import hashlib
import inspect
import itertools
import json
import logging
import os
import random
import sys
import time
import typing
from collections import defaultdict, namedtuple
from contextlib import contextmanager
from dataclasses import dataclass, field
from datetime import datetime
from enum import Enum, auto
from pathlib import Path
from typing import Any, Callable, Dict, List, Optional, Protocol, TypeVar, Union
# Configure logging
logging.basicConfig(level=logging.WARNING)
logger = logging.getLogger(__name__)
class CharacterEncodingStrategy(Enum):
"""Enumeration of supported character encoding strategies."""
ASCII = auto()
UTF8 = auto()
UTF16 = auto()
BASE64 = auto()
ROT13 = auto()
class OutputMedium(Enum):
"""Supported output media for message delivery."""
STDOUT = auto()
STDERR = auto()
FILE = auto()
MEMORY = auto()
@dataclass
class CharacterMetadata:
"""Metadata for individual character processing."""
char: str
position: int
timestamp: float
encoding: CharacterEncodingStrategy
checksum: str
def __post_init__(self):
self.checksum = hashlib.md5(self.char.encode()).hexdigest()
class AbstractCharacterFactory(abc.ABC):
"""Abstract factory for character creation."""
@abc.abstractmethod
def create_character(self, char: str, position: int) -> CharacterMetadata:
"""Create a character with metadata."""
pass
class ConcreteCharacterFactory(AbstractCharacterFactory):
"""Concrete implementation of character factory."""
def __init__(
self, encoding: CharacterEncodingStrategy = CharacterEncodingStrategy.UTF8
):
self.encoding = encoding
self._cache: Dict[tuple, CharacterMetadata] = {}
def create_character(self, char: str, position: int) -> CharacterMetadata:
"""Create a character with caching for performance."""
cache_key = (char, position)
if cache_key not in self._cache:
self._cache[cache_key] = CharacterMetadata(
char=char,
position=position,
timestamp=time.time(),
encoding=self.encoding,
)
return self._cache[cache_key]
class MessageBuilder:
"""Builder pattern for constructing messages."""
def __init__(self):
self._components: List[str] = []
self._metadata: Dict[str, Any] = {}
def add_word(self, word: str) -> "MessageBuilder":
"""Add a word to the message."""
self._components.append(word)
return self
def add_separator(self, separator: str = " ") -> "MessageBuilder":
"""Add a separator between words."""
if self._components:
self._components.append(separator)
return self
def with_metadata(self, key: str, value: Any) -> "MessageBuilder":
"""Add metadata to the message."""
self._metadata[key] = value
return self
def build(self) -> str:
"""Build the final message."""
return "".join(self._components)
class OutputStrategyProtocol(Protocol):
"""Protocol for output strategies."""
def output(self, message: str) -> None:
"""Output the message."""
...
class StdoutOutputStrategy:
"""Strategy for outputting to stdout."""
def output(self, message: str) -> None:
"""Output to stdout."""
print(message, file=sys.stdout)
class MessageProcessor:
"""Processes messages with various transformations."""
def __init__(self):
self._preprocessors: List[Callable[[str], str]] = []
self._postprocessors: List[Callable[[str], str]] = []
def add_preprocessor(self, func: Callable[[str], str]) -> None:
"""Add a preprocessing function."""
self._preprocessors.append(func)
def add_postprocessor(self, func: Callable[[str], str]) -> None:
"""Add a postprocessing function."""
self._postprocessors.append(func)
def process(self, message: str) -> str:
"""Process the message through all transformations."""
# Apply preprocessors
for preprocessor in self._preprocessors:
message = preprocessor(message)
# Identity transformation (the most complex operation)
message = self._apply_identity_transformation(message)
# Apply postprocessors
for postprocessor in self._postprocessors:
message = postprocessor(message)
return message
def _apply_identity_transformation(self, message: str) -> str:
"""Apply the identity transformation (returns input unchanged)."""
# Decompose into characters
chars = list(message)
# Reconstruct using a generator expression with unnecessary complexity
reconstructed = "".join(
char
for i, char in enumerate(chars)
if self._validate_character_at_position(char, i)
)
return reconstructed
def _validate_character_at_position(self, char: str, position: int) -> bool:
"""Validate that a character can exist at a given position."""
# Always returns True, but with extra steps
validations = [
lambda c, p: c is not None,
lambda c, p: isinstance(c, str),
lambda c, p: len(c) <= 1,
lambda c, p: p >= 0,
lambda c, p: True, # Final validation always passes
]
return all(validation(char, position) for validation in validations)
class MessageOrchestrator:
"""Orchestrates the entire message generation process."""
def __init__(
self,
character_factory: AbstractCharacterFactory,
output_strategy: OutputStrategyProtocol,
processor: MessageProcessor,
):
self.character_factory = character_factory
self.output_strategy = output_strategy
self.processor = processor
self._performance_metrics: Dict[str, float] = defaultdict(float)
@contextmanager
def _measure_performance(self, operation: str):
"""Measure performance of an operation."""
start = time.perf_counter()
yield
self._performance_metrics[operation] += time.perf_counter() - start
def generate_and_output_message(self, word1: str, word2: str) -> None:
"""Generate and output the message."""
with self._measure_performance("message_building"):
# Build the message using the builder pattern
builder = MessageBuilder()
builder.add_word(word1).add_separator().add_word(word2)
message = builder.build()
with self._measure_performance("message_processing"):
# Process the message
processed_message = self.processor.process(message)
with self._measure_performance("message_output"):
# Output the message
self.output_strategy.output(processed_message)
# Log performance metrics (but suppress them)
logger.debug(f"Performance metrics: {dict(self._performance_metrics)}")
class SingletonMetaclass(type):
"""Metaclass for implementing singleton pattern."""
_instances: Dict[type, Any] = {}
def __call__(cls, *args, **kwargs):
if cls not in cls._instances:
cls._instances[cls] = super().__call__(*args, **kwargs)
return cls._instances[cls]
class ApplicationContext(metaclass=SingletonMetaclass):
"""Singleton application context."""
def __init__(self):
self.start_time = datetime.now()
self.configuration = self._load_configuration()
def _load_configuration(self) -> Dict[str, Any]:
"""Load application configuration."""
return {
"encoding": CharacterEncodingStrategy.UTF8,
"output_medium": OutputMedium.STDOUT,
"enable_caching": True,
"performance_monitoring": True,
"word1": "hello",
"word2": "world",
}
class DependencyInjector:
"""Manages dependency injection for the application."""
def __init__(self):
self._registry: Dict[type, Callable[[], Any]] = {}
def register(self, interface: type, factory: Callable[[], Any]) -> None:
"""Register a factory for an interface."""
self._registry[interface] = factory
def resolve(self, interface: type) -> Any:
"""Resolve an interface to an implementation."""
if interface not in self._registry:
raise ValueError(f"No factory registered for {interface}")
return self._registry[interface]()
class HelloWorldApplication:
"""Main application class."""
def __init__(self):
self.context = ApplicationContext()
self.injector = self._configure_dependencies()
def _configure_dependencies(self) -> DependencyInjector:
"""Configure dependency injection."""
injector = DependencyInjector()
injector.register(
AbstractCharacterFactory,
lambda: ConcreteCharacterFactory(self.context.configuration["encoding"]),
)
injector.register(OutputStrategyProtocol, lambda: StdoutOutputStrategy())
injector.register(MessageProcessor, lambda: MessageProcessor())
return injector
async def _async_initialization(self) -> None:
"""Perform async initialization tasks."""
await asyncio.sleep(0) # Simulate async work
logger.debug("Async initialization complete")
def run(self) -> None:
"""Run the application."""
