A blog and website by Peter Bengtsson

Filtered home page!
Currently only showing blog entries under the category: Python. Clear filter

Train your own spell corrector with TextBlob

23 August 2019 0 comments   Python

TextBlob is a wonderful Python library it. It wraps nltk with a really pleasant API. Out of the box, you get a spell-corrector. From the tutorial:

>>> from textblob import TextBlob
>>> b = TextBlob("I havv goood speling!")
>>> str(b.correct())
'I have good spelling!'

The way it works is that, shipped with the library, is this text file: en-spelling.txt It's about 30,000 lines long and looks like this:

;;;   Based on several public domain books from Project Gutenberg
;;;   and frequency lists from Wiktionary and the British National Corpus.
a 21155
aah 1
aaron 5
ab 2
aback 3
abacus 1
abandon 32
abandoned 72
abandoning 27

That gave me an idea! How about I use the TextBlob API but bring my own text as the training model. It doesn't have to be all that complicated.

The challenge

(Note: All the code I used for this demo is available here:

I found this site that lists "Top 1,000 Baby Boy Names". From that list, randomly pick a couple of out and mess with their spelling. Like, remove letters, add letters, and swap letters.

So, 5 random names now look like this:

▶ python
RIGHT: jameson  TYPOED: jamesone
RIGHT: abel     TYPOED: aabel
RIGHT: wesley   TYPOED: welsey
RIGHT: thomas   TYPOED: thhomas
RIGHT: bryson   TYPOED: brysn

Imagine some application, where fat-fingered users typo those names on the right-hand side, and your job is to map that back to the correct spelling.

First, let's use the built in TextBlob.correct. A bit simplified but it looks like this:

from textblob import TextBlob

correct, typo = get_random_name()
b = TextBlob(typo)
result = str(b.correct())
right = correct == result

And the results:

▶ python
ORIGIN         TYPO           RESULT         WORKED?
jesus          jess           less           Fail
austin         ausin          austin         Yes!
julian         juluian        julian         Yes!
carter         crarter        charter        Fail
emmett         emett          met            Fail
daniel         daiel          daniel         Yes!
luca           lua            la             Fail
anthony        anthonyh       anthony        Yes!
damian         daiman         cabman         Fail
kevin          keevin         keeping        Fail
Right 40.0% of the time

Buuh! Not very impressive. So what went wrong there? Well, the word met is much more common than emmett and the same goes for words like less, charter, keeping etc. You know, because English.

The solution

The solution is actually really simple. You just crack open the classes out of textblob like this:

from textblob import TextBlob
from textblob.en import Spelling

path = "spelling-model.txt"
spelling = Spelling(path=path)
# Here, 'names' is a list of all the 1,000 correctly spelled names.
# e.g. ['Liam', 'Noah', 'William', 'James', ...
spelling.train(" ".join(names), path)

Now, instead of corrected = str(TextBlob(typo).correct()) we do result = spelling.suggest(typo)[0][0] as demonstrated here:

correct, typo = get_random_name()
b = spelling.suggest(typo)
result = b[0][0]
right = correct == result

So, let's compare the two "side by side" and see how this works out. Here's the output of running with 20 randomly selected names:

▶ python
ORIGIN         TYPO           RESULT         WORKED?
juan           jaun           juan           Yes!
ethan          etha           the            Fail
bryson         brysn          bryan          Fail
hudson         hudsn          hudson         Yes!
oliver         roliver        oliver         Yes!
ryan           rnyan          ran            Fail
cameron        caeron         carron         Fail
christopher    hristopher     christopher    Yes!
elias          leias          elias          Yes!
xavier         xvaier         xvaier         Fail
justin         justi          just           Fail
leo            lo             lo             Fail
adrian         adian          adrian         Yes!
jonah          ojnah          noah           Fail
calvin         cavlin         calvin         Yes!
jose           joe            joe            Fail
carter         arter          after          Fail
braxton        brxton         brixton        Fail
owen           wen            wen            Fail
thomas         thoms          thomas         Yes!
Right 40.0% of the time

