Question: Edge types losing node fields after SDL round-trip - expected behavior or bug?

[marcdaniels-toast]

Context

I'm working on #1029 in steps and I'm enhancing the widgets.rb schema with an example multi-level interface hierarchy to show index inheritance. While testing this kind of configuration, I came across an issue related to how we round-trip the SDL through ::GraphQL::Schema.from_definition().to_definition() in SchemaArtifactManager. For my multi-level interface hierarchy, some interface Edge types are losing their node fields during this from-to "round-trip". I'm still new to this codebase and GraphQL in general and trying to understand if this can be explained as expected behavior or if it's a bug in the GraphQL gem.

What I'm Seeing

I've isolated this to seemingly relate to the alphabetical ordering of the type names in the hierarchy. In interface hierarchies of 4+ levels, when a child type's name sorts alphabetically before its parent interface name, the grandparent interface's Edge loses its node field after the round-trip. This seems to cascade up the hierarchy.

Example:

• Hierarchy: Alpha (top-most interface) → Bravo → Echo → Delta → Charlie (base concrete type)
• Charlie < Delta alphabetically
• After round-trip: EchoEdge, BravoEdge, and AlphaEdge all lose their node fields
• But: CharlieEdge and DeltaEdge keep their node fields

Why I'm Confused

1. Before round-trip: All Edge types have their node fields
2. After round-trip: Only certain Edges lose them based on alphabetical ordering
3. Transitive interfaces don't help: Even when we explicitly list all transitive interfaces (e.g., Charlie implements Delta & Echo & Bravo & Alpha), the bug still occurs
4. Only affects deep hierarchies: Works fine with 3 levels, breaks at 4+

Test Cases

I've written three minimal test cases using NATO phonetic alphabet names to make the alphabetical relationships clear:

1. ✅ Works: Alpha → Bravo → Charlie → Delta → Echo (all children sort after parents) → all Edges keep nodes
2. ❌ Doesn't work: Alpha → Bravo → Echo → Delta → Charlie (Charlie < Delta) → ancestor Edges lose nodes
3. ❌ Doesn't work: Alpha → Bravo → Hotel → Echo → Foxtrot (Echo < Hotel) → ancestor Edges lose nodes

I'll post the diff that adds these three tests so you can run these locally and see the issue or adjust the hierarchy/debug messages, etc.

I'm looking for advice on how to proceed. I could ignore this issue and just set up an example hierarchy less than 4 levels deep, but it seems more like a bug in ::GraphQL::Schema?

[marcdaniels-toast]

Here's the diff that adds the three demonstration tests mentioned above:

