🐛 RelationConnection loads the same relation twice under AsyncDataloader (2.6.7)

[benschwarz]

Given a query like:

{ posts(first: 20) { edges { node { id } } pageInfo { endCursor } } }

Both fields call load_nodes. 2.6.7 runs them concurrently, earlier releases didn't.

@nodes ||= limited_nodes.to_a   # relation_connection.rb:224

Fiber one waits on the DB before setting @nodes. Fiber two sees empty and loads the same relation again. Whichever finishes first clears ActiveRecord's load state from under the other.

This results in:

NoMethodError: undefined method 'instantiate' for nil, plus a duplicate query on every paginated connection.

[rmosolgo]

Hey, thanks for reporting this and sorry for the trouble 😖

I'm hesitant to address this with dataloader.yield because I expect this to all happen inside GraphQL-Ruby's execution steps -- without handing off to Dataloader to fetch results from Dataloader sources.

Do you have a guess why the Fiber is yielding here? Is this a custom connection implementation?

[benschwarz]

Hey @rmosolgo, fair enough. I am not sure if this is the right fix either, but I was able to get my test suite running with it -- good to open the discussion though.

I'm on 1.6.3, and there are seemingly many issues with async dataloader where db connections get used by old/finished/crashed fibres. On 1.6.3 we have a custom ConnectionReleasingDataloader < GraphQL::Dataloader::AsyncDataloader that calls release_connection and also reclaims abandoned fibres. It's non ideal but is working well.

Given the updates in 1.6.7 I thought it was worthwhile to see if I could get any joy from it. In using the code in this PR, I was able to upgrade, and also remove the connection releasing dataloader 😇

[rmosolgo]

Thanks for sharing a bit more. I'd love to hear more details (eg error messages, backtraces) about what goes wrong without manual connection handling. Maybe there's a clue in there to an underlying bug!

Also, if you find any clues about why load_nodes would cause a fiber yield, please do share -- I'd love to understand this better and find a solid fix.