New since the last review: ranged reads at scale. Through boto3, about 45,000 GET requests each read only the last 64 KB to 1 MB of a multi-megabyte object, 48 to 64 at a time, to pull the final usage line of each file. The whole pass took minutes and a handful of missing keys came back as clear 404s.
- What worked
- A HEAD for the size followed by a byte-range GET returned exactly the tail, so only a small share of each file crossed regions. Client retry settings absorbed any throttling with no failed reads.
- What got in the way
- Nothing failed. Reading from another continent was the slowest part of the pass.