7. Production settings

So far the only thing we’ve done in our production settings was to setup ALLOWED_HOSTS. We still have some work to do. It is absolutely essential to setup email and the secret key, it is a good idea to setup logging, and we may also need to setup caching. Most installations will not need anything beyond these.

7.1. Email

Even if your Django application does not use email at all, you must still set it up. The reason is that your code has bugs. Even if it does not have bugs, your server will eventually run into an error condition, such as no disk space, out of memory, or something else going wrong. In many of these cases, Django will throw a “500 error” to the user and will try to email you. You really need to receive that email.

First, you need a mail server to which you can connect and ask to send an email. Such a mail server is called a “smarthost”. The mechanism with which Django connects to the smarthost is pretty much the same as the one with which your desktop or mobile mail client connects to an outgoing mail server. However, the term “outgoing mail server” is mostly used for mailing software, and “smarthost” is used when some unattended software like your Django app sends email. You can often, but not always, use your outgoing mail server as smarthost.

I’m using Runbox for my email, and I also use it as a smarthost. There are many other providers, one of the most popular being Gmail (I believe, however, that it’s not possible to use Gmail as a smarthost if all you have is a free account, and even if it is possible, it is hard to setup).

Let’s set it up and then we will discuss more. Add the following to /etc/opt/$DJANGO_PROJECT/settings.py:

SERVER_EMAIL = 'noreply@$DOMAIN'
DEFAULT_FROM_EMAIL = 'noreply@$DOMAIN'
ADMINS = [
    ('$ADMIN_NAME', '$ADMIN_EMAIL_ADDRESS'),
]
MANAGERS = ADMINS

EMAIL_HOST = '$EMAIL_HOST'
EMAIL_HOST_USER = '$EMAIL_HOST_USER'
EMAIL_HOST_PASSWORD = '$EMAIL_HOST_PASSWORD'
EMAIL_PORT = 587
EMAIL_USE_TLS = True

SERVER_EMAIL is the email address from which emails with error messages appear to come from. It is set in the “From:” field of the email. The default is “root@localhost”, and while “root” is OK, “localhost” is not, and some mail servers may refuse the email. The domain name where your Django application runs is usually OK, but if this doesn’t work you can use any other valid domain. The domain of your email address should work properly.

If your Django project does not send any emails (other than the error messages Django will send anyway), DEFAULT_FROM_EMAIL does not need to be specified. If it does send emails, it may be using django.core.mail.EmailMessage. In order to specify what will be in the “From:” field of the email, EmailMessage accepts a from_email argument at initialization; if this is unspecified, it will use DEFAULT_FROM_EMAIL. So DEFAULT_FROM_EMAIL is exactly what it says: the default from_email of EmailMessage. It’s a good idea to specify this, because even if your Django project does not send emails today, it may well do so tomorrow, and the default, “webmaster@localhost”, is not a good option. Remember that with EmailMessage you are likely to send email to your users, and it should be something nice. “noreply@$DOMAIN” is usually fine.

ADMINS is a list of people to whom error messages will be sent. Make sure your name and email address are listed there, and also add any fellow administrators. MANAGERS is similar to ADMINS, but for broken link notifications, and usually you just need to set it to the same values as ADMINS.

The settings starting with EMAIL_ describe how Django will connect and authenticate to the mail server. Django will connect to EMAIL_HOST and authenticate using EMAIL_HOST_USER and EMAIL_HOST_PASSWORD. Needless to say, I have used placeholders that start with a dollar sign, and you need to replace these with actual values. Mine are usually these:

EMAIL_HOST = 'mail.runbox.com'
EMAIL_HOST_USER = 'smarthostclient%antonischristofides.com'
EMAIL_HOST_PASSWORD = 'topsecret'

However, the details depend on the provider and the account type you have. I don’t use my personal email, which is antonis@antonischristofides.com (Runbox requires you to change @ to % when you use it as a user name for login), because my personal password would then be in many settings.py files in many deployed Django projects, and I’m not the only administrator of these servers (and even if I were, I wouldn’t know when I would invite another one). So I created another user (subaccount in Runbox parlance), “smarthostclient”, which I use for that purpose.

