Feature Request: CopyNonGround() Optimization for Term Types

[alex60217101990]

Following maintainer feedback on previous PRs to focus on optimizations that demonstrate measurable production impact rather than microbenchmark improvements, I profiled OPA in our production environment. Term copying emerged as one of the dominant memory consumers, particularly for refs, arrays, sets, and objects during policy evaluation.

Observation

Current Copy() methods perform deep copies of all elements, including ground (constant) terms that are immutable and never modified after creation. In production workloads, the majority of terms are ground—copying them unnecessarily wastes CPU and allocates memory that's immediately eligible for GC.

Analysis of topdown/eval.go and ast/compile.go shows numerous call sites where Copy() is invoked on collections with predominantly ground elements, but only non-ground parts actually need deep copying.

Proposal

Implement CopyNonGround() methods for Array, Set, and Object types (matching the existing Ref.CopyNonGround() pattern from PR #7350)

Apply at call sites in:

• topdown/eval.go: 3 locations (getRules(), vcKeyScope.reduce(), namespaceRef())
• ast/compile.go: 2 locations (resolveRef(), resolveRefsInTerm())
• topdown/providers.go: 1 location

Would the OPA maintainers be open to a PR with this optimization? The code changes are minimal, well-tested, and demonstrate measurable production improvements as requested in previous feedback.

This is a focused micro-optimization that reduces unnecessary allocations without architectural changes or API breaks.

[anderseknert]

Indeed, that sounds like a worthwhile optimization to me 👍

[alex60217101990]

@anderseknert Thank you for your feedback. I've prepared the code for this PR in my branch. Could you please take a look at open-policy-agent/opa#8233 when you have a chance? From what I understood from the last comment (open-policy-agent/opa#8233 (comment)), I cannot open a new PR with changes until the previous one is resolved. 🙂