diff --git a/elasticgraph-schema_definition/spec/unit/elastic_graph/schema_definition/graphql_schema/interface_type_spec.rb b/elasticgraph-schema_definition/spec/unit/elastic_graph/schema_definition/graphql_schema/interface_type_spec.rb
index dea8b78d..e57e9af3 100644
--- a/elasticgraph-schema_definition/spec/unit/elastic_graph/schema_definition/graphql_schema/interface_type_spec.rb
+++ b/elasticgraph-schema_definition/spec/unit/elastic_graph/schema_definition/graphql_schema/interface_type_spec.rb
@@ -111,6 +111,115 @@ module ElasticGraph
           include_examples "#implements",
             graphql_definition_keyword: "interface",
             ruby_definition_method: :interface_type
+
+          it "preserves all edge node fields when all children sort after parents (5-level hierarchy)" do
+            # 5-level hierarchy: Alpha < Bravo < Charlie < Delta < Echo
+            # All children > parent
+            intermediate_sdl = define_schema do |api|
+              api.interface_type "Alpha" do |t|
+                t.field "id", "ID!"
+              end
+
+              api.interface_type "Bravo" do |t|
+                t.implements "Alpha"
+                t.field "id", "ID!"
+              end
+
+              api.interface_type "Charlie" do |t|
+                t.implements "Bravo"
+                t.field "id", "ID!"
+              end
+
+              api.interface_type "Delta" do |t|
+                t.implements "Charlie"
+                t.field "id", "ID!"
+              end
+
+              api.object_type "Echo" do |t|
+                t.implements "Delta"
+                t.field "id", "ID!"
+                t.index "echos"
+              end
+            end
+
+            puts "\n=== Edge definitions BEFORE round-trip ==="
+            puts edge_type_from(intermediate_sdl, "Echo")
+            puts edge_type_from(intermediate_sdl, "Delta")
+            puts edge_type_from(intermediate_sdl, "Charlie")
+            puts edge_type_from(intermediate_sdl, "Bravo")
+            puts edge_type_from(intermediate_sdl, "Alpha")
+
+            round_tripped_sdl = ::GraphQL::Schema.from_definition(intermediate_sdl).to_definition
+
+            puts "\n=== Edge definitions AFTER round-trip ==="
+            puts edge_type_from(round_tripped_sdl, "Echo")
+            puts edge_type_from(round_tripped_sdl, "Delta")
+            puts edge_type_from(round_tripped_sdl, "Charlie")
+            puts edge_type_from(round_tripped_sdl, "Bravo")
+            puts edge_type_from(round_tripped_sdl, "Alpha")
+
+            # All edges should retain node fields
+            expect(edge_type_from(round_tripped_sdl, "Echo")).to include("node: Echo")
+            expect(edge_type_from(round_tripped_sdl, "Delta")).to include("node: Delta")
+            expect(edge_type_from(round_tripped_sdl, "Charlie")).to include("node: Charlie")
+            expect(edge_type_from(round_tripped_sdl, "Bravo")).to include("node: Bravo")
+            expect(edge_type_from(round_tripped_sdl, "Alpha")).to include("node: Alpha")
+          end
+
+          it "demonstrates bug: concrete type < parent causes grandparent to lose node (5-level hierarchy)" do
+            # Hierarchy: Alpha → Bravo → Echo → Delta → Charlie
+            # Charlie < Delta → Echo (grandparent) loses node
+            # node loss cascades upward
+            intermediate_sdl = define_schema do |api|
+              api.interface_type "Alpha" do |t|
+                t.field "id", "ID!"
+              end
+
+              api.interface_type "Bravo" do |t|
+                t.implements "Alpha"
+                t.field "id", "ID!"
+              end
+
+              api.interface_type "Echo" do |t|
+                t.implements "Bravo"
+                t.field "id", "ID!"
+              end
+
+              api.interface_type "Delta" do |t|
+                t.implements "Echo"
+                t.field "id", "ID!"
+              end
+
+              api.object_type "Charlie" do |t|
+                t.implements "Delta"
+                t.field "id", "ID!"
+                t.index "charlies"
+              end
+            end
+
+            puts "\n=== Edge definitions BEFORE round-trip ==="
+            puts edge_type_from(intermediate_sdl, "Charlie")
+            puts edge_type_from(intermediate_sdl, "Delta")
+            puts edge_type_from(intermediate_sdl, "Echo")
+            puts edge_type_from(intermediate_sdl, "Bravo")
+            puts edge_type_from(intermediate_sdl, "Alpha")
+
+            round_tripped_sdl = ::GraphQL::Schema.from_definition(intermediate_sdl).to_definition
+
+            puts "\n=== Edge definitions AFTER round-trip ==="
+            puts edge_type_from(round_tripped_sdl, "Charlie")
+            puts edge_type_from(round_tripped_sdl, "Delta")
+            puts edge_type_from(round_tripped_sdl, "Echo")
+            puts edge_type_from(round_tripped_sdl, "Bravo")
+            puts edge_type_from(round_tripped_sdl, "Alpha")
+
+            # All edges should retain node fields
+            expect(edge_type_from(round_tripped_sdl, "Charlie")).to include("node: Charlie")
+            expect(edge_type_from(round_tripped_sdl, "Delta")).to include("node: Delta")
+            expect(edge_type_from(round_tripped_sdl, "Echo")).to include("node: Echo")  # This will fail (bug)
+            expect(edge_type_from(round_tripped_sdl, "Bravo")).to include("node: Bravo") # This will fail (bug)
+            expect(edge_type_from(round_tripped_sdl, "Alpha")).to include("node: Alpha") # This will fail (bug)
+          end
         end
       end
     end

[myronmarston]

Thanks for reporting this! It's a graphql-ruby bug. I've reported it there:

rmosolgo/graphql-ruby#5580

In the meantime, there's a workaround--we can list all the interface ancestors in the implements clause, as per the GraphQL spec.

Want to take a stab at implementing that? While I'm calling it a workaround, we're arguably not in compliance with the GraphQL spec and we should be so it's a good thing to improve, regardless.

[marcdaniels-toast]

Yes sounds good. I'll do a small PR just for the ImplementsInterfaces enhancement/fix to include all ancestor interfaces. If I remember correctly that didn't completely work around all cases here, but I'll take another look with tests.