There are three ports used for sending email: 25, 465, and 587. The sender (Django in our case, or your mail client when you send email) connects to a mail server and gives the email to it; the mail server then delivers the email to another mail server, and so on, until the destination is reached. In the old times both the initial submission and the communication between mail servers was through port 25. Nowadays 25 is mostly used for communication between mail servers only. If you try to use port 25 (which is the default setting for EMAIL_PORT), it’s possible that the request will get stuck in firewalls, and even if does reach the mail server, the mail server is likely to refuse to send the email. This is because spam depends much on port 25, so policies about this port are very tight.

The other two ports for email submission are 465 and 587. 465 uses encryption; just as 80 is for unencrypted HTTP and 443 is for encrypted HTTP, 25 is for unencrypted SMTP and 465 is for encrypted SMTP. However, 465 is deprecated in favour of 587, which can handle both unencrypted and encrypted connections. The client (Django in our case) connects to the server at port 587, they start talking unencrypted, and the client may tell the server “I want to continue with encryption”, and then they continue with encryption. Obviously this is done before authentication, which requires the password to be transmitted.

There are thus two methods to start encryption; one is implicit and the other one is explicit. When you connect to port 465, which always works encrypted, the encryption starts implicitly. When you connect to port 587, the two peers (the client and the server) start talking unencrypted, and at some point the client explicitly tells the server “I want to continue with encryption”. Computer people often use “SSL” for implicit encryption and “TLS” for explicit, however this is inaccurate; SSL and TLS are encryption protocols, and do not refer to the method used to initiate them; you could have implicit TLS or explicit SSL. Django uses this inaccurate parlance in its settings, where EMAIL_USE_TLS and EMAIL_USE_SSL are used to specify whether, respectively, the connection will use explicit or implicit encryption. EMAIL_USE_TLS = True should be used with EMAIL_PORT = 587, and EMAIL_USE_SSL = True with EMAIL_PORT = 465.

To test your settings, start a shell from your Django project:

PYTHONPATH=/etc/opt/$DJANGO_PROJECT:/opt/$DJANGO_PROJECT \
DJANGO_SETTINGS_MODULE=settings \
su $DJANGO_USER -c \
"/opt/$DJANGO_PROJECT/venv/bin/python \
/opt/$DJANGO_PROJECT/manage.py shell"

and enter these commands:

from django.conf import settings
from django.core.mail import send_mail

admin_emails = [x[1] for x in settings.ADMINS]
send_mail("Test1557", "Hello", settings.SERVER_EMAIL,
          admin_emails)

If something goes wrong, send_mail will raise an exception; otherwise you should receive the email.

Because of spam, mail servers are often very picky about which emails they will accept. It’s possible that even if your smarthost accepts the email, the next mail server may refuse it. For example, I made some experiments using from_email='noreply@example.com', EMAIL_HOST = 'mail.runbox.com', and recipient anthony@itia.ntua.gr (an old email address of mine). In that case, Runbox accepted the email and subsequently attempted to deliver it to the mail server of ntua.gr, which rejected it because it didn’t like the sender (noreply@example.com; I literally used “example.com”, and ntua.gr didn’t like that domain). When something like this happens, the test we made above with send_mail will appear to work, because send_mail manages to deliver the email to the smarthost, and the error occurs after that; not only will we never receive the email, but it is also likely that we will not receive the failure notification (the returned email), so it’s often hard to know what went wrong and we need to guess.

One thing you can do to lessen the probability of error is to make sure that the recipient (or at least one of the recipients) has an email address served by the provider who provides the smarthost. In my case, the smarthost is mail.runbox.com, and the recipient is antonis@antonischristofides.com, and the email for domain antonischristofides.com is served by Runbox. It is unlikely that mail.runbox.com would accept an email addressed to antonis@antonischristofides.com if another Runbox server were to subsequently refuse it. If something like this happened, I believe it would be a configuration error on behalf of Runbox. But it’s very normal that mail.runbox.com will accept an email which will subsequently be refused by ntua.gr or Gmail or another provider downstream.