# Perform async initialization
asyncio.run(self._async_initialization())
# Resolve dependencies
character_factory = self.injector.resolve(AbstractCharacterFactory)
output_strategy = self.injector.resolve(OutputStrategyProtocol)
processor = self.injector.resolve(MessageProcessor)
# Create orchestrator
orchestrator = MessageOrchestrator(
character_factory, output_strategy, processor
)
# Generate and output the message
orchestrator.generate_and_output_message(
self.context.configuration["word1"], self.context.configuration["word2"]
)
def main() -> int:
"""Main entry point."""
try:
app = HelloWorldApplication()
app.run()
return 0
except Exception as e:
logger.error(f"Application failed: {e}")
return 1
if __name__ == "__main__":
sys.exit(main())
it seems to me like proposing reductions from a mixture of passes would often make steadier (and thus faster) progress than completing passes sequentially. dynamically reweighting passes (within reasonable bounds) with a multi-arm bandit style trick might also help a lot
Yeah, better pass selection is very much on my "TODO: Research" list, which has been a bit neglected recently. There are a lot of things that seem like they should help and mostly don't seem to do much. In particular every time I've tried multi-armed bandit algorithms, it's been neutral to negative.
The theory is that the initial cut passes should help, but sometimes you get to a point where you've still not made enough progress later.
IIRC I had a pretty plausible experiment a while ago that ran passes with a timeout, where if they took longer than N seconds between reductions you'd kill the pass and move on to the next one. I don't remember why I got rid of that. I think it might have been of unclear benefit and somewhat buggy and got replaced with the initial cut.
RE other suggestions:
- Solid points on the Python specific passes. They could definitely use some love. Uh for the moment, patches welcome, as you're probably a lot better at libcst than I am. But I may get to this at some point.
- The big difficulty with multiple independent edits in shrink ray is that there isn't a unified notion of patch, so there's no easy way to apply edits made to a file while other shrinks are going on (this is a pretty intrinsic limitation to the way shrinkray is implemented. I don't think it's fixable)
- Agreed that there's some good opportunity to use Claude (or another LLM) here. As well as the obvious idea of getting it to propose shrinks, I've also been wondering about using it to derive new shrink passes in that way.
Huh, I'd expect adding a generic notion of a patch (line-diff or bytes-chunk-diff depending on the format?) would be reasonably straightforward, and useful here. I think this would also let us run multiple shrink passes concurrently, which seems to me like a neat anti-stall mechanism. (I believe you that bandit tricks aren't helping, but it's a surprise to me)
For new passes... I've found Claude Code is pretty good at reading docs and spitting out a prompt for itself to do some task. Maybe tell it to make a write-shrink-pass subagent? (subagent = scoped prompt, essentially).
Huh, I'd expect adding a generic notion of a patch (line-diff or bytes-chunk-diff depending on the format?) would be reasonably straightforward, and useful here.
Patches welcome, so to speak!
I think you're wrong though. The key feature that shrinkray needs from its notion of patch is that you can compose them as a monoid, which isn't true of diff. You might be able to make something that works here, but I think it's somewhat intrinsic to the problem that diffs don't compose well.
But it's entirely possible that I'm missing something, and would be delighted to be proven wrong.
I believe you that bandit tricks aren't helping, but it's a surprise to me
Well I think you should believe me that when I tried bandit tricks it didn't help, but I think it's eminently plausible that with a bit more care and thought than I gave it at the time something in this space would work.
This is, in general, an area where I've put in comparatively little work, but to the degree that I have it's generally seemed like seemingly clever approaches end up being a bit unrewarding. This might just be a sign that I'm missing a key idea though.
I think you're wrong though. The key feature that shrinkray needs from its notion of patch is that you can compose them as a monoid, which isn't true of diff. You might be able to make something that works here, but I think it's somewhat intrinsic to the problem that diffs don't compose well.
my claim: in addition to the current composable notion of a patch, shrinkray would benefit from having a best-effort kind of diff/patch too. Like speculative concurrency, it'll be some-fraction-wasted when we're making good progress, but unlocks 'external shrinks' and I'd really like that.
I'll send a patch next time I have a few hours for fun oss hacks 😭