This is a common thing for BigQuery to claim. It's quite fast for an analytics database, but it is not a transactional database. Even for static data, the simplest queries on the tiniest data still take 1/10th of a second or so, and realistic queries run in the 1-5s range.

Basically you could use BigQuery to build tools supporting human-in-the-loop processes (BI reports, exec dashboards), and you could call those "real-time" to the humans involved. But it will not support transactional workloads, and it does not provide SLAs to support that. I don't know about this particular feature, I'm guessing seconds but maybe minutes.

From the documentation (https://cloud.google.com/datastream/docs/behavior-overview#p...):

> When using the default stream settings, the latency from reading the data in the source to streaming it into the destination is between 10 and 120 seconds.