[sjperkins]

Summary

ReadResult.value and KvStore.__getitem__ convert the absl::Cord through a temporary std::string. That copies the payload twice. This PR reuses BytesFromCord from serialization.cc, which copies the cord's chunks straight into a new bytes object.

 py::bytes CordToPython(const absl::Cord& value) {
-  return py::bytes(std::string(value));
+  auto obj = BytesFromCord(value);
+  if (!obj) throw py::error_already_set();
+  return py::reinterpret_steal<py::bytes>(obj.release());
 }

py::bytes has no constructor that takes over a std::string. Even if it had one, a PyBytesObject keeps its data inline, so the second copy can't be avoided by moving.

Evidence

Both builds use -c opt --copt=-O3, CPython 3.14 (with the GIL) on 12-core x86-64 Linux. A local HTTP server serves 15.5 MB objects.

	master	this PR
ReadResult.value, median per 15.5 MB object	10.2–11.9 ms	1.20–1.28 ms
page faults over 200 .value calls	1.59 M	78 K
http kvstore reads at 32 in flight, calling .value on each	1.03–1.27 GB/s	4.05–4.14 GB/s
the same reads, without .value	6.98 GB/s	6.87 GB/s

The speed-up is bigger than removing one copy alone would give. The temporary string and the bytes object were each mapped fresh on every call, costing about 7,900 page faults per 15.5 MB, and this PR avoids most of those.

python/tensorstore/tests/kvstore_test.py: 16 passed.

Merge Danger

Door: two-way

Blast Radius: KvStore.__getitem__ and ReadResult.value. Both still return the same bytes.

🤖 Generated with Claude Code

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[sjperkins]

Thanks for including this change. Could you please provide a pypi release with the above fix?