Uploaded image for project: 'Beam'
  1. Beam
  2. BEAM-2699

AppliedPTransform is used as a key in hashmaps but PTransform is not hashable/equality-comparable



    • Type: Bug
    • Status: Open
    • Priority: P3
    • Resolution: Unresolved
    • Affects Version/s: None
    • Fix Version/s: None
    • Component/s: runner-core
    • Labels:


      There's plenty of occurrences in runners-core of Map or BiMap where the key is an AppliedPTransform.

      However, PTransform does not advertise that it is required to implement equals/hashCode, and some transforms can't do it properly anyway - for example, transforms that capture a ValueProvider which is also not hashable/eq-comparable. I'm surprised that things aren't already very broken because of this.

      Fundamentally, I don't see why we should ever compare two PTransform's for equality.

      I looked at the code and wondered "can AppliedPTransform simply be identity-hashable", but right now the answer is no because we can create an AppliedPTransform for the same transform applied to the same thing multiple times.

      Fixing that appears to be not very easy, but definitely possible. Ideally TransformHierarchy.Node would just know its AppliedPTransform, however a Node can be constructed when there's yet no Pipeline. Suppose there's gotta be some way to propagate a Pipeline into Node.finishSpecifying() (which should be called exactly once on the Node, and this should be enforced), and have finishSpecifying() return the AppliedPTransform, and have the caller use that instead of potentially repeatedly calling .toAppliedPTransform() on the same Node.

      Kenneth Knowles is on vacation but perhaps Thomas Groh can help with this meanwhile?

      CC: Reuven Lax


          Issue Links



              • Assignee:
                jkff Eugene Kirpichov
              • Votes:
                0 Vote for this issue
                5 Start watching this issue


                • Created: