This is the third part in a series of posts regarding CurateND’s Library Collection development. The series will dive into the details of implementing a new type of collections within our Hydra ecosystem.

Before reading this post, please read the previous post for planning information.

Here is a link to the final repository state for this solution.

Building a Gem for Rapid Testing

As I stated earlier, fast tests were a requirement. So instead of attempting to incorporate this logic into the existing CurateND code-base, I opted to create a small gem to iterate through these solutions.

One metric to consider is that simply loading the test suite for CurateND requires 10 seconds. Whereas loading the test suite for Curate::Indexer requires about 1/10th of a second.

With the gem in place, I wired up Rubocop and SimpleCov to help ensure that I’m writing consistent code and that its tested.

Test Driving Lap #1

I knew that I was going to need different objects:

  • Preservation object (i.e. the Fedora object)
  • Index object (i.e. the SOLR document)

I also suspected, based on our initial modeling that we would need a Processing Object; Something that could track the state of what had been visited.

Here was my initial guiding scenario (its a bit chatty):

module Curate
  RSpec.describe Indexer do
    before { Indexer::Persistence.clear! }

    context 'Graph Scenario 1' do
      let!(:collection_a) { 'a') }
      let!(:collection_b) { 'b', is_member_of: [,]) }
      let!(:collection_c) { 'c', is_member_of: []) }
      let!(:collection_d) { 'd') }
      let!(:collection_e) { 'e') }
      let!(:collection_f) { 'f') }
      let!(:collection_g) { 'g') }
      let!(:work_1) { '1', is_member_of: [,]) }
      let!(:work_2) { '2', is_member_of: []) }
      let!(:work_3) { '3', is_member_of: []) }
      let!(:work_4) { '4', is_member_of: []) }
      let!(:work_5) { '5', is_member_of: []) }
      let!(:work_6) { '6') }

      context 'when building index for Work 2' do
        it 'will be direct in Collection C and transitive in B, A, D' do
          # Do the work
          response = Indexer.reindex(pid:

          expect(response.is_member_of).to eq([])
          expect(response.is_transitive_member_of).to eq([,,])
          expect(response.has_collection_members).to eq([])
          expect(response.has_transitive_collection_members).to eq([])

          indexed_collection_b = Indexer::Index::Query.find(
          expect(indexed_collection_b.is_transitive_member_of).to eq([,])
          expect(indexed_collection_b.is_member_of).to eq([,])
          expect(indexed_collection_b.has_collection_members).to eq([])
          expect(indexed_collection_b.has_transitive_collection_members).to eq([])

          indexed_collection_a = Indexer::Index::Query.find(
          expect(indexed_collection_a.is_transitive_member_of).to eq([])
          expect(indexed_collection_a.is_member_of).to eq([])
          expect(indexed_collection_a.has_collection_members).to eq([])
          expect(indexed_collection_a.has_transitive_collection_members.sort).to eq([,].sort)

          indexed_collection_d = Indexer::Index::Query.find(
          expect(indexed_collection_d.is_transitive_member_of).to eq([])
          expect(indexed_collection_d.is_member_of).to eq([])
          expect(indexed_collection_d.has_collection_members).to eq([])
          expect(indexed_collection_d.has_transitive_collection_members.sort).to eq([,].sort)

From my initial working commit to the final state of this solution, there were lots of collaborating objects that I leveraged to break apart the problem:

  • Reindexer: Coordinates the reindexing of the entire direct relationship graph
  • IndexingDocument: Responsible for representing an index document
  • RuntimeError: Namespacing for common errors
  • ReindexingReachedMaxLevelError: An exception thrown when a possible cycle is detected in the graph.
  • Queue: An assistive class in the breadth first search.
  • Index: Represents the interaction with the index

    • Index::Rebuilder: Responsible for co-ordinating the rebuild of the index
    • Index::Document: Responsible for representing an index document (extends IndexingDocument)
    • Index::Query: Contains the Query interactions with the Index
  • Processing: Responsible for coordinating all of the building process of the new index data.

    • Processing::Builder: Responsible for building a processing document by “smashing” together a persisted document and its index representation.
    • Index::Document: Represents a document under processing (extends IndexingDocument)
  • Persistence Responsible for being a layer between Fedora and the heavy lifting of the reindexing processor. It has aspects that will need to change.

    • Persistence::Document: This is a disposable intermediary between Fedora and the processing system for reindexing. I believe it is a good idea to keep separation from the persistence layer and the processing. Unlike the IndexingDocument, the Persistence Document should only have the direct relationship. (extends ProcessingDocument)
    • Persistence::Collection: (extends Persistence::Document)
    • Persistence::Work: (extends Persistence::Document)
  • CachingModule: There are several layers of caching involved, this provides some of the common behavior.

There are a lot of objects to consider; but the fundamental thing to remember is that we are:

  • Loading a Fedora object and its ancestors
  • Retrieving from SOLR each loaded Fedora object
  • Processing the subgraph
  • Then writing back to SOLR with the updated relationship information.
  • Recursive methods
  • Possible cycles in the graph
  • Writing bi-directional graph information to each index document

Along the way Rubocop and SimpleCov were providing guidance. There were points in which methods were too long or complex. I spent a bit of time refactoring those, and what emerged was more descriptive method (see “Refactoring to appease rubocop” commit).

Key Abstraction

The key abstraction in this process was the Index cache and Persistence cache. Instead of interacting with CurateND’s full Persistence layer (i.e. Fedora) and Index layer (i.e. SOLR), I wrote an in memory document store that I could run my tests against.

I had custom documents with what I thought would be the bare interface for interaction (i.e. a PID and relationships attributes).

I found this abstraction key in focusing on the harder problem of graph traversal and reindexing. It allowed for very rapid feedback cycles.

At its slowest the entire build suite – Rubocop, RSpec, and SimpleCov - takes about 1.5 seconds of elapsed time (on my machine). This meant I could refactor, extend, refine, and explore in very rapid fire.

It also meant, when I hit a brick wall with this solution, my code was not already entwined in the application. And did I hit a brick wall.

Up Next

In the next installment, I’ll talk about the Pivot.