7.2. Debug

After you have configured email and verified it works, you can now turn off DEBUG:

DEBUG = False

Now it’s good time to verify that error emails do indeed get sent properly. You can do so by deliberately causing an internal server error. A favourite way of mine is to temporarily rename a template file and make a related request, which will raise a TemplateDoesNotExist exception. Your browser should show the “server error” page. Don’t forget to rename the template file back to what it was. By the time you finish doing that, you should have received the email with the full trace.

7.3. Using a local mail server

Usually I don’t configure Django to deliver to the smarthost; instead, I install a mail server locally, have Django deliver to the local mail server, and configure the local mail server to send the emails to the smarthost. There are several reasons why installing a local mail server is better:

  1. Your server, like all Unix systems, has a scheduler, cron, which is configured to run certain programs at certain times. For example, directory /etc/cron.daily contains scripts that are executed once per day. Whenever a program run by cron throws an error message, cron emails that error message to the administrator. cron always works with a local mail server. If you don’t install a local mail server, you will miss these error messages. We will later use cron to clear sessions and to backup the server, and we don’t want to miss any error messages.
  2. While Django attempts to send an error email, if something goes wrong, it fails silently. This behaviour is appropriate (the system is in error, it attempts to email its administrator with the exception, but sending the email also results in an error; can’t do much more). Suppose, however, that when you try to verify, as we did in the previous section, that error emails work, you find out they don’t work. What has gone wrong? Nothing is written in any log. Intercepting the communication with ngrep won’t work either, because it’s usually encrypted. If you use a locally installed mail server, you will at least be able to look at the local mail server’s logs.
  3. Sending an error email might take long. The communication line might be slow, or a firewall or the DNS could be misbehaving, and it might take several seconds, or even a minute, before Django manages to establish a connection to the remote mail server. During this time, the browser will be in a waiting state, and a Gunicorn process will be occupied. Some people will recommend to send emails from celery workers, but this is not possible for error emails. In addition, there is no reason to install and program celery just for this reason. If we use a local mail server, Django will deliver the email to it very fast and finish its job, and the local mail server will queue it and send it when possible.

While the most popular mail servers for Debian and Ubuntu are exim and postfix, I don’t recommend them. Mail servers are strange beasts. They have large and tricky configuration files, because they can do a hell of things. You will have a hard time understanding the necessary configuration (which is buried under a hell of other configuration), and if something goes wrong you will have a hard time debugging it. I also see no great educational value in learning it. I used to run mail servers for years but I’ve got ridden of all of them; it’s not worth the effort when I can do the same thing at Runbox for € 30 per year.

Instead, we are going to use dma (nothing to do with direct memory access; this is the DragonFly Mail Agent). It’s a small mail server that only does what we want; it collects messages in a queue, and sends them to a smarthost. It is much easier to configure than the real thing. Install it like this:

apt install dma

It will ask you a couple of questions:

System mail name
You should probably use $DOMAIN here. If that doesn’t work, you can try to use the domain of your email address.
Smarthost
This is the remote mail server, the smarthost, that is; the one we had specified in Django’s EMAIL_HOST.

Next, open /etc/dma/dma.conf in an editor, and uncomment or edit these directives:

PORT 587
AUTHPATH /etc/dma/auth.conf
SECURETRANSFER
STARTTLS

(If your smarthost uses implicit encryption, you need to specify PORT 465 instead, and omit the STARTTLS.)

Next, open /etc/dma/auth.conf and add this line:

$EMAIL_USER|$EMAIL_HOST:$EMAIL_PASSWORD

(These are placeholders of course, which you need to replace.)

Next, open /etc/aliases and add this line:

root: $ADMIN_EMAIL_ADDRESS

Finally, open /etc/mailname in an editor and make sure it contains a single line which contains your domain ($DOMAIN).

