When creating a dataframe from spark.sql and then using select throws an error. Expected to be able to select the column from the data.

Commands I ran (I am a total n00b):

• install rust via rustup
• Created a new crate: cargo new spark-connect-gist
• edited Cargo.toml to include spark connect rs and tokio dependencies:

[dependencies]
spark-connect-rs = "0.0.1-beta.3"
tokio = { version = "1", features = ["full"] }

• Install cmake: arch -arm64 brew install cmake
• Install protobuf: brew install protobuf
• Copy the example gist in main.rs: https://gist.github.com/sjrusso8/2b4e43af462367a15f91db5a33627449
• Run cargo run

Here is the error message:

   Compiling spark-connect-gist v0.1.0 (/Users/matthew.powers/Documents/code/my_apps/spark-connect-gist)
    Finished dev [unoptimized + debuginfo] target(s) in 1.00s
     Running `target/debug/spark-connect-gist`
Error: tonic::transport::Error(Transport, hyper::Error(Connect, ConnectError("tcp connect error", Os { code: 61, kind: ConnectionRefused, message: "Connection refused" })))

Hi @sjrusso8 can you tell me the rationale behind exposing a camel cased API?

Seems like we should be fine just using idiomatic rust snake casing.

On a separate note, do you have an IM channel of some kind to chat about this project? I would love to iterate more rapidly with you.

Description

Create similar bindings as with Rust but available in server side js (node, deno, bun, ...). The sdk should closely resemble the rust one, and only deviate when either necessary due to napi limitations, or when it is unidiomatic in JS.

napi.rs seems to be good crate to leverage and is relatively easy to use.

Early Experiment

The branch feat/napi contains a super quick pass at creating the bindings. The experiment only covers these areas

1. Create a remote SparkSession
2. Create a dataframe with .sql
3. Modify the dataframe with select, and filter
4. Perform "action" with count()
5. Perform “action” with show()

There is a lot of use of clone() and some not great implementations to create a new empty dataframe to satisfy the napi requirements. The polars js interop is a good example of how the bindings might function.

The collect() panics when returning a large result. The arrow-ipc streamreader is not parsing the data correctly.

Example:

use spark_connect_rs;

use spark_connect_rs::{SparkSession, SparkSessionBuilder};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut spark: SparkSession =
        SparkSessionBuilder::remote("sc://127.0.0.1:15002/;user_id=example_rs")
            .build()
            .await?;

    spark
        .clone()
        .range(None, 100000, 1, Some(1))
        .collect()
        .await
        .unwrap();

     Ok(())
}

This results in a panic

thread 'main' panicked at /home/sjrusso/Documents/code/projects/rust-projects/spark-connect-rs/src/session.rs:191:30:
called `Result::unwrap()` on an `Err` value: IpcError("Not expecting a schema when messages are read")
stack backtrace:
   0: rust_begin_unwind
             at /rustc/07dca489ac2d933c78d3c5158e3f43beefeb02ce/library/std/src/panicking.rs:645:5
   1: core::panicking::panic_fmt
             at /rustc/07dca489ac2d933c78d3c5158e3f43beefeb02ce/library/core/src/panicking.rs:72:14
   2: core::result::unwrap_failed
             at /rustc/07dca489ac2d933c78d3c5158e3f43beefeb02ce/library/core/src/result.rs:1649:5
   3: core::result::Result<T,E>::unwrap
             at /rustc/07dca489ac2d933c78d3c5158e3f43beefeb02ce/library/core/src/result.rs:1073:23
   4: spark_connect_rs::session::SparkSession::consume_plan::{{closure}}
             at ./src/session.rs:191:12
   5: spark_connect_rs::dataframe::DataFrame::collect::{{closure}}
             at ./src/dataframe.rs:99:14
   6: sql::main::{{closure}}
             at ./examples/sql.rs:21:10
   7: tokio::runtime::park::CachedParkThread::block_on::{{closure}}
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/park.rs:282:63
   8: tokio::runtime::coop::with_budget
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/coop.rs:107:5
   9: tokio::runtime::coop::budget
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/coop.rs:73:5
  10: tokio::runtime::park::CachedParkThread::block_on
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/park.rs:282:31
  11: tokio::runtime::context::blocking::BlockingRegionGuard::block_on
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/context/blocking.rs:66:9
  12: tokio::runtime::scheduler::multi_thread::MultiThread::block_on::{{closure}}
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/scheduler/multi_thread/mod.rs:87:13
  13: tokio::runtime::context::runtime::enter_runtime
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/context/runtime.rs:65:16
  14: tokio::runtime::scheduler::multi_thread::MultiThread::block_on
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/scheduler/multi_thread/mod.rs:86:9
  15: tokio::runtime::runtime::Runtime::block_on
             at /home/sjrusso/.cargo/registry/src/index.crates.io-6f17d22bba15001f/tokio-1.32.0/src/runtime/runtime.rs:349:45
  16: sql::main
             at ./examples/sql.rs:46:5
  17: core::ops::function::FnOnce::call_once
             at /rustc/07dca489ac2d933c78d3c5158e3f43beefeb02ce/library/core/src/ops/function.rs:250:5
note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace.

