Most Google Cloud Datastore queries return whole entities as their results, but often an application is actually interested in only a few of the entity's properties. Projection queries allow you to query Cloud Datastore for just those specific properties of an entity that you actually need, at lower latency and cost than retrieving the entire entity.
Projection queries are similar to SQL queries of the form:
SELECT name, email, phone FROM CUSTOMER
You can use all of the filtering and sorting features available for standard entity queries, subject to the limitations described below. The query returns abridged results with only the specified properties (
phone in the example) populated with values; all other properties have no data.
Using projection queries in PythonConsider the following model:
class Article(ndb.Model): title = ndb.StringProperty() author = ndb.StringProperty() tags = ndb.StringProperty(repeated=True)
You specify a projection this way:
def print_author_tags(): query = Article.query() articles = query.fetch(20, projection=[Article.author, Article.tags]) for article in articles: print(article.author) print(article.tags) # article.title will raise a ndb.UnprojectedPropertyError
You handle the results of these queries just as you would for a standard entity query: for example, by iterating over the results.
for article in articles: print(article.author) print(article.tags) # article.title will raise a ndb.UnprojectedPropertyError
You can project indexed sub-properties from a structured property. To get only the
city property of a contact's
address structured property, you could use a projection like:
class Address(ndb.Model): type = ndb.StringProperty() # E.g., 'home', 'work' street = ndb.StringProperty() city = ndb.StringProperty()
class Contact(ndb.Model): name = ndb.StringProperty() addresses = ndb.StructuredProperty(Address, repeated=True)
Contact.query().fetch(projection=["name", "addresses.city"]) Contact.query().fetch(projection=[Contact.name, Contact.addresses.city])
Projection queries can use the
distinct keyword to ensure that only completely unique results will be returned in a result set. This will only return the first result for entities which have the same values for the properties that are being projected.
Article.query(projection=[Article.author], group_by=[Article.author]) Article.query(projection=[Article.author], distinct=True)
Both queries are equivalent and will produce each author's name only once.
Limitations on projections
Projection queries are subject to the following limitations:
Only indexed properties can be projected.
The same property cannot be projected more than once.
Properties referenced in an equality (
=) or membership (
IN) filter cannot be projected.
SELECT A FROM kind WHERE B = 1
is valid (projected property not used in the equality filter), as is
SELECT A FROM kind WHERE A > 1
(not an equality filter), but
SELECT A FROM kind WHERE A = 1
(projected property used in equality filter) is not.
Results returned by a projection query cannot be saved back to Cloud Datastore.
Because the query returns results that are only partially populated, you cannot write them back to Cloud Datastore.
Projections and multiple-valued properties
Projecting a property with multiple values will not populate all values for that property. Instead, a separate entity will be returned for each unique combination of projected values matching the query. For example, suppose you have an entity of kind
Foo with two multiple-valued properties,
entity = Foo(A=[1, 1, 2, 3], B=['x', 'y', 'x'])
Then the projection query
SELECT A, B FROM Foo WHERE A < 3
will return four entities with the following combinations of values:
Indexes for projections
Projection queries require all properties specified in the projection to be included in a Cloud Datastore index. The App Engine development server automatically generates the needed indexes for you in the index configuration file,
index.yaml, which is uploaded with your application.
One way to minimize the number of indexes required is to project the same properties consistently, even when not all of them are always needed. For example, these queries require two separate indexes:
SELECT A, B FROM Kind SELECT A, B, C FROM Kind
However, if you always project properties
C, even when
C is not required, only one index will be needed.
Converting an existing query into a projection query may require building a new index if the properties in the projection are not already included in another part of the query. For example, suppose you had an existing query like
SELECT * FROM Kind WHERE A > 1 ORDER BY A, B
which requires the index
Index(Kind, A, B)
Converting this to either of the projection queries
SELECT C FROM Kind WHERE A > 1 ORDER BY A, B SELECT A, B, C FROM Kind WHERE A > 1 ORDER BY A, B
introduces a new property (
C) and thus will require building a new index
C). Note that the projection query
SELECT A, B FROM Kind WHERE A > 1 ORDER BY A, B
would not change the required index, since the projected properties
B were already included in the existing query.