ORIGIN         TYPO           RESULT         WORKED?
landon         landlon        landon         Yes
sebastian      sebstian       sebastian      Yes
evan           ean            ian            Fail
isaac          isaca          isaac          Yes
matthew        matthtew       matthew        Yes
waylon         ywaylon        waylon         Yes
sebastian      sebastina      sebastian      Yes
adrian         darian         damian         Fail
david          dvaid          david          Yes
calvin         calivn         calvin         Yes
jose           ojse           jose           Yes
carlos         arlos          carlos         Yes
wyatt          wyatta         wyatt          Yes
joshua         jsohua         joshua         Yes
anthony        antohny        anthony        Yes
christian      chrisian       christian      Yes
tristan        tristain       tristan        Yes
theodore       therodore      theodore       Yes
christopher    christophr     christopher    Yes
joshua         oshua          joshua         Yes
Right 90.0% of the time

See, with very little effort you can got from 40% correct to 90% correct.

Note, that the output of something like spelling.suggest('darian') is actually a list like this: [('damian', 0.5), ('adrian', 0.5)] and you can use that in your application. For example:

<li><a href="?name=damian">Did you mean <b>damian</b></a></li>
<li><a href="?name=adrian">Did you mean <b>adrian</b></a></li>

Bonus and conclusion

Ultimately, what TextBlob does is a re-implementation of Peter Norvig's original implementation from 2007. I too, have written my own implementation in 2007. Depending on your needs, you can just figure out the licensing of that source code and lift it out and implement in your custom ways. But TextBlob wraps it up nicely for you.

When you use the textblob.en.Spelling class you have some choices. First, like I did in my demo:

path = "spelling-model.txt"
spelling = Spelling(path=path)
spelling.train(my_space_separated_text_blob, path)

What that does is creating a file spelling-model.txt that wasn't there before. It looks like this (in my demo):

▶ head spelling-model.txt
aaron 1
abel 1
adam 1
adrian 1
aiden 1
alexander 1
andrew 1
angel 1
anthony 1
asher 1

The number (on the right) there is the "frequency" of the word. But what if you have a "scoring" number of your own. Perhaps, in your application you just know that adrian is more right than damian. Then, you can make your own file:

Suppose the text file ("spelling-model-weighted.txt") contains lines like this:

adrian 8
damian 3

Now, the output becomes:

>>> import os
>>> from textblob.en import Spelling
>>> import os
>>> path = "spelling-model-weighted.txt"
>>> assert os.path.isfile(path)
>>> spelling = Spelling(path=path)
>>> spelling.suggest('darian')
[('adrian', 0.7272727272727273), ('damian', 0.2727272727272727)]

Based on the weighting, these numbers add up. I.e. 3 / (3 + 8) == 0.2727272727272727

I hope it inspires you to write your own spelling application using TextBlob.

For example, you can feed it the names of your products on an e-commerce site. The .txt file might bloat if you have too much but note that the 30K lines en-spelling.txt is only 314KB and it loads in...:

>>> from textblob import TextBlob
>>> from time import perf_counter
>>> b = TextBlob("I havv goood speling!")
>>> t0 = perf_counter(); right = b.correct() ; t1 = perf_counter()
>>> t1 - t0

...70ms for 30,000 words.

SongSearch autocomplete rate now 2+ per second

11 July 2019 0 comments   Redis, Nginx, Python, Django

By analyzing my Nginx logs, I've concluded that SongSearch's autocomplete JSON API now gets about 2.2 requests per second. I.e. these are XHR requests to /api/search/autocomplete?q=....

Roughly, 1.8 requests per second goes back to the Django/Elasticsearch backend. That's a hit ratio of 16%. These Django/Elasticsearch requests take roughly 200ms on average. I suspect about 150-180ms of that time is spent querying Elasticsearch, the rest being Python request/response and JSON "paperwork".

