The 30 second answer
At a million requests a second the log grew about 6 GB a minute and reached 175 GB in thirty minutes. The framework's async logger dropped lines when its buffer overflowed, rather than slowing requests down. So I'd size the buffer, flush on an interval, and alert on dropped lines. In the video, a 4 MB buffer flushed every two seconds fixed it, at the cost of a two-second delay.
The volume. One log line per request. In the 60-second test the log reached about 6 GB, so about 6 GB per minute; after the 30-minute run the file was 175 GB. Counting its lines with wc -l took a long time because it has to read the whole file. The half-terabyte disk was sized for this.
The silent loss. In the first 30-minute test the log had fewer lines than requests. The framework's logger is configured to drop log lines when its buffer overflows instead of slowing request handling, and it logs how many it dropped. Grepping for that message showed the gap.
The fix and its cost. A 4 MB buffer flushed every 2 seconds. A log line can now take up to two seconds to reach disk, which the speaker accepts. In the second 30-minute run no lines were dropped.
The design choice behind it. Drop-or-block is a real decision: blocking protects completeness but lets logging slow every request; dropping protects latency but loses data. Either way, count and alert on drops, and treat the log as part of the verified output.
Alternate interview scenarios
"Should request logs go to local disk at this rate?"
Local files are the cheapest write, but you need rotation (the app splits logs by date), disk headroom, and shipping to central storage. At this volume, sampling or aggregating most success lines is a fair product question to raise.