Peterbe.com

This an optimization story that should not surprise anyone using the Django ORM. But I thought I'd share because I have numbers now! The origin of this came from a real requirement. For a given parent model, I'd like to extract the value of the name column of all its child models, and the turn all these name strings into 1 MD5 checksum string.

Variants

The first attempted looked like this:

artist = Artist.objects.get(name="Bad Religion")
names = []
for song in Song.objects.filter(artist=artist):
    names.append(song.name)
return hashlib.md5("".join(names).encode("utf-8")).hexdigest()

The SQL used to generate this is as follows:

SELECT "main_song"."id", "main_song"."artist_id", "main_song"."name", 
"main_song"."text", "main_song"."language", "main_song"."key_phrases", 
"main_song"."popularity", "main_song"."text_length", "main_song"."metadata", 
"main_song"."created", "main_song"."modified", 
"main_song"."has_lastfm_listeners", "main_song"."has_spotify_popularity" 
FROM "main_song" WHERE "main_song"."artist_id" = 22729;

Clearly, I don't need anything but just the name column, version 2:

artist = Artist.objects.get(name="Bad Religion")
names = []
for song in Song.objects.filter(artist=artist).only("name"):
    names.append(song.name)
return hashlib.md5("".join(names).encode("utf-8")).hexdigest()

Now, the SQL used is:

SELECT "main_song"."id", "main_song"."name" 
FROM "main_song" WHERE "main_song"."artist_id" = 22729;

But still, since I don't really need instances of model class Song I can use the .values() method which gives back a list of dictionaries. This is version 3:

names = []
for song in Song.objects.filter(artist=a).values("name"):
    names.append(song["name"])
return hashlib.md5("".join(names).encode("utf-8")).hexdigest()

This time Django figures it doesn't even need the primary key value so it looks like this:

SELECT "main_song"."name" FROM "main_song" WHERE "main_song"."artist_id" = 22729;

Last but not least; there is an even faster one. values_list(). This time it doesn't even bother to map the column name to the value in a dictionary. And since I only need 1 column's value, I can set flat=True. Version 4 looks like this:

names = []
for name in Song.objects.filter(artist=a).values_list("name", flat=True):
    names.append(name)
return hashlib.md5("".join(names).encode("utf-8")).hexdigest()

Same SQL gets used this time as in version 3.

The benchmark

Hopefully this little benchmark script speaks for itself:

from songsearch.main.models import *

import hashlib


def f1(a):
    names = []
    for song in Song.objects.filter(artist=a):
        names.append(song.name)
    return hashlib.md5("".join(names).encode("utf-8")).hexdigest()


def f2(a):
    names = []
    for song in Song.objects.filter(artist=a).only("name"):
        names.append(song.name)
    return hashlib.md5("".join(names).encode("utf-8")).hexdigest()


def f3(a):
    names = []
    for song in Song.objects.filter(artist=a).values("name"):
        names.append(song["name"])
    return hashlib.md5("".join(names).encode("utf-8")).hexdigest()


def f4(a):
    names = []
    for name in Song.objects.filter(artist=a).values_list("name", flat=True):
        names.append(name)
    return hashlib.md5("".join(names).encode("utf-8")).hexdigest()


artist = Artist.objects.get(name="Bad Religion")
print(Song.objects.filter(artist=artist).count())

print(f1(artist) == f2(artist))
print(f2(artist) == f3(artist))
print(f3(artist) == f4(artist))

# Reporting
import time
import random
import statistics

functions = f1, f2, f3, f4
times = {f.__name__: [] for f in functions}

for i in range(500):
    func = random.choice(functions)
    t0 = time.time()
    func(artist)
    t1 = time.time()
    times[func.__name__].append((t1 - t0) * 1000)

for name in sorted(times):
    numbers = times[name]
    print("FUNCTION:", name, "Used", len(numbers), "times")
    print("\tBEST", min(numbers))
    print("\tMEDIAN", statistics.median(numbers))
    print("\tMEAN  ", statistics.mean(numbers))
    print("\tSTDEV ", statistics.stdev(numbers))

I ran this on my PostgreSQL 11.1 on my MacBook Pro with Django 2.1.7. So the database is on localhost.

