GCP Data Engineer Question 12

Added: 2026-05-17

[00:00:00]GCP data engineering professional certification question 12 of 30. Your data flow streaming pipeline reads from Pub/Sub and computes per minute the total revenue. Some messages arrive late up to 5 minutes after the event time.

[00:00:14]This late data is causing incorrect totals in your output. What data flow feature would you use to fix this? Let us try to understand this question a bit more. The pipeline uses windowing to group events into per minute buckets.

[00:00:26]The issue is that some events arrive at 5-minute delay. By which point their window might have crossed and the output is already produced. The good news is we know the exact maximum delay that is 5 minutes. That means we can configure a precise tolerance for it. Here are our four choices given in the exam. Increase the Pub/Sub message retention period.

[00:00:45]That's option number one. Option number two is use data flow windowing with allowed lateness and triggers. Option number three is switch to batch processing to avoid late data. Option number four is add a 5-minute sleep delay before processing. Let's go through option one by one. Pub/Sub message retention controls how long messages wait in the queue before actually expiring. Once a message enters the data flow pipeline, the retention settings no longer apply here. This does not touch the late data problem. So, we are going to rule this out. Windowing with allowed lateness and triggers. This is the data flow feature that proposes how we can solve this problem. Now, let's keep this and maybe this is a good option. Let's go through the other two options for now. Switching to batch eliminates the streaming pipeline entirely. This is not a fix. This is really not solving the problem and moving to a complete new solution. Let's rule this option out. The fourth option is a 5-minute sleep delay without artificially holding every single message before processing, not just the late data ones. Again, here we are assuming that every message in every different flow can be delayed. So, this is not solving the problem. So, this option is out. So, we are left with one option. That is the windowing option.

[00:01:52]Let's [music] read that again. Now, windowing with allowed latency and trigger is correct. Configure a 5-minute allowed lateness on your window. The late messages arriving within their window get added to the right time bucket and trigger an update result.

[00:02:06]Watermarks handle event time tracking through the pipeline. This is the standard pattern for late data in the streaming [music] actually. So, the correct option is option B, dataflow windowing that allowed lateness and triggers. Check out our GCP data engineering course on codecloud.com.