laskoviymishka opened a new pull request, #2623:
URL: https://github.com/apache/iceberg-go/pull/2623

   `iceberg.VisitBoundPredicateRef` was exported and gated only by an 
`internal.BoundPredicateRef` parameter, but a caller in another module could 
still call it by letting generic type inference supply the zero value of the 
token, the same hole the decision log describes for `Schema.FieldsRef`. This PR 
removes the function and moves the zero-copy IN / NOT IN dispatch, including 
the precomputed extrema path through `VisitInWithExtrema`, into 
`VisitBoundPredicate`. That path is taken only for visitors with a 
`BorrowBoundSets(internal.BoundPredicateRef)` marker method, which code outside 
the module cannot declare because it cannot name the parameter type. The four 
evaluators in `table/evaluators.go` get the marker; they are unexported and not 
returned by any exported API, so they cannot be embedded from outside to 
inherit it. No exported signature mentions the token, and `releases.md` gets a 
compatibility note.
   
   The no-allocation and extrema tests now run through `VisitBoundPredicate` 
with a marked test visitor, plus a case showing that an unmarked visitor 
declaring `VisitInWithExtrema` gets `VisitIn`. A new test in `iceberg_test` 
checks that an external visitor receives detached sets for IN and NOT IN over 
string and binary literals: `Add` and mutating a `BinaryLiteral` do not reach 
the predicate, each visit gets its own set, and the extrema method is never 
called. I checked that the trusted tests fail when the marker is ignored and 
the external test fails when every visitor is treated as trusted. `go build 
./...`, `go vet ./...` and the tests of the root, `internal` and `table` 
packages pass. The change also reflows one composite literal in 
`catalog/dynamodb/dynamodb_integration_test.go`, which already failed 
`golangci-lint fmt --diff` on main.
   
   Benchmarks with `-count=10 -benchmem`, main against this branch, for 
`BenchmarkManifestEvaluatorInPredicate`, 
`BenchmarkInclusiveMetricsEvalInPredicate`, 
`BenchmarkManifestEvaluatorBuiltInPartitions` and 
`BenchmarkInclusiveMetricsUnfiltered`: B/op and allocs/op are identical for 
every case. The time geomeans, before and after, are 134.3ns and 133.9ns for 
the manifest IN benchmark (-0.27%), 6.870ms and 7.092ms for the inclusive 
metrics IN benchmark (+3.22%), 21.06µs and 20.33µs for the built-in partitions 
benchmark (-3.46%), and 18.74ns and 19.06ns for the unfiltered inclusive 
metrics benchmark (+1.71%), or 42.42µs and 42.90µs (+1.14%) across all of them. 
Individual cases move by up to about 15% in both directions with wide 
confidence intervals on both sides, so I read this as noise rather than a cost 
of the extra type assertion. The marker method is exported in name only, since 
the types are unexported; reaching it through `reflect` grants nothing because 
it is a no-op.
   
   Part of #2062.
   
   AI-assisted, reviewed by a human.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to