Adding Function Score Query about tantivy

[adamreichold]

Other language bindings (like tantivy-py) can hardly inherit the tantivy's Query struct due to different language implementations. For example, for tantivy-py, though pyO3 can turn python objects into rust's struct, we have to first define all the utility classes before tantivy-py can consume.

Out of curiosity, do you plan to pass a Python callable to this eventually? If so, I fear from personal experience this might prohibitively slow due to the overhead of getting the GIL and crossing the Python-Rust boundary.

[alex-au-922]

Out of curiosity, do you plan to pass a Python callable to this eventually? If so, I fear from personal experience this might prohibitively slow due to the overhead of getting the GIL and crossing the Python-Rust boundary.

Yes, that's my ultimate goal. Actually I have tried to compile the source in tantivy-py as follow:

@numba.njit
def score_add_10(score: float) -> float:
    return score + 10

function_score_query = Query.function_score_query(
    const_score_query, lambda _, score, __: score_add_10(score)
)

and the numba trick works. I understand that currently pyO3 needs you to acquire a GIL in Python, but I'm not sure if this is still the case in the future. If this PR passes, more investigations should be done on the performance issue when calling Python from Rust with JIT / other compiled code. Other than that, I think this feature should exist while providing templates for other languages are just an extra benefit.

Also, as ElasticSearch's documentation said (which also same as Lucene), the function score query should be called only after a majority of documents are filtered out. I also expect users only map the scoring after the retrieval stage as iterating through all the documents is slow.

[adamreichold]

I understand that currently pyO3 needs you to acquire a GIL in Python, but I'm not sure if this is still the case in the future.

It certainly still does and even though we aware of nogil CPython builds, there are a lot of issues around that still unresolved, so I wouldn't hold my breath. This is particularly problematic as tantivy-py explicitly releases the GIL during search (to allow multi-threaded server in a Python application) which means that invoking a callback does not just mean checking that the GIL is held, but really acquiring the lock bouncing it around all search threads in the worst case.

Other than that, I think this feature should exist while providing templates for other languages are just an extra benefit.

I did not add this to argue against the feature itself, just wanted to share some unhappy experiences trying to inject behaviour as Python code into Rust code.

[alex-au-922]

I think for integrating python code there could be some alternatives. The first thing come up in my mind is that we can create some pre-built 'function factory' that perform function currying, so user just plug-in their parameters and the function is executed in rust. Say users want a y = m * pow(score, n) + C function.

For more complicated usecase, they might just create their own pyO3 distribution with the additional function signature that suits their case. Although this seems quite similar to implementing their own Query Struct, but still their work is much less that they don't need figure out the whole querying logic like Weight and Scorer.

[fulmicoton]

Closing

The business need is not strong enough, and the implementation details are not clear.
Please comment here if something like if you need something like this (scaling score in a non linear manner).