$ the-wire · showcase
JAX 0.12.0 ships, and BigQuery stops rereading Arrow results
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
JAX cut a new minor release with new lax and scipy primitives while three BigQuery Arrow fixes land in google-cloud-python, the largest of which removes a redundant full re-read of query results.
fix(bigquery): keep the Arrow first page and wait without rows googleapis/google-cloud-python
The client now keeps the `jobs.query` first page of Arrow results and streams only the remainder, instead of dropping that page and rereading every row; it decides whether more rows remain from `pageToken` and waits with `maxResults=0`. Reading the results twice now returns the same rows.
JAX v0.12.0 google/jax
JAX v0.12.0 adds `jax.scipy.special.erfcinv` and `jax.lax.polynomial`, with the release notes stating the polynomial primitive is faster and uses less memory when computing gradients. It also adds `jax.custom_vjp.defremat` for customizing rematerialization under `jax.remat`, which requires the `jax_custom_vjp3` and `jax_remat3` implementations.
feat(bigquery): add timestamp_precision for Arrow query results googleapis/google-cloud-python
`timestamp_precision` reaches `query_and_wait()` for Arrow results, matching `Client.query()` and `Client.list_rows()`, so `TimestampPrecision.PICOSECOND` returns full-precision ISO 8601 strings rather than truncating to microseconds. Setting it without `ARROW` raises `ValueError`.
fix(bigquery): send the same Arrow options on jobs.query and ReadRows googleapis/google-cloud-python
`arrowSerializationOptions` now goes to both `jobs.query` and `ReadRows`, so the first page and the stream share one encoding. Previously `compression_codec` sat in `formatOptions`, which `jobs.query` ignores, and never reached `ReadRows`.
Add jax.scipy.special.gammaincinv and gammainccinv google/jax
New `jax.scipy.special.gammaincinv` and `gammainccinv` provide the quantile functions behind `chi2.ppf(p, df) == 2 * gammaincinv(df / 2, p)`, built on a DiDonato & Morris (1986) initial guess adapted from TFP's `tfp.math.igammainv` and refined with three Halley steps. Elsewhere: XLA's `cosh` on CPU and TPU now gets a vectorized C++ intrinsic reaching 1.0 ULP on float32, and python-genai's share...