Skip to content

Latest commit

 

History

History
297 lines (209 loc) · 9.18 KB

README.md

File metadata and controls

297 lines (209 loc) · 9.18 KB
Raw

Hotload your Python

hotload enables exploratory programming with Python by providing a super-fast feedback loop and continuous program state.

Why hotload?

Hotload lets you reload a single file, rather than rerunning your application. That really makes a difference when your dependencies are heavy. When I'm testing this script:

# script_with_deps.py

import numpy as np
import pandas as pd
import altair as alt
import psycopg2

print("Done!")

python script_with_deps.py completes in 480 ms, whereas a hotload reload completes in 4 ms. For big projects, this can allow you to stay focused, and test your changes continuously. In addition, you don't have to switch out to a terminal, a single control+s in the file you're developing is enough.

hotload is simple:

  • No dependencies other than Python 2/3.
  • Small: at ~200 lines, you can read the whole source.

Quick start with CLI (Linux / Mac)

Windows users: you'll currently have to use the Python API (next section). Sorry for that!

With the CLI, you'll be running hotload as a binary script. This is nice for testing small things, and works as an alternative to calling python manually on the command line each time you want to run the script.

  1. Install the script. Download hotload.py manually and make it executable, or use this one-liner:

     curl https://raw.githubusercontent.com/teodorlu/hotload/master/install.sh | bash

  2. Ensure that you have ~/.local/bin on your path.

  3. Create hello.py, and try echo hello.py | hotload hello.py

Make some changes to hello.py. Can you see it change?

Note: Windows support for the CLI would be nice, but I rarely (never) use the Windows command line to run scripts. If you're using Windows Susbsystem for Linux, the install instructions should work just fine from a bash shell. Ideas / discussion / PRs for proper Windows CLI support is interesting, though.

Using the Python API (Linux / Mac / Windows)

With the Python API, you'll have more flexibility in how you run Hotload. You may configure what files should be watched, whether you want the screen to refresh, and you may load multiple modules.

Hotload is distributed as a single file

  1. Create lib.py and launch.py in a new folder
  2. Copy hotload.py in there as well.

Our task is to develop lib.py. We're going to do this by creating a really fast feedback loop, where saves trigger re-runs.

# In lib.py

x = 3
y = 4

print(x*x + y*y)

To hotload lib.py, we're going to use launch.py as a load script. This script sets up and runs hotload. We'll use this start:

# In launch.py

import hotload

hotload.hotload(
    watch=[
        hotload.listfiles(".", ext=".py")
    ],
    steps=[
        hotload.ClearTerminal(),                               # (1) 
        hotload.ReloadedPythonModule.from_module_name("lib"),  # (2)
    ]
)

# (1) hotload.ClearTerminal() ensures a "dashboard"-like experience when we work, so
#     that the terminal doesn't keep scrolling down.

# (2) then we set up a reloaded python module.

Now, use it! Run the launch script with

$ python launch.py

... and edit lib.py and save! Each save triggers a new reload.

Supported platforms

Hotload aims to support:

  • Python 2.7.3 and later 2.7.x versions
  • Python 3.7, and later 3.x versions
  • Mac, Linux, Windows
  • Abaqus Python 2.7.x

We want to support Python 2.7 because hotloading shines when working with old, clunky systems. We can reload a single piece of the system to test our change.

hotload has been tested on the following systems:

  • Source-built Python 3.8.1 on Ubuntu 18.04
  • system-installed Python 3.7.3 on Ubuntu 18.04
  • Anaconda Python 3.7.3 on Ubuntu 18.04
  • Anaconda Python 3.7.3 on Windows 10
  • Anaconda Python 2.7.15 on Windows 10
  • Abaqus Python 2.7.4 on Windows 10

Please report a bug in an issue if hotload doesn't work on any supported platform.

Advanced usage

Did you get the reloaded experience? Good! Would you like some more? Then, please read on! This is going to cover some limitations with naïve reloading, and how we can overcome them.

