binaries/runtime/src/operator/python.rs

#![allow(clippy::borrow_deref_ref)] // clippy warns about code generated by #[pymethods]

use super::{OperatorEvent, StopReason};
use dora_core::{
    config::{NodeId, OperatorId},
    descriptor::{source_is_url, Descriptor, PythonSource},
};
use dora_download::download_file;
use dora_node_api::Event;
use dora_operator_api_python::PyEvent;
use dora_operator_api_types::DoraStatus;
use eyre::{bail, eyre, Context, Result};
use pyo3::{
    pyclass,
    types::{IntoPyDict, PyAnyMethods, PyDict, PyTracebackMethods},
    Py, PyAny, Python,
};
use std::{
    panic::{catch_unwind, AssertUnwindSafe},
    path::Path,
};
use tokio::sync::{mpsc::Sender, oneshot};
use tracing::{error, field, span, warn};

fn traceback(err: pyo3::PyErr) -> eyre::Report {
    let traceback = Python::with_gil(|py| err.traceback_bound(py).and_then(|t| t.format().ok()));
    if let Some(traceback) = traceback {
        eyre::eyre!("{traceback}\n{err}")
    } else {
        eyre::eyre!("{err}")
    }
}

#[tracing::instrument(skip(events_tx, incoming_events), level = "trace")]
pub fn run(
    node_id: &NodeId,
    operator_id: &OperatorId,
    python_source: &PythonSource,
    events_tx: Sender<OperatorEvent>,
    incoming_events: flume::Receiver<Event>,
    init_done: oneshot::Sender<Result<()>>,
    dataflow_descriptor: &Descriptor,
) -> eyre::Result<()> {
    let path = if source_is_url(&python_source.source) {
        let target_path = Path::new("build")
            .join(node_id.to_string())
            .join(format!("{}.py", operator_id));
        // try to download the shared library
        let rt = tokio::runtime::Builder::new_current_thread()
            .enable_all()
            .build()?;
        rt.block_on(download_file(&python_source.source, &target_path))
            .wrap_err("failed to download Python operator")?;
        target_path
    } else {
        Path::new(&python_source.source).to_owned()
    };

    if !path.exists() {
        bail!("No python file exists at {}", path.display());
    }
    let path = path
        .canonicalize()
        .wrap_err_with(|| format!("no file found at `{}`", path.display()))?;
    let module_name = path
        .file_stem()
        .ok_or_else(|| eyre!("module path has no file stem"))?
        .to_str()
        .ok_or_else(|| eyre!("module file stem is not valid utf8"))?;
    let path_parent = path.parent();

    let send_output = SendOutputCallback {
        events_tx: events_tx.clone(),
    };

    let init_operator = move |py: Python| {
        if let Some(parent_path) = path_parent {
            let parent_path = parent_path
                .to_str()
                .ok_or_else(|| eyre!("module path is not valid utf8"))?;
            let sys = py
                .import_bound("sys")
                .wrap_err("failed to import `sys` module")?;
            let sys_path = sys
                .getattr("path")
                .wrap_err("failed to import `sys.path` module")?;
            let sys_path_append = sys_path
                .getattr("append")
                .wrap_err("`sys.path.append` was not found")?;
            sys_path_append
                .call1((parent_path,))
                .wrap_err("failed to append module path to python search path")?;
        }

        let module = py.import_bound(module_name).map_err(traceback)?;
        let operator_class = module
            .getattr("Operator")
            .wrap_err("no `Operator` class found in module")?;

        let locals = [("Operator", operator_class)].into_py_dict_bound(py);
        let operator = py
            .eval_bound("Operator()", None, Some(&locals))
            .map_err(traceback)?;
        operator.setattr(
            "dataflow_descriptor",
            pythonize::pythonize(py, dataflow_descriptor)?,
        )?;

