$ the-wire · showcase
Vertex AI client timeout fix, Compute v1 API regenerated
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
The day's most consequential shift is in generated clients: the Vertex AI interactions client stops defaulting to a 5s timeout, and the Compute v1 surface is rebuilt against the 2026-09-01 API with fields and enums removed.
fix: ensure Vertex AI interactions client uses default timeout=None instead of 5s. googleapis/python-genai
The Vertex AI interactions client used to default to a 5s timeout; it now passes timeout=None, removing the implicit short deadline that long-running interactions were hitting. If you had worked around it by setting an explicit timeout on every call, that workaround is no longer load-bearing.
feat: update Compute v1 API to version 2026-09-01 (#18523) googleapis/google-cloud-python
Compute v1 generated clients are rebuilt against the 2026-09-01 API. All AggregatedList RPCs now behave as return_partial_success=true and their return_partial_success fields are removed, several enums and fields that no longer exist in the API are gone, which is a breaking change for code that set or read those fields.
v2.26.0 googleapis/python-genai
by release-please[bot]
python-genai 2.26.0 adds Gemini 3.8 Flash TTS and Gemini 3.8 Flash Lite TTS models to the SDKs, plus gcs_metrics_uri for exporting tuning job metrics and an environments dedicated module. Pull it in if you are wiring TTS or shipping tuning metrics to GCS.
feat(gapic-generator): add rpc_name metadata to mixin schema definitions googleapis/google-cloud-python
MixinMethod now carries an rpc_name attribute (empty string by default) holding the canonical full gRPC name, such as google.longrunning.Operations/GetOperation. Templates that generate telemetry, span names, or interceptor routing for mixins no longer have to reconstruct that name themselves.
feat(bigquery): Declare Python 3.15 support (#18525) googleapis/google-cloud-python
bigquery declares Python 3.15 support; it had been held out of the main 3.15 rollout only because of its pyarrow dependency, which is optional. The classifier is now in the setup file and tests are adjusted.
[internal] refactor to remove private array._chunk_iter google/jax
Quiet on the rest of the stack: JAX removed the private array._chunk_iter workaround in favor of implementing __iter__ and __reversed__ directly now that dynamic slice is exposed through array.at[...].get(strategy='dynamic_slice'), while the Pallas tutorials and tests dropped the deprecated device_id_type= argument and the numerics and reducer test files were split and expanded.