The results

276
True
True
True
FUNCTION: f1 Used 135 times
    BEST 6.309986114501953
    MEDIAN 7.531881332397461
    MEAN   7.834429211086697
    STDEV  2.03779968066591
FUNCTION: f2 Used 135 times
    BEST 3.039121627807617
    MEDIAN 3.7298202514648438
    MEAN   4.012803678159361
    STDEV  1.8498943539073027
FUNCTION: f3 Used 110 times
    BEST 0.9920597076416016
    MEDIAN 1.4405250549316406
    MEAN   1.5053835782137783
    STDEV  0.3523240470133114
FUNCTION: f4 Used 120 times
    BEST 0.9369850158691406
    MEDIAN 1.3251304626464844
    MEAN   1.4017681280771892
    STDEV  0.3391019435930447

Discussion

I guess the hashlib.md5("".join(names).encode("utf-8")).hexdigest() stuff is a bit "off-topic" but I checked and it's roughly 300 times faster than building up the names list.

It's clearly better to ask less of Python and PostgreSQL to get a better total time. No surprise there. What was interesting was the proportion of these differences. Memorize that and you'll be better equipped if it's worth the hassle of not using the Django ORM in the most basic form.

Also, do take note that this is only relevant in when dealing with many records. The slowest variant (f1) takes, on average, 7 milliseconds.

Summarizing the difference with percentages compared to the fastest variant:

• f1 - 573% slower
• f2 - 225% slower
• f3 - 6% slower
• f4 - 0% slower

UPDATE Feb 25 2019

James suggested, although a bit "missing the point", that it could be even faster if all the aggregation is pushed into the PostgreSQL server and then the only thing that needs to transfer from PostgreSQL to Python is the final result.

By the way, name column in this particular benchmark, when concatenated into one big string, is ~4KB. So, with variant f5 it only needs to transfer 32 bytes which will/would make a bigger difference if the network latency is higher.

Here's the whole script: https://gist.github.com/peterbe/b2b7ed95d422ab25a65639cb8412e75e

And the results:

276
True
True
True
False
False
FUNCTION: f1 Used 92 times
    BEST 5.928993225097656
    MEDIAN 7.311463356018066
    MEAN   7.594626882801885
    STDEV  2.2027017044658423
FUNCTION: f2 Used 75 times
    BEST 2.878904342651367
    MEDIAN 3.3979415893554688
    MEAN   3.4774907430013022
    STDEV  0.5120246550765524
FUNCTION: f3 Used 88 times
    BEST 0.9310245513916016
    MEDIAN 1.1944770812988281
    MEAN   1.3105544176968662
    STDEV  0.35922655625999383
FUNCTION: f4 Used 71 times
    BEST 0.7879734039306641
    MEDIAN 1.1661052703857422
    MEAN   1.2262606284987758
    STDEV  0.3561764250427344
FUNCTION: f5 Used 90 times
    BEST 0.7929801940917969
    MEDIAN 1.0334253311157227
    MEAN   1.1836051940917969
    STDEV  0.4001442703048186
FUNCTION: f6 Used 84 times
    BEST 0.80108642578125
    MEDIAN 1.1119842529296875
    MEAN   1.2281338373819988
    STDEV  0.37146893005516973

Result: f5 is takes 0.793ms and (the previous "winner") f4 takes 0.788ms.

I'm not entirely sure why f5 isn't faster but I suspect it's because the dataset is too small for it all to matter.

Compare:

songsearch=# explain analyze SELECT "main_song"."name" FROM "main_song" WHERE "main_song"."artist_id" = 22729;
                                                             QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------
 Index Scan using main_song_ca949605 on main_song  (cost=0.43..229.33 rows=56 width=16) (actual time=0.014..0.208 rows=276 loops=1)
   Index Cond: (artist_id = 22729)
 Planning Time: 0.113 ms
 Execution Time: 0.242 ms
(4 rows)

with...

songsearch=# explain analyze SELECT md5(STRING_AGG("main_song"."name", '')) AS "names_hash" FROM "main_song" WHERE "main_song"."artist_id" = 22729;
                                                                QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------
 Aggregate  (cost=229.47..229.48 rows=1 width=32) (actual time=0.278..0.278 rows=1 loops=1)
   ->  Index Scan using main_song_ca949605 on main_song  (cost=0.43..229.33 rows=56 width=16) (actual time=0.019..0.204 rows=276 loops=1)
         Index Cond: (artist_id = 22729)
 Planning Time: 0.115 ms
 Execution Time: 0.315 ms
(5 rows)

I ran these two SQL statements about 100 times each and recorded their best possible execution times:

1) The plain SELECT - 0.99ms
2) The STRING_AGG - 1.06ms

So that accounts from ~0.1ms difference only! Which kinda matches the results seen above. All in all, I think the dataset is too small to demonstrate this technique. But, considering the chance that the complexity might not be linear with the performance benefit, it's still interesting.

Even though this tangent is a big off-topic, it is often a great idea to push as much work into the database as you can if applicable. Especially if it means you can transfer a lot less data eventually.

1. Create a create-react-app first:

create-react-app myapp

2. Enter it and install node-sass and bulmaswatch

cd myapp
yarn add bulma bulmaswatch node-sass

3. Edit the src/index.js to import index.scss instead:

-import "./index.css";
+import "./index.scss";

4. "Rename" the index.css file:

git rm src/index.css 
touch src/index.scss
git add src/index.scss

5. Now edit the src/index.scss to look like this:

@import "node_modules/bulmaswatch/darkly/bulmaswatch";

This assumes your favorite theme was the darkly one. You can obviously change that later.

6. Run the app:

BROWSER=none yarn start

7. Open the browser at http://localhost:3000

That's it! However, the create-react-app default look doesn't expose any of the cool stuff that Bulma can style. So let's rewrite our src/App.js by copying the minimal starter HTML from the Bulma documentation. So make the src/App.js component look something like this:

class App extends Component {
  render() {
    return (
      <section className="section">
        <div className="container">
          <h1 className="title">Hello World</h1>
          <p className="subtitle">
            My first website with <strong>Bulma</strong>!
          </p>
        </div>
      </section>
    );
  }
}

Now it'll look like this:

Yes, it's not much but it's a great start. Over to you to take this to infinity and beyond!

Not So Secret Sauce

In the rushed instructions above the choice of theme was darkly. But what you need to do next is go to https://jenil.github.io/bulmaswatch/, click around and eventually pick the one you like. Suppose you like spacelab, then you just change that @import ... line to be:

@import "node_modules/bulmaswatch/spacelab/bulmaswatch";

TEMPORARY:

h1 {  
    color: red;
    font-size: 5em;
}

TEST CHANGE 3.

tl;dr; minimalcss 0.8.2 introduces a 20% post-processing optimization by lumping many CSS selectors to their parent CSS selectors as a pre-emptive cache.

In minimalcss the general core of it is that it downloads a DOM tree, as HTML, parses it and parses all the CSS stylesheets associated. These might be from <link ref="stylesheet"> or <style> tags.
Once the CSS stylesheets are turned into an AST it loops over each and every CSS selector and asks a simple question; "Does this CSS selector exist in the DOM?". The equivalent is to open your browser's Web Console and type:

>>> document.querySelectorAll('div.foo span.bar b').length > 0
false

For each of these lookups (which is done with cheerio by the way), minimalcss reduces the CSS, as an AST, and eventually spits the AST back out as a CSS string. The only problem is; it's slow. In the case of view-source:https://semantic-ui.com/ in the CSS it uses, there are 6,784 of them. What to do?

First of all, there isn't a lot you can do. This is the work that needs to be done. But one thing you can do is be smart about which selectors you look at and use a "decision cache" to pre-emptively draw conclusions. So, if this is what you have to check:

1. #example .alternate.stripe
2. #example .theming.stripe
3. #example .solid .column p b
4. #example .solid .column p

As you process the first one you extract that the parent CSS selector is #example and if that doesn't exist in the DOM, you can efficiently draw conclusion about all preceeding selectors that all start with #example .... Granted, if they call exist you will pay a penalty of doing an extra lookup. But that's the trade-off that this optimization is worth.

Check out the comments where I tested a bloated page that uses Semantic-UI before and after. Instead of doing 3,285 of these document.querySelector(selector) calls, it's now able too come to the exact same conclusion with just 1,563 lookups.

Sadly, the majority of the time spent processing lies in network I/O and other overheads but this work did reduce something that used to take 6.3s (median) too 5.1s (median).

React 16.8.0 with Hooks was released today. A big deal. Executive summary; components as functions is all the rage now.

What used to be this:

class MyComponent extends React.Component {
  ...

  componentDidMount() {
    ...
  }
  componentDidUpdate() {
    ...
  }

  render() { STUFF }
}

...is now this:

function MyComponent() {
  ...

  React.useEffect(() => {
    ...
  })

  return STUFF
}

Inside the useEffect "side-effect callback" you can actually update state. But if you do, and this is no different that old React.Component.componentDidUpdate, it will re-run the side-effect callback. Here's a simple way to cause an infinite recursion:

// DON'T DO THIS

function MyComponent() {
  const [counter, setCounter] = React.useState(0);

  React.useEffect(() => {
    setCounter(counter + 1);
  })

  return <p>Forever!</p>
}

The trick is to pass a second argument to React.useEffect that is a list of states to exclusively run on.

Here's how to fix the example above:

function MyComponent() {
  const [counter, setCounter] = React.useState(0);
  const [times, setTimes] = React.useState(0);

  React.useEffect(
    () => {
      if (times % 3 === 0) {
        setCounter(counter + 1);
      }
    },
    [times]  // <--- THIS RIGHT HERE IS THE KEY!
  );

  return (
    <div>
      <p>
        Divisible by 3: {counter}
        <br />
        Times: {times}
      </p>
      <button type="button" onClick={e => setTimes(times + 1)}>
        +1
      </button>
    </div>
  );
}

You can see it in this demo.

Note, this isn't just about avoiding infinite recursion. It can also be used to fit your business logic and/or an optimization to avoid executing the effect too often.

tl;dr; You can use error instanceof window.Response to distinguish between fetch exceptions and fetch responses.

When you do something like...

const response = await fetch(URL);

...two bad things can happen.

1. The XHR request fails entirely. I.e. there's not even a response with a HTTP status code.
2. The response "worked" but the HTTP status code was not to your liking.

Either way, your React app needs to deal with this. Ideally in a not-too-clunky way. So here is one take on this challenge/opportunity which I hope can inspire you to extend it the way you need it to go.

The trick is to "clump" exceptions with responses. Then you can do this:

function ShowServerError({ error }) {
  if (!error) {
    return null;
  }
  return (
    <div className="alert">
      <h3>Server Error</h3>
      {error instanceof window.Response ? (
        <p>
          <b>{error.status}</b> on <b>{error.url}</b>
          <br />
          <small>{error.statusText}</small>
        </p>
      ) : (
        <p>
          <code>{error.toString()}</code>
        </p>
      )}
    </div>
  );
}

The greatest trick the devil ever pulled was to use if (error instanceof window.Reponse) {. Then you know that error thing is the outcome of THIS = await fetch(URL) (or fetch(URL).then(THIS) if you prefer). Another good trick the devil pulled was to be aware that exceptions, when asked to render in React does not naturally call its .toString() so you have to do that yourself with {error.toString()}.

This codesandbox demonstrates it quite well. (Although, at the time of writing, codesandbox will spew warnings related to testing React components in the console log. Ignore that.)

If you can't open that codesandbox, here's the gist of it:

React.useEffect(() => {
  url &&
    (async () => {
      let response;
      try {
        response = await fetch(url);
      } catch (ex) {
        return setServerError(ex);
      }
      if (!response.ok) {
        return setServerError(response);
      }
      // do something here with `await response.json()`
    })(url);
}, [url]);

By the way, another important trick is to be subtle with how you put the try { and } catch(ex) {.

// DON'T DO THIS

try {
  const response = await fetch(url);
  if (!response.ok) {
    setServerError(response);
  }
  // do something here with `await response.json()`
} catch (ex) {
  setServerError(ex);
}

Instead...