Autocomplete counts in Datadog

Caching strategy

Caching is hard because the queries are so vastly different over time. Had I put a Redis cache decorator on the autocomplete Django view function I'd quickly bloat Redis memory and cause lots of evictions.

What I used to do was something like this:

def search_autocomplete(request):
   q = request.GET.get('q') 

   cache_key = None
   if len(q) < 10:
      cache_key = 'autocomplete:' + q
      results = cache.get(cache_key)
      if results is not None:
          return http.JsonResponse(results)

   results = _do_elastisearch_query(q)
   if cache_key:
       cache.set(cache_key, results, 60 * 60)

   return http.JsonResponse(results)   

However, after some simple benchmarking it was clear that using Nginx' uwsgi_cache it was much faster to let the cacheable queries terminate already at Nginx. So I changed the code to something like this:

def search_autocomplete(request):
   q = request.GET.get('q') 
   results = _do_elastisearch_query(q)
   response = http.JsonResponse(results)   

   if len(q) < 10:
       patch_cache_control(response, public=True, max_age=60 * 60)

   return response

The only annoying thing about Nginx caching is that purging is hard unless you go for that Nginx Plus (or whatever their enterprise version is called). But more annoying, to me, is that fact that I can't really see what this means for my server. When I was caching with Redis I could just use redis-cli and...

# Memory

Nginx Amplify

My current best tool for keeping an eye on Nginx is Nginx Amplify. It gives me some basic insights about the state of things. Here are some recent screenshots:

NGINX Requests/s

NGINX Memory Usage


Thoughts and conclusion

Caching is hard. But it's also fun because it ties directly into performance work.

In my business logic, I chose that autocomplete queries that are between 1 and 9 characters are cacheable. And I picked a TTL of 60 minutes. At this point, I'm not sure exactly why I chose that logic but I remember doing some back-of-envelope calculations about what the hit ratio would be and roughly what that would mean in bytes in RAM. I definitely remember picking 60 minutes because I was nervous about bloating Nginx's memory usage. But as of today, I'm switching that up to 24 hours and let's see what that does to my current 16% Nginx cache hit ratio. At the moment, /var/cache/nginx-cache/ is only 34MB which isn't much.

Another crux with using uwsgi_cache (or proxy_cache) is that you can't control the cache key very well. When it was all in Python I was able to decide about the cache key myself. A plausible implementation is cache_key = q.lower().strip() for example. That means you can protect your Elasticsearch backend from having to do {"q": "A"} and {"q": "a"}. Who knows, perhaps there is a way to hack this in Nginx without compiling in some Lua engine.

The ideal would be some user-friendly diagnostics tool that I can point somewhere, towards Nginx, that says how much my uwsgi_cache is hurting or saving me. Autocomplete is just one of many things going on on this single DigitalOcean server. There's also a big PostgreSQL server, a node-express cluster, a bunch of uwsgi workers, Redis, lots of cron job scripts, and of course a big honking Elasticsearch 6.

UPDATE (July 12 2019)

Currently, and as mentioned above, I only set Cache-Control headers (which means Nginx snaps it up) for queries that at max 9 characters long. I wanted to appreciate and understand how ratio of all queries are longer than 9 characters so I wrote a report and its output is this:

Sum show 75646 32.2%
Sum rest 159321 67.8%

Sum show 83702 35.6%
Sum rest 151265 64.4%

Sum show 90870 38.7%
Sum rest 144097 61.3%

Sum show 98384 41.9%
Sum rest 136583 58.1%

Sum show 106093 45.2%
Sum rest 128874 54.8%

Sum show 113905 48.5%
Sum rest 121062 51.5%

It means that (independent of time expiry) 38.7% of queries are 9 characters or less.

