[Optimization] Avoid disk I/O for STRLEN command on tiered storage strings

[asherlaau]

Description

While exploring the codebase for optimization opportunities, I noticed a TODO in src/server/string_family.cc:133 regarding the STRLEN command.

Currently, when a string is offloaded to tiered storage (SSD), the implementation triggers an asynchronous disk read (ReadTiered) just to determine the string length.

// src/server/string_family.cc:133
// TODO(vlad): Optimize to return co.Size() if no modify operations are present
// TODO(vlad): Omit decoding string to just query its length
if (const auto& co = it_res.value()->second; co.IsExternal()) {
  auto cb = [](string_view s) { return s.size(); };

  TieredStorage::TResult<size_t> fut = ReadTiered<size_t>(
      op_args.db_cntx.db_index, key, co, std::move(cb),
      op_args.shard->tiered_storage());
  return {std::move(fut)};
}

Proposed Optimization

For tiered storage objects, the CompactObject metadata already includes the size of the
string. We can optimize this by returning co.Size() directly when the object is external,
effectively bypassing the disk I/O latency.

I'm interested in working on this!
I have already set up the development environment, built Dragonfly locally, and passed the
string_family_test. Please let me know if I should proceed with a Pull Request.

[romange]

@dranikpg can comment more - we would implement it right away if it was so simple. I think the problem is with RMW operations like APPEND where your approach would return stale results. BTW - the comment there hints it.

having said that - it's still possible to implement the optimization, but you should be mindful of sequences that involve RMW and read (aka APPEND and STRLEN).