[tuxedomik]

Summary

Fixes an issue where exporting Java sampler data could mutate the live profiling tree.

JavaNodeExporter merges stack trace nodes representing the same method before serializing them. Previously, this merge was performed directly on nodes from the sampler's live data:

other.merge(child);

Because toProto() can be called multiple times during a profiling session, such as by the live viewer and again when the profiler stops, repeated exports could progressively inflate descendant timings.

This could result in invalid profiles where child timings exceed their parents, percentages exceed 100%, and calculated self time becomes negative.

Fix

Use copy-on-write when merging nodes during export.

Nodes are reused normally. When a merge is required, the target node is first detached from the live sampler tree by copying it. The merge is then performed on that copy.

This avoids copying the entire profiling tree while ensuring serialization cannot modify the underlying sampling data.

Tests

Added a regression test which verifies that:

• nodes representing the same method are still merged correctly in the exported profile
• exporting the same sampler tree multiple times produces identical output
• exporting does not modify the original sampler tree

./gradlew :spark-common:test passes.

Note

this was fixed by AI

[lucko]

Thanks!