Build an XML sitemap of XML sitemaps

01 June 2019 0 comments   Python, Django

Suppose that you have so many thousands of pages that you can't just create a single /sitemap.xml file that has all the URLs (aka <loc>) listed. Then you need to make a /sitemaps.xml that points to the other sitemap files. And if you're in the thousands, you'll need to gzip these files.

The blog post demonstrates how Song Search generates a sitemap file that points to 63 sitemap-{M}-{N}.xml.gz files which spans about 1,000,000 URLs. The context here is Python and the getting of the data is from Django. Python is pretty key here but if you have something other than Django, you can squint and mentally replace that with your own data mapper.

Generate the sitemap .xml.gz file(s)

Here's the core of the work. A generator function that takes a Django QuerySet instance (that is ordered and filtered!) and then starts generating etree trees and dumps them to disk with gzip.

import gzip

from lxml import etree

outfile = "sitemap-{start}-{end}.xml"
batchsize = 40_000

def generate(self, qs, base_url, outfile, batchsize):
    # Use `.values` to make the query much faster
    qs = qs.values("name", "id", "artist_id", "language")

    def start():
        return etree.Element(
            "urlset", xmlns=""

    def close(root, filename):
        with, "wb") as f:
            f.write(b'<?xml version="1.0" encoding="utf-8"?>\n')
            f.write(etree.tostring(root, pretty_print=True))

    root = filename = None

    count = 0
    for song in qs.iterator():
        if not count % batchsize:
            if filename:  # not the very first loop
                close(root, filename)
                yield filename
            filename = outfile.format(start=count, end=count + batchsize)
            root = start()
        loc = "{}{}".format(base_url, make_song_url(song))
        etree.SubElement(etree.SubElement(root, "url"), "loc").text = loc
        count += 1
    close(root, filename)
    yield filename

The most important lines in terms of lxml.etree and sitemaps are:

root = etree.Element("urlset", xmlns="")
etree.SubElement(etree.SubElement(root, "url"), "loc").text = loc

Another important thing is the note about using .values() . If you don't do that Django will create a model instance for every single row it returns of the iterator. That's expensive. See this blog post.

Another important thing is to use a Django ORM iterator as that's much more efficient than messing around with limits and offsets.

Generate the map of sitemaps

Making the map of maps doesn't need to be gzipped since it's going to be tiny.

def generate_map_of_maps(base_url, outfile):
    root = etree.Element(
        "sitemapindex", xmlns=""

    with open(outfile, "wb") as f:
        f.write(b'<?xml version="1.0" encoding="UTF-8"?>\n')
        files_created = sorted(glob("sitemap-*.xml.gz"))
        for file_created in files_created:
            sitemap = etree.SubElement(root, "sitemap")
            uri = "{}/{}".format(base_url, os.path.basename(file_created))
            etree.SubElement(sitemap, "loc").text = uri
            lastmod = datetime.datetime.fromtimestamp(
            etree.SubElement(sitemap, "lastmod").text = lastmod
        f.write(etree.tostring(root, pretty_print=True))

And that sums it up. On my laptop, it takes about 60 seconds to generate 39 of these files (e.g. sitemap-1560000-1600000.xml.gz) and that's good enough.

Bonus and Thoughts

The bad news is that this is about as good as it gets in terms of performance. The good news is that there are no low-hanging fruit fixes. I know, because I tried. I experimented with not using pretty_print=True and I experimented with not writing with and instead gzipping the files on later. Nothing made any significant difference. The lxml.etree part of this, in terms of performance, is order of maginitude marginal in comparison to the cost of actually getting the data out of the database plus later writing to disk. I also experimenting with generating the gzip content with zopfli and it didn't make much of a difference.

I originally wrote this code years ago and when I did, I think I knew more about sitemaps. In my implementation I use a batch size of 40,000 so each file is called something like sitemap-40000-80000.xml.gz and weighs about 800KB. Not sure why I chose 40,000 but perhaps not important.

