feat(search): FT.PROFILE

[dranikpg]

Adds FT.PROFILE commands that runs the search query with profiling to see a detailed execution plan with timings. Does not comply with RS's FtProfile.

Example output for a simple query. First, it prints general results, then the detailed execution plan for each shard

[dranikpg]

I know we can compute more helpful metrics, but for now I'd like to output just the raw data

Everything (like non-leaf node overhead) can be calculated on top of that

[chakaz]

Nice!

[chakaz]

Consider using absl::Duration or the std equivalent

[dranikpg]

I was using absl::Now() first and its not monotonic. I actually detected it just by checking the results visually 😆 So I switched to chrono's steady clock.

I think the micros is fine. I don't see what the benefit is of using a scaringly long chrono type is if I just count the number of microseconds at the end

[chakaz]

What do you mean by non-monotonic? It represents a moment in time, regardless of timezones, and should never go back / forward in time
https://github.com/abseil/abseil-cpp/blob/fc44fa053cc91193d2bb60fb9356bcecb301242e/absl/time/time.h#L751-L752

Generally, I dislike having the units as a convention instead of part of the type, as it's error prone. Your call as always.

[dranikpg]

absl::Now queries the system clock which is not monotonic, that's why I switched to chrono (and converting chrono timepoint to asbl duration isn't that nice as well)

abseil/abseil-cpp#172 (comment)

[romange]

it can go back in time since it uses a real clock. It's better using steady/monotonic clocks when measuring latency.
+1 to @dranikpg preference to simplicity over standard. if to count microseconds I need to write:
std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::steady_clock::now() - start).count()
then maybe it's the standard committee that should do a better job.

[chakaz]

I didn't think of the case of the system changing time (I did think of daylight saving time changes, but that should be OK with absl::Now).
Anyway, I would use std::chrono::microseconds but I realize I'm in the minority. I saw so many bugs caused by bad unit conversions, some costlier than others. Sure, the type is verbose, but there are std typedefs to assist, and we can create one ourselves.
And.. in any case we need to write the long statement above (since absl::Now() is non monotonic), it's just a matter of storing a raw integer type vs a type-safe one
Having written that, @dranikpg you decide :)

[dranikpg]

We're not sending spaceships to Mars, we're debugging profiling queries 😆

Its has nothing to do with dates. The system clock is non monotonic on a microsecond level and jumps around a little. Because its implemented by some low level device / caches values / corrects itself 🤷🏻‍♂️

[kostasrim]

We're not sending spaceships to Mars, we're debugging profiling queries 😆

Not yet. SpaceFly on the way to Mars 🚀