Limitations

The following boundaries should always stay visible:

• not all query shapes benefit equally from range reads
• the write path and immutable-artifact model introduce their own tradeoffs
• performance depends heavily on data organization
• latency can increase if too many range reads are needed

Example limitation from the benchmarks

One key finding was that indexed_top can be byte-efficient while still being latency-risky if it requires too many separate range-read requests.

Original remote result:

indexed_top
  rangeReads: 46
  bytesRead: 5603
  worker p50: 571 ms

After clustered indexed layout and string-index page cache:

indexed_top
  rangeReads: 8
  bytesRead: 2563
  duration: 365 ms

Conclusion

Object size is no longer the primary optimization problem. The next optimization problem is R2 range-read latency.

What should not be generalized

• not every query will automatically work with just a few reads
• low bytes read alone does not guarantee a good user experience
• a read-only architecture does not solve write-path or ad hoc query problems