feat: sql jobs outputting to s3 + streaming for high-number of rows

backend/parsers/windmill-parser-sql/src/lib.rs

@@ -120,6 +120,60 @@ pub fn parse_db_resource(code: &str) -> Option<String> {
     cap.map(|x| x.get(1).map(|x| x.as_str().to_string()).unwrap())
 }
 
+#[derive(Clone, Copy, Debug)]
+pub enum S3ModeFormat {
+    Json,
+    Csv,
+    Parquet,
+}
+pub fn s3_mode_extension(format: S3ModeFormat) -> &'static str {
+    match format {
+        S3ModeFormat::Json => "json",
+        S3ModeFormat::Csv => "csv",
+        S3ModeFormat::Parquet => "parquet",
+    }
+}
+pub struct S3ModeArgs {
+    pub prefix: Option<String>,
+    pub storage: Option<String>,
+    pub format: S3ModeFormat,
+}
+pub fn parse_s3_mode(code: &str) -> anyhow::Result<Option<S3ModeArgs>> {
+    let cap = match RE_S3_MODE.captures(code) {
+        Some(x) => x,
+        None => return Ok(None),
+    };
+    let args_str = cap
+        .get(1)
+        .map(|x| x.as_str().to_string())
+        .unwrap_or_default();
+
+    let mut prefix = None;
+    let mut storage = None;
+    let mut format = S3ModeFormat::Json;
+
+    for kv in args_str.split(' ').map(|kv| kv.trim()) {
+        if kv.is_empty() {
+            continue;
+        }
+        let mut it = kv.split('=');
+        let (Some(key), Some(value)) = (it.next(), it.next()) else {
+            return Err(anyhow!("Invalid S3 mode argument: {}", kv));
+        };
+        match (key.trim(), value.trim()) {
+            ("prefix", _) => prefix = Some(value.to_string()),
+            ("storage", _) => storage = Some(value.to_string()),
+            ("format", "json") => format = S3ModeFormat::Json,
+            ("format", "parquet") => format = S3ModeFormat::Parquet,
+            ("format", "csv") => format = S3ModeFormat::Csv,
+            ("format", format) => return Err(anyhow!("Invalid S3 mode format: {}", format)),
+            (_, _) => return Err(anyhow!("Invalid S3 mode argument: {}", kv)),
+        }
+    }
+
+    Ok(Some(S3ModeArgs { prefix, storage, format }))
+}
+
 pub fn parse_sql_blocks(code: &str) -> Vec<&str> {
     let mut blocks = vec![];
     let mut last_idx = 0;
@@ -147,6 +201,7 @@ lazy_static::lazy_static! {
     static ref RE_NONEMPTY_SQL_BLOCK: Regex = Regex::new(r#"(?m)^\s*[^\s](?:[^-]|$)"#).unwrap();
 
     static ref RE_DB: Regex = Regex::new(r#"(?m)^-- database (\S+) *(?:\r|\n|$)"#).unwrap();
+    static ref RE_S3_MODE: Regex = Regex::new(r#"(?m)^-- s3( (.+))?*(?:\r|\n|$)"#).unwrap();
 
     // -- $1 name (type) = default
     static ref RE_ARG_MYSQL: Regex = Regex::new(r#"(?m)^-- \? (\w+) \((\w+)\)(?: ?\= ?(.+))? *(?:\r|\n|$)"#).unwrap();

backend/windmill-common/Cargo.toml

@@ -12,7 +12,7 @@ tantivy = []
 prometheus = ["dep:prometheus"]
 loki = ["dep:tracing-loki"]
 benchmark = []
-parquet = ["dep:object_store", "dep:aws-config", "dep:aws-sdk-sts"]
+parquet = ["dep:object_store", "dep:aws-config", "dep:aws-sdk-sts", "dep:datafusion"]
 aws_auth = ["dep:aws-sdk-sts", "dep:aws-config"]
 otel = ["dep:opentelemetry-semantic-conventions", "dep:opentelemetry-otlp", "dep:opentelemetry_sdk", 
     "dep:opentelemetry", "dep:tracing-opentelemetry", "dep:opentelemetry-appender-tracing", "dep:tonic"]
@@ -44,6 +44,9 @@ tracing = { workspace = true }
 axum = { workspace = true }
 hyper = { workspace = true }
 tokio = { workspace = true }
+tokio-stream.workspace = true
+tokio-util.workspace = true
+datafusion = { workspace = true, optional = true}
 reqwest = { workspace = true }
 tracing-subscriber = { workspace = true }
 lazy_static.workspace = true
@@ -67,6 +70,7 @@ async-stream.workspace = true
 const_format.workspace = true
 crc.workspace = true
 windmill-macros.workspace = true
+windmill-parser-sql.workspace = true
 jsonwebtoken.workspace = true
 backon.workspace = true
 