Let’s test it to see if it works:

sendmail $ADMIN_EMAIL_ADDRESS

This will pause for input. Type a short email message, and end it with a line that contains a single fullstop. Check /var/log/mail.log to verify it has been delivered to the smarthost (if it says “delivery successful” it’s OK, even if it’s preceded by a warning message about the authentication mechanism), and verify that you have received it.

The next step is to configure Django. You might think that we would set EMAIL_HOST = 'localhost' and EMAIL_PORT = 25, but this is not what we will do. dma does not listen on port 25 or on any other port. The only way to send emails with it is by using the sendmail command. Traditionally this has been the easiest and most widely available way to send emails in Unix, and it is also what cron uses. (In the old times, when sendmail was the only existing mail server, the practice of using the sendmail command was standardized, so today all mail servers create a sendmail command when they are installed, which is usually a symbolic link to something else). We will install a Django email backend that sends emails in the same way.

/opt/$DJANGO_PROJECT/venv/bin/pip install django-sendmail-backend

The only Django configuration we need is this:

EMAIL_BACKEND = 'django_sendmail_backend.backends.EmailBackend'

The dma configuration should have been obvious, except for /etc/aliases and /etc/mailname. These are not dma-specific, they are also used by exim, postfix, and most other mail servers, and /etc/mailname may also be used by other programs.

/etc/aliases specifies aliases for email addresses. If cron decides it needs to send an email, the recipient will most likely be a mere root. The line we added specifies that root should be translated to your actual email address. For Django, /etc/aliases doesn’t matter, since Django will get the recipient email address from the ADMINS and MANAGERS settings.

If a program somehow needs to know the domain used for the email of the system, it usually takes it from /etc/mailname. Setting that to $DOMAIN should be fine, but if this doesn’t work, you can try setting it to the domain of your email address.

7.4. Secret key

Django uses the SECRET_KEY in several cases, for example, when digitally signing sessions in cookies. If it leaks, attackers might be able to compromise your system. You should not use the SECRET_KEY you use in development, because that one is easy to leak, and because many developers often have access to it, whereas they should not have access to the production SECRET_KEY.

You can create a secret key in this way:

import sys

from django.utils.crypto import get_random_string

sys.stdout.write(get_random_string(50))

7.5. Logging

Even if your Django apps do no logging, they eventually will. At some point one of your users is going to cause an error which you will be unable to reproduce in the development environment, so you will introduce some logging calls. It makes sense to configure logging so that it is ready for that time. You need a configuration that will write log messages in /var/log/$DJANGO_PROJECT/$DJANGO_PROJECT.log, and here it is:

LOGGING = {
    'version': 1,
    'disable_existing_loggers': False,
    'formatters': {
        'default': {
            'format': '[%(asctime)s] %(levelname)s: '
                      '%(message)s',
        }
    },
    'handlers': {
        'file': {
            'class': 'logging.handlers.'
                     'TimedRotatingFileHandler',
            'filename': '/var/log/$DJANGO_PROJECT/'
                        '$DJANGO_PROJECT.log',
            'when': 'midnight',
            'backupCount': 60,
            'formatter': 'default',
        },
    },
    'root': {
        'handlers': ['file'],
        'level': 'INFO',
    },
}

Here is the meaning of the various items:

version
This is reserved for the future; for now, it should always be 1.
disable_existing_loggers
Django already has a default logging configuration. If disable_existing_loggers is True (the default), then this configuration will override Django’s default, otherwise it will work in addition to the default. We really want Django’s default configuration, which is to email critical errors to the administrators.
root
This defines the root logger. You can specify very complicated logging schemes, where different loggers will be logging using different handlers and different formatters. However, as long as our system is small, we only need to specify a single logger, the root logger, which uses a single handler (the “file” handler) with a single formatter (the “default” formatter). In this example I have specified 'level': 'INFO', which means the logger will ignore messages with a lower priority (the only lower priority is DEBUG, and the higher priorities are WARNING, ERROR and CRITICAL). You can change this as needed, however INFO is reasonable to begin with.
handlers
Here we define the “file” handler, whose class is logging.TimedRotatingFileHandler. This essentially logs to a file, but it has the added benefit that each midnight it starts a new log file, renames the old one, and deletes log files older than 60 days. In this way it is very unlikely that your disk will fill up because of the growing log files escaping your attention.
formatters