Generate a random IP address in Python

01 June 2019 0 comments   Python, Django

I have a commenting system where people can type in a comment and optionally their name and email if they like.
In production, where things are real, the IP address that can be collected are all interestingly different. But when testing this manually on my laptop, since the server is running http://localhost:8000, the request.META.get('REMOTE_ADDR') always becomes Boring! So I fake it. Like this:

import random
from ipaddress import IPv4Address

def _random_ip_address(seed):
    return str(IPv4Address(random.getrandbits(32)))

# Here's the code deep inside the POST handler just before storing 
# the form submission the database.

if settings.DEBUG and metadata.get("REMOTE_ADDR") == "":
    # Make up a random one!
    metadata["REMOTE_ADDR"] = _random_ip_address(
        str(form.cleaned_data["name"]) + str(form.cleaned_data["email"])

It's pretty rough but it works and makes me happy.

How I simulate a CDN with Nginx

15 May 2019 1 comment   Nginx, Python

Usually, a CDN is just a cache you put in front of a dynamic website. You set up the CDN to be the first server your clients get data from, the CDN quickly decides if it was a copy cached or otherwise it asks the origin server for a fresh copy. So far so good, but if you really care about squeezing that extra performance out you need to worry about having a decent TTL and as soon as you make the TTL more than a couple of minutes you need to think about cache invalidation. You also need to worry about preventing certain endpoints from ever getting caught in the CDN which could be very bad.

For this site,, I'm using KeyCDN which I've blogged out here: "I think I might put my whole site behind a CDN" and here: "KeyCDN vs. DigitalOcean Nginx". KeyCDN has an API and a python client which I've contributed to.

The next problem is; how do you test all this stuff on your laptop? Unfortunately, you can't deploy a KeyCDN docker image or something like that, that attempts to mimic how it works for reals. So, to simulate a CDN locally on my laptop, I'm using Nginx. It's definitely pretty different but it's not the point. The point is that you want something that acts as a reverse proxy. You want to make sure that stuff that's supposed to be cached gets cached, stuff that's supposed to be purged gets purged and that things that are always supposed to be dynamic is always dynamic.

The Configuration

First I add peterbecom.local into /etc/hosts like this:

▶ cat /etc/hosts | grep peterbecom.local       peterbecom.local origin.peterbecom.local
::1             peterbecom.local origin.peterbecom.local

Next, I set up the Nginx config (running on port 80) and the configuration looks like this:

proxy_cache_path /tmp/nginxcache  levels=1:2    keys_zone=STATIC:10m
    inactive=24h  max_size=1g;

server {
    server_name peterbecom.local;
    location / {
        proxy_cache_bypass $http_secret_header;
        add_header X-Cache $upstream_cache_status;
        proxy_set_header x-forwarded-host $host;
        proxy_cache STATIC;
        # proxy_cache_key $uri;
        proxy_cache_valid 200  1h;
        proxy_pass http://origin.peterbecom.local;
    access_log /tmp/peterbecom.access.log combined;
    error_log /tmp/peterbecom.error.log info;

By the way, I've also set up origin.peterbecom.local to be run in Nginx too but it could just be proxy_pass http://localhost:8000; to go straight to Django. Not relevant for this context.

The Purge

Without the commercial version of Nginx (Plus) you can't do easy purging just for purging sake. But with proxy_cache_bypass $http_secret_header; it's very similar to purging except that it immediately makes a request to the origin.

First, to test that it works, I start up Nginx and Django and now I can run:

▶ curl -v http://peterbecom.local/about > /dev/null
< HTTP/1.1 200 OK
< Server: nginx/1.15.10
< Cache-Control: public, max-age=3672
< X-Cache: MISS

(Note the X-Cache: MISS which comes from add_header X-Cache $upstream_cache_status;)

This should trigger a log line in /tmp/peterbecom.access.log and in the Django runserver foreground logs.