        Result::<_, eyre::Report>::Ok(Py::from(operator))
    };

    let python_runner = move || {
        let mut operator =
            match Python::with_gil(init_operator).wrap_err("failed to init python operator") {
                Ok(op) => {
                    let _ = init_done.send(Ok(()));
                    op
                }
                Err(err) => {
                    let _ = init_done.send(Err(err));
                    bail!("Could not init python operator")
                }
            };

        let mut reload = false;
        let reason = loop {
            #[allow(unused_mut)]
            let Ok(mut event) = incoming_events.recv() else {
                break StopReason::InputsClosed;
            };

            if let Event::Reload { .. } = event {
                reload = true;
                // Reloading method
                #[allow(clippy::blocks_in_conditions)]
                match Python::with_gil(|py| -> Result<Py<PyAny>> {
                    // Saving current state
                    let current_state = operator
                        .getattr(py, "__dict__")
                        .wrap_err("Could not retrieve current operator state")?;
                    let current_state = current_state
                        .extract::<&PyDict>(py)
                        .wrap_err("could not extract operator state as a PyDict")?;
                    // Reload module
                    let module = py
                        .import_bound(module_name)
                        .map_err(traceback)
                        .wrap_err(format!("Could not retrieve {module_name} while reloading"))?;
                    let importlib = py
                        .import_bound("importlib")
                        .wrap_err("failed to import `importlib` module")?;
                    let module = importlib
                        .call_method("reload", (module,), None)
                        .wrap_err(format!("Could not reload {module_name} while reloading"))?;
                    let reloaded_operator_class = module
                        .getattr("Operator")
                        .wrap_err("no `Operator` class found in module")?;

                    // Create a new reloaded operator
                    let locals = [("Operator", reloaded_operator_class)].into_py_dict_bound(py);
                    let operator: Py<pyo3::PyAny> = py
                        .eval_bound("Operator()", None, Some(&locals))
                        .map_err(traceback)
                        .wrap_err("Could not initialize reloaded operator")?
                        .into();

                    // Replace initialized state with current state
                    operator
                        .getattr(py, "__dict__")
                        .wrap_err("Could not retrieve new operator state")?
                        .extract::<&PyDict>(py)
                        .wrap_err("could not extract new operator state as a PyDict")?
                        .update(current_state.as_mapping())
                        .wrap_err("could not restore operator state")?;

                    Ok(operator)
                }) {
                    Ok(reloaded_operator) => {
                        operator = reloaded_operator;
                    }
                    Err(err) => {
                        error!("Failed to reload operator.\n {err}");
                    }
                }
            }

            let status = Python::with_gil(|py| -> Result<i32> {
                let span = span!(tracing::Level::TRACE, "on_event", input_id = field::Empty);
                let _ = span.enter();

                // Add metadata context if we have a tracer and
                // incoming input has some metadata.
                #[cfg(feature = "telemetry")]
                if let Event::Input {
                    id: input_id,
                    metadata,
                    ..
                } = &mut event
                {
                    use dora_tracing::telemetry::{deserialize_context, serialize_context};
                    use tracing_opentelemetry::OpenTelemetrySpanExt;
                    span.record("input_id", input_id.as_str());

                    let cx = deserialize_context(&metadata.parameters.open_telemetry_context);
                    span.set_parent(cx);
                    let cx = span.context();
                    let string_cx = serialize_context(&cx);
                    metadata.parameters.open_telemetry_context = string_cx;
                }

                let py_event = PyEvent::from(event);

                let status_enum = operator
                    .call_method1(py, "on_event", (py_event, send_output.clone()))
                    .map_err(traceback);
                match status_enum {
                    Ok(status_enum) => {
                        let status_val = Python::with_gil(|py| status_enum.getattr(py, "value"))
                            .wrap_err("on_event must have enum return value")?;
                        Python::with_gil(|py| status_val.extract(py))
                            .wrap_err("on_event has invalid return value")
                    }
                    Err(err) => {
                        if reload {
                            // Allow error in hot reloading environment to help development.
                            warn!("{err}");
                            Ok(DoraStatus::Continue as i32)
                        } else {
                            Err(err)
                        }
                    }
                }
            })?;
            match status {
                s if s == DoraStatus::Continue as i32 => {} // ok
                s if s == DoraStatus::Stop as i32 => break StopReason::ExplicitStop,
                s if s == DoraStatus::StopAll as i32 => break StopReason::ExplicitStopAll,
                other => bail!("on_event returned invalid status {other}"),
            }
        };

        // Dropping the operator using Python garbage collector.
        // Locking the GIL for immediate release.
        Python::with_gil(|_py| {
            drop(operator);
        });