Description

The initial DataFrameWriter is created but there is also another way to write data via DataFrameWriterV2. This method has a slightly different implementation and leverages a different proto command message.

I think the methods should mirror the ones found on the Spark API guide, and a new method should be added onto the DataFrame for writeTo.

Description

Being able to compile the rust bindings into wasm32-unknown-unknown and/or wasm32-wasi would be interesting. This could allow some interesting interactions between a browser and spark. A wasm32-wasi target would allow spark programs to run on any runtime.

Early Experiments

A feature flag under the core bindings for wasm already exists and does compile successfully to those targets mentioned above. The issue arises when trying to send a HTTP1.1 request with grpc-web to the Spark Connect server, which only accepts normal HTTP2 grpc requests. There are methods of standing up a proxy server with envoy to forward the gRPC browser request to the backend server. But this feels like a lot of effort for the client to do.

The branch feat/wasm contains the early experiment and trying to run wasm with wasmtime. Issue arises with using async code in wasm. There is probably a way to code it correctly, but I don't have time to finish the experiment

Description

There are many functions for Spark, and most of them are created via a macro. However, not all of them have unit test coverage. Create additional unit tests based on similar test conditions from the exiting Spark API test cases.

I have been mirror the docstring tests from the PySpark API for reference.

Description

The overall documentation needs to be reviewed and matched against the Spark Core Classes and Functions. For instance, the README should be accurate to what functions and methods are currently implemented compared to the existing Spark API.

However, there are probably a few misses that are either currently implemented but marked as open, or were accidentally excluded. Might consider adding a few sections for other classes like StreamingQueryManager, DataFrameNaFunctions, DataFrameStatFunctions, etc.

Description

Implement the ability to use positional/keyword args with sql. Because of the differences between python and rust, the function arguments need to be clearly implemented.

The pyspark process for sql allows for literals and dataframes to be in one argument. However, rust probably won't take to kindly to that input arg. If a user passes in a DataFrame it will need to be handled with a SubqueryAlias and if it's a literal it will be passed in as either a positional or a keyword argument.

We might want to only allow for to inputs parameters are options. Something like this?

sql<T: ToLiteral>(self, sql_query: &str, col_args: Option<HashMap<String, T>>, df_args: Option<HashMap<String, DataFrame>>) -> DataFrame 

This could allow a user to do these variations.

spark.sql("SELECT * FROM table", None, None).await?;

let df = spark.range(...);

// create the hashmap

spark.sql("SELECT * FROM {df}" None, Some(df_hashmap)).await?;

let col = "name";

// create the hashmap

spark.sql("SELECT {col} FROM {df}", Some(col_hashmap), Some(df_hashmap)).await?;

Or should positional SQL be a completely different method all together? like sql_params? So that a user doesn't need to fuss with adding None x2 to all their sql statements.