4 Search backend implementation notes
Matt Westcott edytuje tę stronę 2020-05-15 00:45:58 +01:00

Search backend implementation notes

The SearchBackend class

Each search backend module defines a SearchBackend class inheriting from wagtail.search.backends.base.BaseSearchBackend which serves as the entry point for all code interfacing with wagtail.search. A search backend is obtained through the get_search_backend function:

wagtail.search.backends.get_search_backend(backend_name='default', **kwargs)

This instantiates the named backend as listed in WAGTAILSEARCH_BACKENDS, passing the (backend-specific) configuration options from there.

The reset_index, add_type, refresh_index, add, add_bulk and delete methods on SearchBackend are obsolete, and replaced by Rebuilder objects.

Indexing

Indexing (as performed by the update_index management command) is handled by the rebuilder class given by backend.rebuilder_class. If this is None, this backend does not require indexing (e.g. the database backend).

An Index object represents a storage location where we store the searchable data for some subset of the project's models. There can be multiple indexes, and the details of how models are allocated to indexes (and, indeed, what the concept of an index means for a particular backend implementation) is internal to that backend. For Postgres, there is a single index for all models; for Elasticsearch there is one index per base model class (Page, Document, Image etc). The index for a given model can be retrieved with backend.get_index_for_model(model).

To insert data into an index, update_index makes the following sequence of method calls:

# Perform initialisation so that the index is ready to receive data
rebuilder = backend.rebuilder_class(index)
index = rebuilder.start()

# Configure the index to accept data for a given model
index.add_model(model)

# Insert data for a queryset (or other iterable) of model instances; update_index takes care of
# doing this in appropriately sized batches
index.add_items(model, items)

# Perform any necessary cleanup / finalisation after all data has been inserted
rebuilder.finish()

Indexing - Elasticsearch-specific notes

The main logic of index.add_model / index.add_items is delegated to an internal ElasticsearchMapping class which wraps an individual model's search_fields configuration and serves as an adapter between that model and Elasticsearch:

  • add_model(model) calls the adapter's get_mapping() method, which returns an Elasticsearch-readable JSON-like configuration
  • add_item(item) calls the adapter's get_document(item) method, which returns a JSON-like representation of the indexable data.

The ElasticsearchMapping object will also be used when constructing search queries.

Indexing - PostgreSQL-specific notes

Postgres search only allows text to be indexed under one of four boost levels labelled A, B, C and D (which are assigned numeric values at query time), so extra logic is required at index time to inspect all boost values defined across all models' search_fields definitions and select a suitable set of four.

Support for field length normalisation (to give matches in shorter fields higher weighting than matches in longer ones) is also limited - we are only able to normalise against the length of the full indexed content, not an individual field. PR #6013 works around this by special-casing fields named 'title' and placing them in their own index field.

Searching

The standard entry point for performing search queries is the .search(query) and .autocomplete(query) methods on QuerySets that inherit wagtail.search.queryset.SearchableQuerySetMixin. Here query is one of:

  • a query string
  • a query expression object (supported as of PR #5953) - these allow complex queries to be built up as a tree structure, much like Q() expressions in the ORM. These objects do not implement any search logic of their own - they just define a backend-agnostic data structure for individual backends to interpret.
  • None (which is treated as "return all results")

These methods simply call get_search_backend and delegate to the search / autocomplete method on that backend, which (in BaseSearchBackend) delegate to a shared implementation in the _search method. This proceeds as follows:

  • Creates an instance of backend.query_compiler_class (a subclass of BaseSearchQueryCompiler, which is responsible for transforming the query and queryset into a backend-specific search query)
  • Calls check() on this compiler instance, to verify that the query complies with the search_fields configuration, e.g. we are not performing a full-text search on a field that is only registered as a FilterField. (This is aimed at enforcing consistent behaviour across backends, so that a site using the database backend in development but Elasticsearch in production doesn't end up defining invalid queries that are possible in SQL but unsupported on Elasticsearch.)
  • Constructs and returns an instance of backend.results_class (a subclass of BaseSearchResults, which provides the results in iterable format much like an ORM queryset) wrapping this query compiler instance

The SearchQueryCompiler class

What this class does is ultimately backend-specific; it just needs to provide an internal API that allows the (similarly backend-specific) SearchResults class to serve up results. However, one thing it will need to do is deconstruct the passed queryset so that any filters applied on it can be re-applied to the mechanism that's performing the search query. BaseSearchQueryCompiler provides base functionality for this, which all backends are expected to call, even if they run the final search query through an ORM call (and could thus use the queryset as-is rather than deconstructing it) - as with SearchQueryCompiler.check(), this ensures that all backends support the same common subset of ORM operations. (TODO: define what this subset is. Just filter and exclude?)

Subclasses of BaseSearchQueryCompiler need to implement the methods _process_lookup(field, lookup, value) (which returns a backend-specific value representing the operation of filtering on a single field) and _connect_filters(filters, connector, negated) (which returns a backend-specific value representing the logical AND/OR/NOT of the sub-operations passed in filters). The _get_filters_from_queryset() method will then return the backend-specific representation of the complete filter expression defined on the queryset.

The SearchResults class

SearchResults objects are constructed from a SearchQueryCompiler instance and provide the search results in iterable form. BaseSearchResults provides a base implementation that replicates the "lazy evaluation" behaviour of ORM querysets: queries are not performed until the results are explicitly accessed, slicing a not-yet-evaluated result set works by applying an offset and count to the query, and results are cached so that iterating over them multiple times does not re-run the query. Subclasses only need to provide a _do_search() and _do_count() method and optionally facet(field_name) (in which case they should also define supports_facet = True).

The result of _do_search() should be an iterable of model instances; search backends that do not natively return ORM objects will thus need to follow the search query with a standard ORM query to fetch those objects by ID, and ensure that the ordering on the search query is preserved.

Searching - Elasticsearch-specific notes

The Elasticsearch implementation of _get_filters_from_queryset returns a JSON-like Elasticsearch query expression to be inserted into the final query. The code for building the final query is split into get_inner_query() (which converts the query string/expression into a JSON-like query expression) and get_query() (which combines this with the filters from the queryset).

Elasticsearch places a limit on the number of items that can be returned in a standard query; result sets of > 100 items (or ones that do not set a limit when querying) thus need to use Elasticsearch's scroll API. ElasticsearchSearchResults._do_search implements both methods, and transparently selects one or the other as required.

Searching - PostgreSQL-specific notes

The Postgres implementation of _get_filters_from_queryset returns a Q() ORM expression.