This defines a formatter named “default”. In a system where I’m using this logging configuration, I have this code:

import logging

# ...

logging.info('Notifying user {} about the agrifields of '
             'user {}'.format(user, owner))

and it produces this line in the log file:

[2016-11-29 04:40:02,880] INFO: Notifying user aptiko about the agrifields of user aptiko

7.6. Caching

The only other setting I expect you to set to a different value from development is CACHES. How you will set it depends on your needs. I usually want my caches to persist across reboots, so I specify this:

CACHES = {
    'default': {
        'BACKEND': 'django.core.cache.backends.filebased.'
                   'FileBasedCache',
        'LOCATION': '/var/cache/$DJANGO_PROJECT/cache',
    }
}

You also need to create the directory and give it the necessary permissions:

mkdir /var/cache/$DJANGO_PROJECT/cache
chown $DJANGO_USER /var/cache/$DJANGO_PROJECT/cache

7.7. Recompile your settings

Remember that Django runs as $DJANGO_USER and does not (and should not) have permission to write in directory /etc/opt/$DJANGO_PROJECT, which is owned by root. Therefore it can’t write the Python 2 compiled file settings.pyc, or the Python 3 compiled files directory __pycache__. In theory you should be compiling it each time you make a change to your settings:

/opt/$DJANGO_PROJECT/venv/bin/python -m compileall \
    /etc/opt/$DJANGO_PROJECT

Of course it’s not possible to remember to do this every single time you change something in the settings. There are two solutions to this. The first solution, which is fine, is to ignore the problem. If the compiled file is absent or outdated, Python will compile the source file on the spot. This will happen whenever each gunicorn worker starts, which is only when you start or restart gunicorn, and it costs less than 1 ms. It’s really negligible.

The second solution is to create a script /usr/local/sbin/restart-$DJANGO_PROJECT, with the following contents:

#!/bin/bash
set -e
/opt/$DJANGO_PROJECT/venv/bin/python -m compileall -q \
     -x /opt/$DJANGO_PROJECT/venv/ /opt/$DJANGO_PROJECT \
     /etc/opt/$DJANGO_PROJECT
service $DJANGO_PROJECT restart

You must make that script executable:

chmod 755 /usr/local/sbin/restart-$DJANGO_PROJECT

You might object that we don’t want users other than root to be able to recompile the Python files or to restart the gunicorn service. The answer is that they won’t be able. They will be able to execute the script, but when the script arrives at the point where it compiles the Python files, they will be denied permission to write the compiled Python files to the directory; and if the script ever arrives at the last line, again systemd will deny to restart the service. Making a script non-executable doesn’t achieve anything security-wise; a malicious user could simply copy it and make the copy executable.

From now on, whenever you want to restart gunicorn, instead of service $DJANGO_PROJECT restart, you can be using restart-$DJANGO_PROJECT, which will run the above script. The set -e command tells bash to stop executing the script when an error occurs, and the -q parameter to compileall tells to not print the list of files compiled.

7.8. Clearing sessions

If you use django.contrib.sessions, Django stores session data in the database (unless you use using a different SESSION_ENGINE). Django does not automatically clean up the sessions table, so most of the sessions remain in the database even after they expire. I’ve seen sessions tables in small deployments of only a few requests per minute grow to several hundreds of GB through the years. You can manually remove expired sessions by executing python manage.py clearsessions.