Exceptions and interrupts

The balance between liveness and reliability is important. Therefore, it's recommended to understand the exception model of hotload.

  1. C-c (Control-c) interrupts the reload loop
  2. For all other exceptions, the stacktrace is printed and the reload loop continues.

Why aren't recursive dependencies reloaded?

The following setup won't reload mymath.py:

# in lib.py

import mymath

print(mymath.f(1,2))
# in mymath.py

def f(x,y):
    return x*x + y*y

To trigger reloads in mymath.py from the lib.py entry point, we need a manual reload. Reloading differs between Python 2 and 3.

# In lib.py, Python 3

from importlib import reload

import mymath
reload(mymath)

print(mymath.f(2,3))

A design note. The first version of this library tried to infer what modules it would have to reload. It maintained a list of all the modules it had previously reloaded, and watched those for files. Lots of automation. Lots of complexity. In the end, I didn't find it useful at all. Instead, I add in reloads in the modules I'm developing at the top like this. The advantage? It's all dynamic -- I don't have to restart hotload -- it can just keep on runnning. If you've chosen to watch a folder, each file change will trigger a reload to the top-level entry point.

Calling a specific function in a module after reload

Just developing modules by having a reload "be your test" works fine for small modules, and is quite nice to work with. However, I've found it problematic for larger systems. Then I'd like to create an init function in the bottom of the file instead, and call the init function from hotload.

To use this, you can extend hotload.ReloadedPythonModule and override pre_reload_hook (for cleanup) or post_reload_hook (for initialization).

This allows you to let the your reloadable application code stay clean and nice, and you can provide different configurations for running different parts of your code.

See the post-reload-hook example for more details!

How do I persist state across reloads?

Usage cases:

  • Keep a connection to an external system open under reloads
  • Avoid having to re-run time-consuming external calls on each reload

You can use NameError to trigger the computation in the first place:

# in lib.py

import time

def expensive_computation():
    # ...
    time.sleep(5)
    return 42

try:
    constant = constant
except NameError:
    constant = expensive_computation()

This will trigger expensive_computation() on the first run, and cache the result on the next reload. Why? Because reload doesn't flush the module's namespace. It still has access to everything that was there before! Which can be neat ...

Inspiration

  • Bruce Hauman's Figwheel revolutionized web development by hotloading new code into a running JavaScript environment. Thanks!

Future work

Hotload's command line interface may need some work. In its current form, it's quite limited. We could try to go for multiple arguments:

find . | grep ".py" | hotload.py --clear-terminal --hotload script --run echo "That's a reload!"

Message from 2021-10-21 - I've never really felt the need for a more complex CLI. So that's perhaps over engineering.

It might be useful to run hotload in the background. Why? Because then we could interact with a hotloaded Python REPL that (at all times) is aware of our code.

Tests might be nice. To be able to ensure that there's no primitive errors across platforms. Challenge: hotload typically operates on the "outer" layer of a code base, so we'd have to build an outer-outer layer.

Other options

  • Jupyter Notebooks enable a similar workflow, where you import your dependencies once and stay in-process afterwards. Notebooks are a good option when you don't want to consider your script as a whole.
  • entr provides this workflow as a command-line, language-agnostic tool. I use entr all the time on Linux. Limitations: you have to hop out of your Python context, and you'll have to install it on a target environment which supports it.
  • Test runners for Python. You can use hotload with a test runner, and have your lib.py run the test runner, thereby avoid having to restart the Python interpreter. Or you might have a test runner with a built in reloader; in which case you might not need this.
  • Other Python reloaders like hoh/reloadr.

FAQ

  • Q: Does hotload use polling?
  • A: Yes. That means it asks the OS whether all watched files have changed each iteration (by default 144 iterations per second). Using polling is simple, and works across systems. If you need lower response times or less system load, you might want to use something like inotify on Linux. Search for existing Python libraries! Note that these are going to be more complex than hotload.