PointStruct is very slow

[dacox]

At first I thought the bottleneck in my indexing code was data access, but it turns out that it is actually the python_client PointStruct.

from qdrant_client.http.models import PointStruct

def f(embeddings):
    points = []
    for embedding in embeddings:
        points.append({
            'id': embedding.id,
            'vector': embedding.vector,
            'payload': {
                "asset_uuid": embedding.asset_uuid,
                "type": embedding.get_type(),
                "source": embedding.get_source(),
            },
        })
    return points

def g(embeddings):
    points = []
    for embedding in embeddings:
        points.append(PointStruct(
            id=embedding.id,
            vector=embedding.vector,
            payload={
                "asset_uuid": embedding.asset_uuid,
                "type": embedding.get_type(),
                "source": embedding.get_source(),
            },
        ))
    return points

Here are the timing comparisons:

%timeit _ = f(embeddings)
12.8 ms ± 243 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

%timeit _ = g(embeddings)
2.02 s ± 48.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

It's over 150x slower!

[joein]

Hi, could you please check which pydantic version are you using?

[dacox]

Hi, could you please check which pydantic version are you using?

@joein

Looking at our poetry.lock, it seems to be

[[package]]
name = "pydantic"
version = "1.10.13"
description = "Data validation and settings management using python type hints"
optional = false
python-versions = ">=3.7"

[joein]

pydantic v1 is indeed very slow, however I would not expect it to be this slow. Could you install pydantic v2 or do you have some blocking dependency?

[dacox]

@joein looking at the lockfile, its not immediately clear to me why were on v1. Will try with v2 and report back

[dacox]

@joein ok, i ran some tests with v2. I'm seeing g execute in more like 100ms now.

However, f is down to 2ms 😆 So its still 50x slower, but I'm not really sure how to explain the speedup of the original method after upgrading pydantic

[joein]

serialization and deserialization in pydantic v2 is written in rust

I think it is the cost of data validation

If you don't want to validate your data in runtime, you can use PointStruct.construct, then you will have static type checking in your IDE and no runtime validation

[joein]

Do you mind if I convert this issue into a discussion?

[dacox]

Do you mind if I convert this issue into a discussion?

go for it!