[fulmicoton]

There might be much better alternative but it really depends on the nature of your key.

Because the crux of S3 is the latency you can also decide to encode the docs in blocks, and retrieve more data than is actually needed.

For this demo, the index from DocID to offset in S3 takes 1.2 bytes per doc. For a log corpus, we end up with 0.2 bytes per doc.