# coding: utf-8 from django.db import models CHOICES = ( (1, 'first'), (2, 'second'), ) class Article(models.Model): headline = models.CharField(max_length=100, default='Default headline') pub_date = models.DateTimeField() status = models.IntegerField(blank=True, null=True, choices=CHOICES) misc_data = models.CharField(max_length=100, blank=True) article_text = models.TextField() class Meta: ordering = ('pub_date','headline') # A utf-8 verbose name (Ångström's Articles) to test they are valid. verbose_name = "\xc3\x85ngstr\xc3\xb6m's Articles" def __unicode__(self): return self.headline class Movie(models.Model): #5218: Test models with non-default primary keys / AutoFields movie_id = models.AutoField(primary_key=True) name = models.CharField(max_length=60) class Party(models.Model): when = models.DateField(null=True) class Event(models.Model): when = models.DateTimeField() class Department(models.Model): id = models.PositiveIntegerField(primary_key=True) name = models.CharField(max_length=200) def __unicode__(self): return self.name class Worker(models.Model): department = models.ForeignKey(Department) name = models.CharField(max_length=200) def __unicode__(self): return self.name class BrokenUnicodeMethod(models.Model): name = models.CharField(max_length=7) def __unicode__(self): # Intentionally broken (trying to insert a unicode value into a str # object). return 'Názov: %s' % self.name __test__ = {'API_TESTS': """ (NOTE: Part of the regression test here is merely parsing the model declaration. The verbose_name, in particular, did not always work.) An empty choice field should return None for the display name. >>> from datetime import datetime >>> a = Article(headline="Look at me!", pub_date=datetime.now()) >>> a.save() >>> a.get_status_display() is None True Empty strings should be returned as Unicode >>> a2 = Article.objects.get(pk=a.id) >>> a2.misc_data u'' # TextFields can hold more than 4000 characters (this was broken in Oracle). >>> a3 = Article(headline="Really, really big", pub_date=datetime.now()) >>> a3.article_text = "ABCDE" * 1000 >>> a3.save() >>> a4 = Article.objects.get(pk=a3.id) >>> len(a4.article_text) 5000 # Regression test for #659 >>> import datetime >>> p = Party.objects.create(when = datetime.datetime(1999, 12, 31)) >>> p = Party.objects.create(when = datetime.datetime(1998, 12, 31)) >>> p = Party.objects.create(when = datetime.datetime(1999, 1, 1)) >>> [p.when for p in Party.objects.filter(when__month=2)] [] >>> [p.when for p in Party.objects.filter(when__month=1)] [datetime.date(1999, 1, 1)] >>> [p.when for p in Party.objects.filter(when__month=12)] [datetime.date(1999, 12, 31), datetime.date(1998, 12, 31)] >>> [p.when for p in Party.objects.filter(when__year=1998)] [datetime.date(1998, 12, 31)] # Regression test for #8510 >>> [p.when for p in Party.objects.filter(when__day='31')] [datetime.date(1999, 12, 31), datetime.date(1998, 12, 31)] >>> [p.when for p in Party.objects.filter(when__month='12')] [datetime.date(1999, 12, 31), datetime.date(1998, 12, 31)] >>> [p.when for p in Party.objects.filter(when__year='1998')] [datetime.date(1998, 12, 31)] # Date filtering was failing with NULL date values in SQLite (regression test # for #3501, amongst other things). >>> _ = Party.objects.create() >>> p = Party.objects.filter(when__month=1)[0] >>> p.when datetime.date(1999, 1, 1) >>> l = Party.objects.filter(pk=p.pk).dates("when", "month") >>> l[0].month == 1 True # Check that get_next_by_FIELD and get_previous_by_FIELD don't crash when we # have usecs values stored on the database # # [It crashed after the Field.get_db_prep_* refactor, because on most backends # DateTimeFields supports usecs, but DateTimeField.to_python didn't recognize # them. (Note that Model._get_next_or_previous_by_FIELD coerces values to # strings)] # >>> e = Event.objects.create(when = datetime.datetime(2000, 1, 1, 16, 0, 0)) >>> e = Event.objects.create(when = datetime.datetime(2000, 1, 1, 6, 1, 1)) >>> e = Event.objects.create(when = datetime.datetime(2000, 1, 1, 13, 1, 1)) >>> e = Event.objects.create(when = datetime.datetime(2000, 1, 1, 12, 0, 20, 24)) >>> e.get_next_by_when().when datetime.datetime(2000, 1, 1, 13, 1, 1) >>> e.get_previous_by_when().when datetime.datetime(2000, 1, 1, 6, 1, 1) # Check Department and Worker >>> d = Department(id=10, name='IT') >>> d.save() >>> w = Worker(department=d, name='Full-time') >>> w.save() >>> w # Models with broken unicode methods should still have a printable repr >>> b = BrokenUnicodeMethod(name="Jerry") >>> b.save() >>> BrokenUnicodeMethod.objects.all() [] """ }