$ the-wire · showcase
Bigtable batchers stop hanging at exit, genai resumes continuation streams
By RepoJournal · Filed · About Google · Composed from the cited sources · methodology
Two Bigtable fixes remove a process-exit hang and a noisy metrics-exporter failure, while python-genai makes continuation-token resumption opt-in and stops recording half-finished stream turns in curated history.
fix(bigtable): avoid hang when batcher not closed on process exit googleapis/google-cloud-python
The sync batcher's internal ThreadPoolExecutor waited on flush tasks at process exit, and if the batcher was never closed the internal timer never got an exit signal, so the process could hang instead of terminating. CrossSync.create_task is now split so regular jobs still route through the ThreadPoolExecutor while shutdown work takes a separate path.
feat: allow automatic resumption of continuation token for in generate content and generate content stream. Default behavior for this feature is disabled googleapis/python-genai
generate_content and generate_content_stream can now resume automatically from a continuation token, addressing the case where a long response is cut short. The default is disabled, so existing callers see no behavior change until they turn it on.
fix: do not record incomplete continuation stream turns in curated chat history googleapis/python-genai
Incomplete continuation stream turns are no longer written into curated chat history, so a stream that dies mid-turn no longer leaves a broken entry that later turns build on.
fix(bigtable): fix authentication scopes for metrics exporter googleapis/google-cloud-python
The metrics client nested inside the Bigtable client shares its credentials, and when those scopes don't cover the metrics service every metric write fails and logs. The fix adds the metrics scope before client creation and disables the exporter after the first auth failure instead of logging on each write.
fix(interactions): restore `super().model_construct` in `Interaction` model validator googleapis/python-genai
Also worth noting: the Interaction model validator went back to super().model_construct, fixing an AttributeError: __pydantic_fields__ raised on every call because calling pydantic.BaseModel.model_construct bound cls to a class with no fields. The rest of the day was cleanup, including removal of the redundant TMABarrier class in jax [mgpu] and dead-code pruning in pallas pipelining and fuser, ...