Onnx Feature - Loading Model from Object Storage

[chloeminkyung]

Current Onnx Model Configuration Format

udfs:
   - name: is_fraudulent
     config: !Onnx
       path: ./models/model_file

Proposed New Format which is reusable

udfs:
   - type: !Onnx
     name: is_fraudulent
     config: !LocalStorage
       details:
         path: data/zones/taxi_zone_lookup.csv

   - type: !Onnx
     name: is_fraudulent
     config: !S3Stroage
       details:
         # optional
         access_key_id: {{YOUR_ACCESS_KEY}}
         # optional
         secret_access_key: {{YOUR_SECRET_KEY}}
         region: {{YOUR_REGION}}
         bucket_name: {{YOUR_BUCKET_NAME}}
       path: ./models/model_file

Original Object Storage Configuration Format

  - config : !LocalStorage
      details:
        path: data
      tables:
        - !Table
          name: yellow_trips
          config: !Parquet
            path: yellow_trips
            extension: .parquet
    name: yellow_taxi

  - config : !S3Storage
      details:
        access_key_id: {{YOUR_ACCESS_KEY}}
        secret_access_key: {{YOUR_SECRET_KEY}}
        region: {{YOUR_REGION}}
        bucket_name: {{YOUR_BUCKET_NAME}}
      tables:
        - !Table
          name: stocks
          config: !CSV
            path: .
            extension: .csv
    name: s3

[chubei]

1. I prefer if access_key_id and secret_access_key are not required, to make the use of public buckets easier.
2. I prefer option 1, but the enum name S3Onnx and LocalOnnx must be changed so this configuration format can be used in other places.
3. Option 2 is obviously not reusable, because onnx is embedded in the format.

[snork-alt]

I don't like this format. I would standardize the file loading across Dozer, whether is onnx or wasm and use URI formats.

Also I would consider using https://github.com/apache/incubator-opendal or something similar so we can support any type of file storage at once.

[Xuanwo]

Thank you for mentioning OpenDAL! I'm one of the opendal committers. Please let me know if there is any assistance we can offer.

[chloeminkyung]

Utilizing opendal will consist of following steps:

1. logical changes
2. configuration format changes (optional)

list of supported object store services - https://github.com/apache/incubator-opendal/tree/main/core/src/services

Opendal uses its own Result

enum Result<T, E> {
   Ok(T),
   Err(E),
}

opendal	require config changes	logical changes	dozer supported
s3	optional	supported	y
local	required	supported	y
gcs	required	supported	n
azblob	required	supported	n
..

async fn read_data(op: Operator) -> Result<()> {
    let bs = op.read("test").await?;
    println!("data: {}", String::from_utf8_lossy(&bs));

    Ok(())
}

S3

fn init() -> Result<Operator> {
    let mut builder = S3::default();
    builder.bucket("example");
    builder.access_key_id("access_key_id");
    builder.secret_access_key("secret_access_key");
    builder.region("us-east-1");

    let op = Operator::new(builder)?.finish();
    Ok(op)
}

config changes (optional)

config: !S3Stroage
     details:
         # optional
         access_key_id: {{YOUR_ACCESS_KEY}}
         # optional
         secret_access_key: {{YOUR_SECRET_KEY}}
         region: {{YOUR_REGION}}
         bucket_name: {{YOUR_BUCKET_NAME}}
     path: ./models/model_file

In-memory

fn init() -> Result<Operator> {
    let mut builder = Memory::default();
    builder.root("/tmp");

    let op: Operator = Operator::new(builder)?.finish();
    Ok(op)
}

config changes (optional)

config: !LocalStorage
      details:
        path: data/zones/taxi_zone_lookup.csv

[chubei]

Looks good to me

[Xuanwo]

The code of read(&self, path: &str, args: OpRead) is OpenDAL S3 Service internal implementaion detail and not expose to users. All services will be used in the same way like:

async fn read_data(op: Operator) -> Result<()> {
    let bs = op.read("test").await?;
    println!("data: {}", String::from_utf8_lossy(&bs));

    Ok(())
}

So ideally, we will have code like:

fn init_s3() -> Result<Operator> {
    let mut builder = S3::default();
    builder.bucket("example");
    builder.access_key_id("access_key_id");
    builder.secret_access_key("secret_access_key");
    builder.region("us-east-1");

    let op = Operator::new(builder)?.finish();
    Ok(op)
}

fn init_memory() -> Result<Operator> {
    let mut builder = Memory::default();
    builder.root("/tmp");

    let op: Operator = Operator::new(builder)?.finish();
    Ok(op)
}

async fn read_data(op: Operator) -> Result<()> {
    let bs = op.read("test").await?;
    println!("data: {}", String::from_utf8_lossy(&bs));

    Ok(())
}

And there is an ongoing issue to allow users init OpenDAL Operator in uri format like s3://bucket/path/to/root: apache/opendal#3022