The subscription type decides which failures you will see, so it is worth choosing from the failure modes, not from the tutorial.
StreamingPull keeps a long-lived stream open per subscriber and is the default in the high-level client libraries. It gives the best throughput and latency, and the client library handles lease extension. Its failure modes are client-side: too much outstanding data held in memory, streams reset by proxies, and lease extension that runs out. When messages come back repeatedly, we read the client's flow-control settings first.
Pull (unary) gives the most control: you decide when to request and when to ack. It suits batch-style consumers and workers that scale from the backlog. Failure modes are polling too slowly, a pool of workers each holding a batch, and acking after a long downstream call.
Push hands delivery to an HTTPS endpoint, which suits Cloud Run and functions. The service controls the rate based on your endpoint's responses, so endpoint latency becomes delivery rate. Failure modes are slow responses, authentication mismatch on the OIDC token audience, and cold starts that miss the ack deadline. Push subscriptions do not support exactly-once delivery (Google Cloud documentation, read 17 November 2026).
BigQuery and Cloud Storage subscriptions write messages straight to a table or bucket with no subscriber code. Failure modes are schema mismatch (messages that do not fit the table end in the dead-letter path or are retried), quota on the destination, and cost.
Our rule of thumb: use StreamingPull for sustained services, pull for scheduled or autoscaled workers, push when the consumer is already a request-driven service with predictable latency, and a direct export subscription when no transformation is needed. We review each subscription against the traffic shape and its handler, and say plainly when moving a subscription type would remove a class of incidents.