To make sure your sessions are being cleared regularly, create file /etc/cron.daily/$DJANGO_PROJECT-clearsessions with the following contents:

#!/bin/bash
export PYTHONPATH=/etc/opt/$DJANGO_PROJECT:/opt/$DJANGO_PROJECT
export DJANGO_SETTINGS_MODULE=settings
su $DJANGO_USER -c "/opt/$DJANGO_PROJECT/venv/bin/python \
    /opt/$DJANGO_PROJECT/manage.py clearsessions"

Make the file executable:

chmod 755 /etc/cron.daily/$DJANGO_PROJECT-clearsessions

In Unix-like systems, cron is the standard scheduler; it executes tasks at specified times. Scripts in /etc/cron.daily are executed once daily, starting at 06:25 (am) local time. The time to which this actually refers depends on the system’s time zone, which you can find by examining the contents of the file /etc/timezone. In most of my servers, I use UTC. The time during which these scripts are run doesn’t really matter much, but it’s better to do it when the system is not very busy—especially if some of the scripts are intensive, such as backup (which we will see in a later chapter). For time zones with a positive UTC offset, 06:25 UTC could be a busy time, so you might want to change the system time zone with this command:

dpkg-reconfigure tzdata

There is a way to tell cron exactly at what time you want a task to run, but I won’t go into that as throwing stuff into /etc/cron.daily should be sufficient for most use cases.

Cron expects all the programs it runs to be silent, i.e., to not display any output. If they do display output, cron emails that output to the administrator. This is very neat, because if your tasks only display output when there is an error, you will be emailed only when there is an error. However, for this to work, you must setup a local mail server as explained in Using a local mail server.

7.9. Chapter summary

  • Install dma and (in the virtualenv) django-sendmail-backend

  • Make sure /etc/dma/dma.conf has these contents:

    SMARTHOST $EMAIL_HOST
    PORT 587
    AUTHPATH /etc/dma/auth.conf
    SECURETRANSFER
    STARTTLS
    MAILNAME /etc/mailname
    

    Also make sure /etc/dma/auth.conf has these contents:

    $EMAIL_HOST_USER|$EMAIL_HOST:$EMAIL_HOST_PASSWORD
    

    Make sure /etc/mailname contains $DOMAIN.

  • Create the cache directory:

    mkdir /var/cache/$DJANGO_PROJECT/cache
    chown $DJANGO_USER /var/cache/$DJANGO_PROJECT/cache
    
  • Create file /etc/cron.daily/$DJANGO_PROJECT-clearsessions with the following contents:

    #!/bin/bash
    export PYTHONPATH=/etc/opt/$DJANGO_PROJECT:/opt/$DJANGO_PROJECT
    export DJANGO_SETTINGS_MODULE=settings
    su $DJANGO_USER -c "/opt/$DJANGO_PROJECT/venv/bin/python \
        /opt/$DJANGO_PROJECT/manage.py clearsessions"
    

    Make the file executable:

    chmod 755 /etc/cron.daily/$DJANGO_PROJECT-clearsessions
    
  • Finally, this is the whole settings.py file:

     from django_project.settings import *
    
     debug = false
     allowed_hosts = ['$domain', 'www.$domain']
     databases = {
         'default': {
             'engine': 'django.db.backends.sqlite3',
             'name': '/var/opt/$django_project/$django_project.db',
         }
     }
    
     server_email = 'noreply@$domain'
     default_from_email = 'noreply@$domain'
     admins = [
         ('$admin_name', '$admin_email_address'),
     ]
     managers = admins
     email_backend = 'django_sendmail_backend.backends.' \
                     'emailbackend'
    
     logging = {
         'version': 1,
         'disable_existing_loggers': false,
         'formatters': {
             'default': {
                 'format': '[%(asctime)s] %(levelname)s: '
                           '%(message)s',
             }
         },
         'handlers': {
             'file': {
                 'class': 'logging.timedrotatingfilehandler',
                 'filename': '/var/log/$django_project/'
                             '$django_project.log',
                 'when': 'midnight',
                 'backupcount': 60,
                 'formatter': 'default',
             },
         },
         'root': {
             'handlers': ['file'],
             'level': 'info',
         },
     }
    
    caches = {
        'default': {
            'backend': 'django.core.cache.backends.filebased.'
                       'filebasedcache',
            'location': '/var/cache/$django_project/cache',
        }
    }