Key encoding unit test flakes on wall-clock performance bound #6745

Expected Behavior

The parallel unit suite should not fail because a shared CI runner briefly provides less CPU time to one pytest worker. Key encoding performance should be tracked by the benchmark suite.

Current Behavior

test_performance_bounds_single_entity measures 1,000 serialization and deserialization calls with time.perf_counter() and requires each loop to finish within 0.2 seconds. The unit suite runs with 8 pytest-xdist workers, so elapsed time includes time that the worker is descheduled.

The test failed on master in run 31768603920:

AssertionError: Serialization too slow: 0.2062s
assert 0.20623705900004552 < 0.2

This limit has already been raised twice to address CI failures. It started at 0.02 seconds in #5981, changed to 0.1 seconds in #6040, then changed to 0.2 seconds in #6043. Raising it again would postpone the next failure without making the measurement deterministic.

Steps to reproduce

1. Run make test-python-unit, which starts 8 pytest workers.
2. Let test_performance_bounds_single_entity execute while the runner is under contention.
3. The wall-clock loop can exceed 0.2 seconds even though the implementation has not regressed.

The intermittent failure is captured in the linked master run. Local execution passes when the host is idle, which is consistent with runner contention.

Specifications

• Version: current master at f771ea4
• Failing platform: Python 3.10 on ubuntu-latest
• Subsystem: Python unit tests, entity key encoding
• File: sdk/python/tests/unit/infra/test_key_encoding_utils.py

Possible Solution

Remove test_performance_bounds_single_entity. The same change that introduced it also added dedicated pytest-benchmark coverage in sdk/python/tests/benchmarks/test_key_encoding_benchmarks.py for single-entity string and integer serialization and deserialization. Those benchmarks run in the separate benchmark workflow and preserve the performance signal without turning shared-runner latency into a unit-test failure.

The existing test_single_entity_fast_path continues to cover correctness of the optimized path.