Reference processing

GitLab Flavored Markdown includes the ability to process references to a range of GitLab domain objects. This is implemented by two abstractions in the Banzai pipeline: ReferenceFilter and ReferenceParser. This page explains what these are, how they are used, and how you would implement a new filter/parser pair.

Note: Each ReferenceFilter must have a corresponding ReferenceParser.

It is possible to share reference parsers between filters - if two filters find and link the same type of objects (as specified by the data-reference-type attribute), then we only need one reference parser for that type of domain object.

Reference filters

The first way that references are handled is by reference filters. These are the tools that identify short-code and URI references from markup documents and transform them into structured links to the resources they represent.

For example, the class Banzai::Filter::IssueReferenceFilter is responsible for handling references to issues, such as gitlab-org/gitlab#123 and

All reference filters are instances of HTML::Pipeline::Filter, and inherit (often indirectly) from Banzai::Filter::ReferenceFilter.

HTML::Pipeline::Filter has a simple interface consisting of #call, a void method that mutates the current document. ReferenceFilter provides methods that make defining suitable #call methods easier. Most reference filters however do not inherit from either of these classes directly, but from AbstractReferenceFilter, which provides a higher-level interface.

Subclasses of AbstractReferenceFilter generally do not override #call; instead, a minimum implementation of AbstractReferenceFilter should define:

  • .reference_type: The type of domain object.

    This is usually a keyword, and is used to set the data-reference-type attribute on the generated link, and is an important part of the interaction with the corresponding ReferenceParser (see below).

  • .object_class: a reference to the class of the objects a filter refers to.

    This is used to:

    • Find the regular expressions used to find references. The class should include Referable and thus define two regular expressions: .link_reference_pattern and .reference_pattern, both of which should contain a named capture group named the value of ReferenceFilter.object_sym.
    • Compute the .object_name.
    • Compute the .object_sym (the group name in the reference patterns).
  • .parse_symbol(string): parse the text value to an object identifier (#to_i by default).
  • #record_identifier(record): the inverse of .parse_symbol, that is, transform a domain object to an identifier (#id by default).
  • #url_for_object(object, parent_object): generate the URL for a domain object.
  • #find_object(parent_object, id): given the parent (usually a Project) and an identifier, find the object. For example, this in a reference filter for merge requests, this might be project.merge_requests.where(iid: iid).


This default implementation is not very efficient, because we need to call #find_object for each reference, which may require issuing a DB query every time. For this reason, most reference filter implementations will instead use an optimization included in AbstractReferenceFilter:

AbstractReferenceFilter provides a lazily initialized value #records_per_parent, which is a mapping from parent object to a collection of domain objects.

To use this mechanism, the reference filter must implement the method: #parent_records(parent, set_of_identifiers), which must return an enumerable of domain objects.

This allows such classes to define #find_object (as IssuableReferenceFilter does) as:

def find_object(parent, iid)

This makes the number of queries linear in the number of projects. We only need to implement parent_records method when we call records_per_parent in our reference filter.

Reference parsers

In a number of cases, as a performance optimization, we render Markdown to HTML once, cache the result and then present it to users from the cached value. For example this happens for notes, issue descriptions, and merge request descriptions. A consequence of this is that a rendered document might refer to a resource that some subsequent readers should not be able to see.

For example, you might create an issue, and refer to a confidential issue #1234, which you have access to. This is rendered in the cached HTML as a link to that confidential issue, with data attributes containing its ID, the ID of the project and other confidential data. A later reader, who has access to your issue might not have permission to read issue #1234, and so we need to redact these sensitive pieces of data. This is what ReferenceParser classes do.

A reference parser is linked to the object that it handles by the link advertising this relationship in the data-reference-type attribute (set by the reference filter). This is used by the ReferenceRedactor to compute which nodes should be visible to users:

def nodes_visible_to_user(nodes)
  per_type = { |h, k| h[k] = [] }
  visible =

  nodes.each do |node|
    per_type[node.attr('data-reference-type')] << node

  per_type.each do |type, nodes|
    parser = Banzai::ReferenceParser[type].new(context)

    visible.merge(parser.nodes_visible_to_user(user, nodes))


The key part here is Banzai::ReferenceParser[type], which is used to look up the correct reference parser for each type of domain object. This requires that each reference parser must:

  • Be placed in the Banzai::ReferenceParser namespace.
  • Implement the .nodes_visible_to_user(user, nodes) method.

In practice, all reference parsers inherit from BaseParser, and are implemented by defining:

  • .reference_type, which should equal ReferenceFilter.reference_type.
  • And by implementing one or more of:
    • #nodes_visible_to_user(user, nodes) for finest grain control.
    • #can_read_reference? needed if nodes_visible_to_user is not overridden.
    • #references_relation an active record relation for objects by ID.
    • #nodes_user_can_reference(user, nodes) to filter nodes directly.
Note: A failure to implement this class for each reference type means that the application will raise exceptions during Markdown processing.