backend/windmill-common/src/s3_helpers.rs

@@ -1,9 +1,18 @@
 #[cfg(feature = "parquet")]
 use crate::error;
+use crate::error::to_anyhow;
+use crate::utils::rd_string;
 #[cfg(feature = "parquet")]
 use aws_sdk_sts::config::ProvideCredentials;
 #[cfg(feature = "parquet")]
 use axum::async_trait;
+use bytes::Bytes;
+use datafusion::arrow::array::{RecordBatch, RecordBatchWriter};
+use datafusion::arrow::error::ArrowError;
+use datafusion::arrow::json::writer::JsonArray;
+use datafusion::arrow::{csv, json};
+use datafusion::parquet::arrow::ArrowWriter;
+use futures::TryStreamExt;
 #[cfg(feature = "parquet")]
 use object_store::aws::AwsCredential;
 #[cfg(feature = "parquet")]
@@ -15,10 +24,14 @@ use object_store::{aws::AmazonS3Builder, ClientOptions};
 #[cfg(feature = "parquet")]
 use reqwest::header::HeaderMap;
 use serde::{Deserialize, Serialize};
+use std::io::Write;
 #[cfg(feature = "parquet")]
 use std::sync::Arc;
+use std::sync::Mutex;
 #[cfg(feature = "parquet")]
 use tokio::sync::RwLock;
+use tokio::task;
+use windmill_parser_sql::S3ModeFormat;
 