        Result::<_, eyre::Report>::Ok(reason)
    };

    let closure = AssertUnwindSafe(|| {
        python_runner().wrap_err_with(|| format!("error in Python module at {}", path.display()))
    });

    match catch_unwind(closure) {
        Ok(Ok(reason)) => {
            let _ = events_tx.blocking_send(OperatorEvent::Finished { reason });
        }
        Ok(Err(err)) => {
            let _ = events_tx.blocking_send(OperatorEvent::Error(err));
        }
        Err(panic) => {
            let _ = events_tx.blocking_send(OperatorEvent::Panic(panic));
        }
    }

    Ok(())
}

#[pyclass]
#[derive(Clone)]
struct SendOutputCallback {
    events_tx: Sender<OperatorEvent>,
}

#[allow(unsafe_op_in_unsafe_fn)]
mod callback_impl {

    use crate::operator::OperatorEvent;

    use super::SendOutputCallback;
    use aligned_vec::{AVec, ConstAlign};
    use arrow::{array::ArrayData, pyarrow::FromPyArrow};
    use dora_core::message::ArrowTypeInfo;
    use dora_node_api::{
        arrow_utils::{copy_array_into_sample, required_data_size},
        ZERO_COPY_THRESHOLD,
    };
    use dora_operator_api_python::pydict_to_metadata;
    use dora_tracing::telemetry::deserialize_context;
    use eyre::{eyre, Context, Result};
    use pyo3::{
        pymethods,
        types::{PyBytes, PyBytesMethods, PyDict},
        Bound, PyObject, Python,
    };
    use tokio::sync::oneshot;
    use tracing::{field, span};
    use tracing_opentelemetry::OpenTelemetrySpanExt;

    /// Send an output from the operator:
    /// - the first argument is the `output_id` as defined in your dataflow.
    /// - the second argument is the data as either bytes or pyarrow.Array for zero copy.
    /// - the third argument is dora metadata if you want to link the tracing from one input into an output.
    /// `e.g.:  send_output("bbox", pa.array([100], type=pa.uint8()), dora_event["metadata"])`
    #[pymethods]
    impl SendOutputCallback {
        fn __call__(
            &mut self,
            output: &str,
            data: PyObject,
            metadata: Option<Bound<'_, PyDict>>,
            py: Python,
        ) -> Result<()> {
            let parameters = pydict_to_metadata(metadata)
                .wrap_err("failed to parse metadata")?
                .into_owned();
            let span = span!(
                tracing::Level::TRACE,
                "send_output",
                output_id = field::Empty
            );
            span.record("output_id", output);

            let cx = deserialize_context(&parameters.open_telemetry_context);
            span.set_parent(cx);
            let _ = span.enter();

            let allocate_sample = |data_len| {
                if data_len > ZERO_COPY_THRESHOLD {
                    let (tx, rx) = oneshot::channel();
                    self.events_tx
                        .blocking_send(OperatorEvent::AllocateOutputSample {
                            len: data_len,
                            sample: tx,
                        })
                        .map_err(|_| eyre!("failed to send output to runtime"))?;
                    rx.blocking_recv()
                        .wrap_err("failed to request output sample")?
                        .wrap_err("failed to allocate output sample")
                } else {
                    let avec: AVec<u8, ConstAlign<128>> = AVec::__from_elem(128, 0, data_len);

                    Ok(avec.into())
                }
            };

            let (sample, type_info) = if let Ok(py_bytes) = data.downcast_bound::<PyBytes>(py) {
                let data = py_bytes.as_bytes();
                let mut sample = allocate_sample(data.len())?;
                sample.copy_from_slice(data);
                (sample, ArrowTypeInfo::byte_array(data.len()))
            } else if let Ok(arrow_array) = ArrayData::from_pyarrow_bound(data.bind(py)) {
                let total_len = required_data_size(&arrow_array);
                let mut sample = allocate_sample(total_len)?;

                let type_info = copy_array_into_sample(&mut sample, &arrow_array);

                (sample, type_info)
            } else {
                eyre::bail!("invalid `data` type, must by `PyBytes` or arrow array")
            };

            py.allow_threads(|| {
                let event = OperatorEvent::Output {
                    output_id: output.to_owned().into(),
                    type_info,
                    parameters,
                    data: Some(sample),
                };
                self.events_tx
                    .blocking_send(event)
                    .map_err(|_| eyre!("failed to send output to runtime"))
            })?;

            Ok(())
        }
    }
}