Python 3.10 is now available in preview.

google.appengine.datastore.datastore_index.CompositeIndexForQuery

Returns the composite index needed for a query.

A query is translated into a tuple, as follows:

  • The first item is the kind string, or None if we're not filtering on kind (see below).

  • The second item is a bool giving whether the query specifies an ancestor.

  • After that come (property, ASCENDING) pairs for those Filter entries whose operator is EQUAL or IN. Since the order of these doesn't matter, they are sorted by property name to normalize them in order to avoid duplicates.

  • After that comes at most one (property, ASCENDING) pair for a Filter entry whose operator is on of the four inequalities. There can be at most one of these.

  • After that come all the (property, direction) pairs for the Order entries, in the order given in the query. Exceptions:

    • (a) If there is a Filter entry with an inequality operator that matches the first Order entry, the first order pair is omitted (or, equivalently, in this case the inequality pair is omitted).

    • (b) If an Order entry corresponds to an equality filter, it is ignored (since there will only ever be one value returned).

    • (c) If there is an equality filter on __key__ all orders are dropped (since there will be at most one result returned).

    • (d) If there is an order on __key__ all further orders are dropped (since keys are unique).

    • (e) Orders on __key__ ASCENDING are dropped (since this is supported natively by the datastore).

  • Finally, if there are Filter entries whose operator is EXISTS, and whose property names are not already listed, they are added, with the direction set to ASCENDING.

This algorithm should consume all Filter and Order entries.

Additional notes:

  • The low-level implementation allows queries that don't specify a kind; but the Python API doesn't support this yet.

  • If there's an inequality filter and one or more sort orders, the first sort order must match the inequality filter.

  • The following indexes are always built in and should be suppressed:

    • Query on kind only;
    • Query on kind and one filter or one order;
    • Query on ancestor only, without kind (not exposed in Python yet);
    • Query on kind and equality filters only, no order (with or without ancestor).
  • While the protocol buffer allows a Filter to contain multiple properties, we don't use this. It is only needed for the IN operator but this is (currently) handled on the client side, so in practice each Filter is expected to have exactly one property.

query A datastore_pb.Query instance.

A tuple of the form (required, kind, ancestor, properties). required: Boolean, whether the index is required; kind: The kind or None; ancestor: True if this is an ancestor query; properties: A tuple consisting of:

  • The prefix, represented by a set of property names.
  • The postfix, represented by a tuple consisting of any number of:
    • Sets of property names or PropertySpec objects: these properties can appear in any order.
    • Sequences of PropertySpec objects: Indicates the properties must appear in the given order, with the specified direction (if specified in the PropertySpec).