A bit of context

Before PostgreSQL 9.4 (released December 2014) introduced jsonb, the only way to store JSON data was as a string: both json and text types from PostgreSQL store strings.

Also, django.contrib.postgres.fields.JSONField did not exist before Django 1.9 (released December 2015). You had to use third-party libraries to handle JSON in your models (or make your own field).

We chose django-jsonfield.

But now it's 2018, we have PostgreSQL 10, our systems are running with Python 3.6 and Django 1.11… and django-jsonfield isn't compatible with any newer version of Django and Python, and so we couldn't upgrade to Django 2.0.

It was time to tackle the issue!

Our constraints

Our systems run 24/7. No disruptions in service when we can prevent it.

This means that we cannot change the code and change the column format all at once. If we did, we would have some server with the old code running at the same time as servers with the new code. This rarely works with schema migrations.

Our basic migration scheme in similar cases is:

• step 1:
  • create a new_field,
  • edit code to write in both fields but read in old_field,
  • migrate all existing data after deploying the new code (in an empty Django migration);
• step 2:
  • read in new_field,
  • drop old_field.

But this takes 2 releases and you sometimes have to edit many lines of code to swap the uses of old_field to new_field everywhere. And then many of your field names are called new_* which is a bit frustrating when you are as perfectionist as we are!

That's a lot of work, and after that, we will still have import jsonfield in our migrations, so we will need to squash migrations to clean that, which is tedious and takes a long time. That would delay our migration to Django 2.

Since we did not want to wait too long, we tried another approach.

How we did it

The nice thing is that writing JSON in a PostgreSQL database uses the same raw SQL whether the column type is jsonb, json or text. This means that we don't need the old_field/new_field dance of our basic migration scheme.

The only thing we need is to have a Django model field that handle both jsonb and text formatted columns when reading.

So our migration scheme becomes:

• deploy code with a MigrationJSONField handling both json and jsonb columns;
• change column type in the database;
• clean the code for the next release to use the default JSONField that does not handle text columns.

class MigrationJSONDecoder(json.JSONDecoder):
    """Handle JSON stored as text or jsonb in PostgreSQL."""

    def decode(self, obj, *args, **kwargs):
        obj = super().decode(obj, *args, **kwargs)
        return self.post_decode(obj)

    @classmethod
    def post_decode(cls, obj):
        if isinstance(obj, str):
            obj = json.loads(obj)
        return obj


class MigrationJSONField(pg_fields.JSONField):
    """Handle migration from jsonfield.JSONField to pg_fields.JSONField."""

    def __init__(self, **kwargs):
        self.decoder = MigrationJSONDecoder
        super().__init__(**kwargs)

    def from_db_value(self, value, _expression, _connection, _context):
        return self.decoder.post_decode(value)

def __init__(self, **kwargs):
    self.decoder = MigrationJSONDecoder
    kwargs['encoder'] = myJSONEncoder
    super().__init__(**kwargs)

ALTER TABLE my_table ALTER COLUMN my_column TYPE JSONB USING my_column::JSONB

from django.db import migrations

sql = "ALTER TABLE t ALTER COLUMN col TYPE JSONB USING col::JSONB"
reverse_sql = "ALTER TABLE t ALTER COLUMN col TYPE TEXT USING col::TEXT"

class Migration(migrations.Migration):
    dependencies = [
        ('app_name', 'previous_migration_name'),
    ]
    operations = [
        migrations.RunSQL(
            sql=sql,
            reverse_sql=reverse_sql,
        ),
    ]

The tricky parts

Conclusion

All in all, this migration went relatively well, considering the number of impacted columns accross all our systems (46). That was our last showstopper before migrating to Django 2, which will be our next major upgrade.