 #[cfg(feature = "parquet")]
 lazy_static::lazy_static! {
@@ -480,3 +493,169 @@ pub fn bundle(w_id: &str, hash: &str) -> String {
 pub fn raw_app(w_id: &str, version: &i64) -> String {
     format!("/home/rfiszel/raw_app/{}/{}", w_id, version)
 }
+
+// Originally used a Arc<Mutex<dyn RecordBatchWriter + Send>>
+// But cannot call .close() on it because it moves the value and the object is not Sized
+enum RecordBatchWriterEnum {
+    Parquet(ArrowWriter<ChannelWriter>),
+    Csv(csv::Writer<ChannelWriter>),
+    Json(json::Writer<ChannelWriter, JsonArray>),
+}
+
+impl RecordBatchWriter for RecordBatchWriterEnum {
+    fn write(&mut self, batch: &RecordBatch) -> Result<(), ArrowError> {
+        match self {
+            RecordBatchWriterEnum::Parquet(w) => w.write(batch).map_err(|e| e.into()),
+            RecordBatchWriterEnum::Csv(w) => w.write(batch),
+            RecordBatchWriterEnum::Json(w) => w.write(batch),
+        }
+    }
+
+    fn close(self) -> Result<(), ArrowError> {
+        match self {
+            RecordBatchWriterEnum::Parquet(w) => w.close().map_err(|e| e.into()).map(drop),
+            RecordBatchWriterEnum::Csv(w) => w.close(),
+            RecordBatchWriterEnum::Json(w) => w.close(),
+        }
+    }
+}
+
+pub async fn convert_json_line_stream<E: Into<anyhow::Error>>(
+    mut stream: impl TryStreamExt<Item = Result<serde_json::Value, E>> + Unpin,
+    output_format: S3ModeFormat,
+) -> anyhow::Result<impl TryStreamExt<Item = anyhow::Result<bytes::Bytes>>> {
+    const MAX_MPSC_SIZE: usize = 1000;
+
+    use datafusion::{execution::context::SessionContext, prelude::NdJsonReadOptions};
+    use futures::StreamExt;
+    use std::path::PathBuf;
+    use tokio::io::AsyncWriteExt;
+
+    let mut path = PathBuf::from(std::env::temp_dir());
+    path.push(format!("{}.json", rd_string(8)));
+    let path_str = path
+        .to_str()
+        .ok_or_else(|| anyhow::anyhow!("Invalid path"))?;
+
+    // Write the stream to a temporary file
+    let mut file: tokio::fs::File = tokio::fs::File::create(&path).await.map_err(to_anyhow)?;
+
+    while let Some(chunk) = stream.next().await {
+        match chunk {
+            Ok(chunk) => {
+                // Convert the chunk to bytes and write it to the file
+                let b: bytes::Bytes = serde_json::to_string(&chunk)?.into();
+                file.write_all(&b).await?;
+                file.write_all(b"\n").await?;
+            }
+            Err(e) => {
+                tokio::fs::remove_file(&path).await?;
+                return Err(e.into());
+            }
+        }
+    }
+
+    file.flush().await?;
+    file.sync_all().await?;
+    drop(file);
+
+    let ctx = SessionContext::new();
+    ctx.register_json(
+        "my_table",
+        path_str,
+        NdJsonReadOptions { ..Default::default() },
+    )
+    .await
+    .map_err(to_anyhow)?;
+
+    let df = ctx.sql("SELECT * FROM my_table").await.map_err(to_anyhow)?;
+    let schema = df.schema().clone().into();
+    let mut datafusion_stream = df.execute_stream().await.map_err(to_anyhow)?;
+
+    let (tx, rx) = tokio::sync::mpsc::channel(MAX_MPSC_SIZE);
+    let writer: Arc<Mutex<Option<RecordBatchWriterEnum>>> =
+        Arc::new(Mutex::new(Some(match output_format {
+            S3ModeFormat::Parquet => RecordBatchWriterEnum::Parquet(
+                ArrowWriter::try_new(ChannelWriter { sender: tx.clone() }, Arc::new(schema), None)
+                    .map_err(to_anyhow)?,
+            ),
+
+            S3ModeFormat::Csv => {
+                RecordBatchWriterEnum::Csv(csv::Writer::new(ChannelWriter { sender: tx.clone() }))
+            }
+            S3ModeFormat::Json => {
+                RecordBatchWriterEnum::Json(json::Writer::<_, JsonArray>::new(ChannelWriter {
+                    sender: tx.clone(),
+                }))
+            }
+        })));
+
+    // This spawn is so that the data is sent in the background. Else the function would deadlock
+    // when hitting the mpsc channel limit
+    task::spawn(async move {
+        while let Some(batch_result) = datafusion_stream.next().await {
+            let batch: RecordBatch = match batch_result {
+                Ok(batch) => batch,
+                Err(e) => {
+                    tracing::error!("Error in datafusion stream: {:?}", &e);
+                    match tx.send(Err(e.into())).await {
+                        Ok(_) => {}
+                        Err(e) => tracing::error!("Failed to write error to channel: {:?}", &e),
+                    }
+                    break;
+                }
+            };
+            let writer = writer.clone();
+            // Writer calls blocking_send which would crash if called from the async context
+            let write_result = task::spawn_blocking(move || {
+                // SAFETY: We await so the code is actually sequential, lock unwrap cannot panic
+                // Second unwrap is ok because we initialized the option with Some
+                writer.lock().unwrap().as_mut().unwrap().write(&batch)
+            })
+            .await;
+            match write_result {
+                Ok(Ok(_)) => {}
+                Ok(Err(e)) => {
+                    tracing::error!("Error writing batch: {:?}", &e);
+                    match tx.send(Err(e.into())).await {
+                        Ok(_) => {}
+                        Err(e) => tracing::error!("Failed to write error to channel: {:?}", &e),
+                    }
+                }
+                Err(e) => tracing::error!("Error in blocking task: {:?}", &e),
+            };
+        }
+        task::spawn_blocking(move || {
+            writer.lock().unwrap().take().unwrap().close()?;
+            drop(writer);
+            Ok::<_, anyhow::Error>(())
+        })
+        .await??;
+        drop(ctx);
+        tokio::fs::remove_file(&path).await?;
+        Ok::<_, anyhow::Error>(())
+    });
+
+    Ok(tokio_stream::wrappers::ReceiverStream::new(rx))
+}
+
+struct ChannelWriter {
+    sender: tokio::sync::mpsc::Sender<anyhow::Result<Bytes>>,
+}
+
+impl Write for ChannelWriter {
+    fn write(&mut self, buf: &[u8]) -> std::io::Result<usize> {
+        let data: Bytes = buf.to_vec().into();
+        self.sender.blocking_send(Ok(data)).map_err(|e| {
+            std::io::Error::new(
+                std::io::ErrorKind::BrokenPipe,
+                format!("Channel send error: {}", e),
+            )
+        })?;
+        Ok(buf.len())
+    }
+
+    fn flush(&mut self) -> std::io::Result<()> {
+        Ok(())
+    }
+}

backend/windmill-worker/Cargo.toml

@@ -90,6 +90,7 @@ deno_tls = { workspace = true, optional = true }
 deno_permissions = { workspace = true, optional = true }
 deno_io = { workspace = true, optional = true }
 deno_error = { workspace = true, optional = true }
+async-stream.workspace = true
 
 postgres-native-tls.workspace = true
 native-tls.workspace = true