// DO THIS

let response;
try {
  response = await fetch(url);
} catch (ex) {
  return setServerError(ex);
}
if (!response.ok) {
  return setServerError(response);
}
// do something here with `await response.json()`

If you don't do that you risk catching other exceptions that aren't exclusively the fetch() call. Also, notice the use of return inside the catch block which will exit the function early leaving you the rest of the code (de-dented 1 level) to deal with the happy-path response object.

Be aware that the test if (!response.ok) is simplistic. It's just a shorthand for checking if the "status in the range 200 to 299, inclusive". Realistically getting a response.status === 400 isn't an "error" really. It might just be a validation error hint from a server, and likely the await response.json() will work and contain useful information. No need to throw up a toast or a flash message that the communication with the server failed.

Conclusion

The details matter. You might want to deal with exceptions entirely differently from successful responses with bad HTTP status codes. It's nevertheless important to appreciate two things:

1. Handle complete fetch() failures and feed your UI or your retry mechanisms.

2. You can, in one component distinguish between a "successful" fetch() call and thrown JavaScript exceptions.

tl;dr; Use f"{number:,}" to thousands format an integer to a string.

I keep forgetting and having to look this up every time. Hopefully by blogging about it, this time it'll stick in my memory. And hopefully in yours too :)

Suppose you have a number, like 1234567890 and you want to display it, here's how you do it:

>>> number = 1234567890
>>> f"{number:,}"
'1,234,567,890'

In the past, before Python 3.6, I've been using:

>>> number = 1234567890
>>> format(number, ",")
'1,234,567,890'

All of this and more detail can be found in PEP 378 -- Format Specifier for Thousands Separator. For example, you can do this beast too:

>>> number = 1234567890
>>> f"{number:020,.2f}"
'0,001,234,567,890.00'

which demonstrates (1) how to do zero-padding (of length 20), (2) the thousands comma, (3) round to 2 significant figures. All useful weapons to be able to draw from the top of your head.

UPDATE

Also, incredibly useful is the equivalent of somestring.ljust(10):

>>> mystr = "peter"
>>> f"{mystr:10}"
'peter     '
>>> f"{mystr:>10}"
'     peter'

Prior to version 0.14.5 hashin would write write down the hashes of PyPI packages in the order they appear in PyPI's JSON response. That means there's a slight chance that two distinct clients/computers/humans might actually get different output when then run hashin Django==2.1.5.

The pull request has a pretty hefty explanation as it demonstrates the fix.

Do note that if the existing order of hashes in a requirements file is not in the "right" order, hashin won't correct it unless any of the hashes are different.

Thanks @SomberNight for patiently pushing for this.

tl;dr; If you use the django.views.decorators.cache.cache_control decorator, consider this one instead to change the max_age depending on the request.

I had/have a Django view function that looks something like this:

@cache_control(public=True, max_age=60 * 60)
def home(request, oc=None, page=1):
    ...

But, that number 60 * 60 I really needed it to be different depending on the request parameters. For example, that oc=None, if that's not None I know the page's Cache-Control header can and should be different.

So I wrote this decorator:

from django.utils.cache import patch_cache_control


def variable_cache_control(**kwargs):
    """Same as django.views.decorators.cache.cache_control except this one will
    allow the `max_age` parameter be a callable.
    """

    def _cache_controller(viewfunc):
        @functools.wraps(viewfunc)
        def _cache_controlled(request, *args, **kw):
            response = viewfunc(request, *args, **kw)
            copied = kwargs
            if kwargs.get("max_age") and callable(kwargs["max_age"]):
                max_age = kwargs["max_age"](request, *args, **kw)
                # Can't re-use, have to create a shallow clone.
                copied = dict(kwargs, max_age=max_age)
            patch_cache_control(response, **copied)
            return response

        return _cache_controlled

    return _cache_controller

Now, I can do this instead:

def _best_max_age(req, oc=None, **kwargs):
    max_age = 60 * 60
    if oc:
        max_age *= 10
    return max_age

@variable_cache_control(public=True, max_age=_best_max_age)
def home(request, oc=None, page=1):
    ...

I hope it inspires.

When Immer first came out I was confused. I kinda understood what I was reading but I couldn't really see what was so great about it. As always, nothing beats actual code you type yourself to experience how something works.

