[rustyconover]

When an hll_union absorbs a sketch, it may copy that sketch into its gadget. The copy helpers keep the source's allocator:

• copyAs(HLL_8), used for a LIST/SET sketch going into an empty gadget
• copy_or_downsample(), used for HLL sketches; it also allocated the downsampled array with src->getAllocator()

In addition, update(hll_sketch&&) moves the input into the gadget outright.

With a stateful allocator, the union then holds memory owned by a different allocator instance than its own. That's a problem when the input's allocator is an arena or pool released before the union, and it breaks accounting that tracks memory per allocator.

This change:

• copies, or takes over a moved sketch, only when the input's allocator compares equal to the union's
• otherwise uses the paths the union already has for building into its own memory: inserting coupons for LIST/SET inputs, or merging into a new Hll8Array allocated with the union's allocator (the existing downsampling path)

std::allocator, and any allocator that always compares equal, behaves exactly as before.

Test

hll union: keeps no memory of an input sketch's allocator uses a small stateful allocator that counts bytes per instance. It covers:

• LIST, SET and HLL inputs
• lg_k below, equal to and above the union's
• HLL_4, HLL_6 and HLL_8 inputs
• empty and non-empty unions
• const& and && updates

In every case, once the input is destroyed, none of its allocator's memory may be left, and the estimate must match a std::allocator union. Without the fix it fails at REQUIRE( pool_bytes[1] == 0 ); with it, all HLL tests pass (65 test cases).

We found this in the DuckDB DataSketches extension. We're moving aggregate states into DuckDB's per-query arena so they count toward its memory limit and can spill to disk, and HLL unions kept heap memory from deserialized input sketches.

🤖 Generated with Claude Code