At this point, I can kill the Django server and run it again:

▶ curl -v http://peterbecom.local/about > /dev/null
< Server: nginx/1.15.10
< HTTP/1.1 200 OK
< Cache-Control: max-age=86400
< Cache-Control: public
< X-Cache: HIT

Cool! It's working without Django running. As expected. This is how to send a "purge request"

▶ curl -v -H "secret-header:true" http://peterbecom.local/about > /dev/null
> GET /about HTTP/1.1
> secret-header:true
< HTTP/1.1 502 Bad Gateway

Clearly, it's trying to go to the origin, which was killed, so you start that up again and you get back to:

▶ curl -v http://peterbecom.local/about > /dev/null
< HTTP/1.1 200 OK
< Server: nginx/1.15.10
< Cache-Control: public, max-age=3672
< X-Cache: MISS

In Python

In my site, there are Django signals that are triggered when a piece of content changes and I'm using python-keycdn-api in production but obviously, that won't work with Nginx. So I have a local setting and my Python code looks like this:

# This function gets called by a Django `post_save` signal
# among other things such as cron jobs and management commands.

def purge_cdn_urls(urls):
    if settings.USE_NGINX_BYPASS:
        # Note! This Nginx trick will not just purge the proxy_cache, it will
        # immediately trigger a refetch.
        x_cache_headers = []
        for url in urls:
            if "://" not in url:
                url = settings.NGINX_BYPASS_BASEURL + url
            r = requests.get(url, headers={"secret-header": "true"})
            x_cache_headers.append({"url": url, "x-cache": r.headers.get("x-cache")})
        print("PURGED:", x_cache_headers)

    ...the stuff that uses keycdn...

Notes and Conclusion

One important feature is that my CDN is a CNAME for but it reaches the origin server on a different URL. When my Django code needs to know the outside facing domain, I need to respect that. The communication between by the CDN and my origin is a domain I don't want to expose. What KeyCDN does is that they send an x-forwarded-host header which I need to take into account when understanding what outward facing absolute URL was used. Here's how I do that:

def get_base_url(request):
    base_url = ["http"]
    if request.is_secure():
    x_forwarded_host = request.headers.get("X-Forwarded-Host")
    if x_forwarded_host and x_forwarded_host in settings.ALLOWED_HOSTS:
    return "".join(base_url)

That's about it. There are lots of other details I glossed over but the point is that this works good enough to test that the cache invalidation works as expected.

Whatsdeployed rewritten in React

15 April 2019 0 comments   Javascript, ReactJS, Python, Web development

A couple of months ago my colleague Michael @mythmon Cooper wanted to add a feature to the front-end code of Whatsdeployed and learned that the whole front-end is spaghetti jQuery code. So, instead, he re-wrote it in React. My only requirements were "Use create-react-app and no redux", i.e. keep it simple.

We also took the opportunity to rewrite some of the ways that URLs are handled. It used to be that a "short link" would redirect. For example GET /s-5HY would return 302 to Location: ?org=mozilla&repo=tecken&name[]=Dev&url[]=[]=Stage... Basically, the short link was just an alias for a redirect. Just like those services like or Now, the short link is a permanent fixture. The short link is included in the XHR calls to the server for getting the relevant data.

All old URLs will continue to work but now the canonical URL becomes /s/5HY/mozilla-services/tecken , for example. The :org/:repo isn't really necessary because the server knows exactly what 5HY (in this example means), but it's nice for the URL bar's memory.

Another thing that changed was how it can recognize "bors commits". When you use bors, you put a bunch of commits into a GitHub Pull Request and then ask the bors bot to merge them into master. Using "bors mode" in Whatsdeployed is optional but we believe it looks a lot more user-friendly. Here is an example of mozilla/normandy with and without bors toggled on and off.

Without "bors mode"
Without "bors mode"

With "bors mode"
With "bors mode"

Thank you mythmon!

Lastly, hopefully this will make it a lot easier to contribute. Check out All you need is Python 3, a PostgreSQL, and almost any version of Node that can run create-react-apps. Ping me if you find it hard to get up and running.