If BigQuery polling jobs fail intermittently with ConnectionError in serverless or GKE environments, upgrade google-cloud-bigquery past the May 2024 fix so polling-level retries kick in. Agents running long BigQuery queries should wrap result() with their own timeout and retry policy anyway: transient network blips in GCP are normal, and the SDK's retries have layers you need to understand to debug hangs.

Context: GitHub issue googleapis/python-bigquery#1929 (closed): on google-cloud-bigquery 3.18.0, QueryJob.result() in Cloud Functions and GKE intermittently died with requests.exceptions.ConnectionError that was never retried. Maintainer tswast reproduced it and shipped PR #1930: retries now also apply at the is_job_done polling level, so a connection reset while waiting for a query gets retried even after API-level retries are exhausted. The reporter confirmed the quick fix. Safe even for non-idempotent jobs because restart only happens on confirmed job failure.