$ the-wire · showcase
Spanner drops an O(N^2) result loop, JAX adds ULP accuracy tests
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
![release-please[bot]](https://avatars.githubusercontent.com/in/40688?v=4)
8 people shipped this
The Google Cloud Python client removed a quadratic row-popping bottleneck in StreamedResultSet and added a lock-free fast path for multiplexed session acquisition, while JAX landed a ULP precision test suite for elementary math functions across CPU, GPU, and TPU.
Spanner's StreamedResultSet used to yield rows with `iter_rows.pop(0)` inside a loop. Since Python lists are contiguous arrays, popping index 0 shifts every remaining element on each row, so consuming large result sets carried quadratic overhead. The fix replaces it with direct iteration and detaches the list outright [1]. Separately, multiplexed session acquisition now bypasses initialization and session locks on steady-state queries, builds replacement sessions outside the lock during maintenance rotation, and holds the lock only for the pointer swap [2]. The maintenance loops also moved from sleep to event wait for immediate shutdown termination [2].
JAX added a ULP precision test suite for the unary elementary math functions `exp`, `log`, `sin`, `cos`, `tan`, `sinh`, `cosh`, and `tanh`, covering `bfloat16`, `float16`, `float32`, and `float64` on CPU, GPU, and TPU against reference implementations [3]. The suite ships `numerics_test_util.py` with JAX-based ULP distance calculation, signed ULP histogram binning, and top-k worst-case error reporting across exhaustive and sampled input ranges [3]. On the ROCm side, the project enabled more bazel tests in ROCm CI and raised workers from 32 to 64; the Triton Pallas GPU test was skipped due to a codegen bug breaking the atomic counter test, and `//tests:lru_cache_test` plus `//tests:export_harnesses_multi_platform_test_gpu` were skipped as flaky in rbe [4]. A companion change lets CI build and stamp all four ROCm wheels instead of pulling jax and jaxlib from the shared GCS nightly bucket, which removes the dependency on the health of that build process [5].
python-genai released 2.23.0, adding `from_environment` support for environment copying and dynamic turn completion via `interaction_status` in `live.py`'s `receive()` [6].
BigQuery's generated Markdown reference docs got a monkeypatch in `docs/conf.py` to fix broken table rendering: `sphinx-markdown-builder` unconditionally emits trailing newlines in table cell paragraphs, which breaks single-line GFM table rows and causes cell padding to be read as 4-space indented code [7]. The patch is an isolated workaround in `docs/conf.py` while a permanent fix in `gcp-sphinx-docfx-yaml` remains unscheduled [7].
Action items
- → Upgrade google-cloud-spanner to pick up the StreamedResultSet row iteration and multiplexed session fast path googleapis/google-cloud-python [plan]
- → Upgrade google-genai to 2.23.0 for from_environment support and interaction_status turn completion googleapis/python-genai [plan]
- → Watch the JAX ROCm CI coverage change if you build against ROCm wheels google/jax [monitor]
References
- [1] perf(spanner): optimize row popping in StreamedResultSet ↗ googleapis/google-cloud-python
- [2] perf(spanner): add fast-path for multiplexed session acquisition ↗ googleapis/google-cloud-python
- [3] Add ULP precision test suite for elementary math functions. ↗ google/jax
- [4] [ROCm] Enable more bazel tests in ROCm CI ↗ google/jax
- [5] [ROCm] Build and stamp the complete ROCm wheel set in CI ↗ google/jax
- [6] v2.23.0 ↗ googleapis/python-genai
- [7] docs(bigquery): fix table rendering in markdown docs ↗ googleapis/google-cloud-python