Here is, I believe, a great example: https://codesandbox.io/s/y2m399pw31

If you're reading this on your mobile it might be hard to see what it does. Basically, it's a very simple React app that displays a "todo list like" thing. The state (aka. this.state.tasks) is a pure JavaScript array. The React components that display the data (e.g. <List tasks={this.state.tasks}/> and <ShowItem item={item} />) are pure (i.e. extends React.PureComponent) meaning React natively protects from re-rendering a component when the props haven't changed. So no wasted render-cycles.

What Immer does is that it helps mutate an object in a smart way. I'm sure you've heard that you're never supposed to mutate state objects (arrays are a form of mutable objects too!) and instead do things like const stuff = Object.assign({}, this.state.stuff); or const things = this.state.things.slice(0);. However, those things are shallow copies meaning any mutable objects within (i.e. nested objects) don't get the clone treatment and can thus cause problems with not re-rendering when they should.

Here's the core gist:

import React from "react";
import produce from "immer";

class App extends React.Component {
  state = {
    tasks: [[false, { text: "Do something", date: new Date() }]]
  };
  onToggleDone = (i, done) => {
    // Immer
    // This is what the blog post is all about...
    const tasks = produce(this.state.tasks, draft => {
      draft[i][0] = done;
      draft[i][1].date = new Date();
    });

    // Pure JS
    // Don't do this!
    // const tasks = this.state.tasks.slice(0);
    // tasks[i][0] = done;
    // tasks[i][1].date = new Date();

    this.setState({ tasks });
  };
  render() {
    // appreviated, but...
    return <List tasks={this.state.tasks}/>
  }
}

class List extends React.PureComponent {
   ...

It just works. Neat!

By the way, here's a code sandbox that accomplishes the same thing but with ImmutableJS which I think is uglier. I think it's uglier because now the rendering components need to be aware that it's rendering immutable.Map objects instead.

Caveats

1. The cost of doing what immer.produce isn't free. It's some smart work that needs to be done. But the alternative is to deep clone the object which is going to be much slower. Immer isn't the fastest kid on the block but unlike MobX and ImmutableJS once you've done this smart stuff you're back to plain JavaScript objects.

2. Careful with doing something like console.log(draft) since it will raise a TypeError in your web console. Just be aware of that or use console.log(JSON.stringify(draft)) instead.

3. If you know with confidence that your mutable object does not, and will not, have nested mutable objects you can use object spread, Object.assign(), or .slice(0) and save yourself the trouble of another dependency.

I only just learned about this today after all these years and thought you might like it too.

The trick is to conveniently turn an argparse.Namespace into keyword arguments that you can send to a function. This is the old/wrong way I've been doing it for years:

# THE OLD WAY

def main(things, option_a, option_n):
    print(locals())  # Debugging 


import argparse

parser = argparse.ArgumentParser()
parser.add_argument("things", help="Bla bla", nargs="*")
parser.add_argument("-o", "--option-a", help="Bla bla", default="Op A")
parser.add_argument("-n", "--option-n", help="Ble ble", default="Op N")
args = parser.parse_args()

main(
    things=args.things,
    option_a=args.option_a,
    option_n=args.option_n
)

That works but the tedious thing is to have to have spell out every single argument, twice!, when sending the argparse Namespace into the function. Here's the much moar betterest way:

# THE NEW WAY

def main(things, option_a, option_n):
    print(locals())  # Debugging 


import argparse

parser = argparse.ArgumentParser()
parser.add_argument("things", help="Bla bla", nargs="*")
parser.add_argument("-o", "--option-a", help="Bla bla", default="Op A")
parser.add_argument("-n", "--option-n", help="Ble ble", default="Op N")
args = parser.parse_args()

# The only difference and the magic sauce...
main(**vars(args))  

What's neat about this is that you don't have to type up every argument defined in the parser to the get it as arguments into a function. And as a bonus, Python will name match keyword arguments to arguments so the order doesn't matter.

Caveat! This "trick" assumes that the arguments in the parser match the arguments in the function. So if the main() function takes an argument called foo_bar you have to have an argument in the parser called --foo-bar.