lib.rs 34.7 KB
Newer Older
1
2
3
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

4
5
use futures::StreamExt;
use once_cell::sync::OnceCell;
6
use pyo3::IntoPyObjectExt;
7
use pyo3::exceptions::PyStopAsyncIteration;
8
use pyo3::types::{PyDict, PyString};
9
use pyo3::{exceptions::PyException, prelude::*};
10
use rand::seq::IteratorRandom as _;
11
use rs::pipeline::network::Ingress;
12
use std::net::{IpAddr, Ipv4Addr, SocketAddr, SocketAddrV4};
13
use std::path::PathBuf;
14
use std::time::Duration;
15
16
use std::{fmt::Display, sync::Arc};
use tokio::sync::Mutex;
17
use tracing::Instrument;
18

Neelay Shah's avatar
Neelay Shah committed
19
use dynamo_runtime::{
Ryan Olson's avatar
Ryan Olson committed
20
    self as rs, logging,
21
    pipeline::{
22
23
        EngineStream, ManyOut, SingleIn, context::Context as RsContext,
        network::egress::push_router::RouterMode as RsRouterMode,
24
    },
25
    protocols::annotated::Annotated as RsAnnotated,
26
    traits::DistributedRuntimeProvider,
27
28
};

Neelay Shah's avatar
Neelay Shah committed
29
use dynamo_llm::{self as llm_rs};
30
31
use dynamo_llm::{entrypoint::RouterConfig, kv_router::KvRouterConfig};

32
33
use crate::llm::local_model::ModelRuntimeConfig;

34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
#[pyclass(eq, eq_int)]
#[derive(Clone, Debug, PartialEq)]
pub enum RouterMode {
    RoundRobin,
    Random,
    KV,
}

impl From<RouterMode> for RsRouterMode {
    fn from(mode: RouterMode) -> Self {
        match mode {
            RouterMode::RoundRobin => Self::RoundRobin,
            RouterMode::Random => Self::Random,
            RouterMode::KV => Self::KV,
        }
    }
}
51

52
mod context;
53
mod engine;
54
mod http;
55
mod llm;
56
mod parsers;
57
mod planner;
58
mod prometheus_metrics;
59
60
61
62
63
64
65
66

type JsonServerStreamingIngress =
    Ingress<SingleIn<serde_json::Value>, ManyOut<RsAnnotated<serde_json::Value>>>;

static INIT: OnceCell<()> = OnceCell::new();

const DEFAULT_ANNOTATED_SETTING: Option<bool> = Some(true);

67
68
// Helper to get appropriate span for instrumentation - always emit spans
fn get_span_for_context(context: &context::Context, operation: &str) -> tracing::Span {
69
70
71
72
73
74
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        None,
    )
75
76
77
78
79
80
81
82
}

// Helper to create span for direct method with instance_id
fn get_span_for_direct_context(
    context: &context::Context,
    operation: &str,
    instance_id: &str,
) -> tracing::Span {
83
84
85
86
87
88
    logging::make_client_request_span(
        operation,
        context.inner().id(),
        context.trace_context(),
        Some(instance_id),
    )
89
90
}

91
92
93
94
95
/// A Python module implemented in Rust. The name of this function must match
/// the `lib.name` setting in the `Cargo.toml`, else Python will not be able to
/// import the module.
#[pymodule]
fn _core(m: &Bound<'_, PyModule>) -> PyResult<()> {
Yan Ru Pei's avatar
Yan Ru Pei committed
96
    m.add_function(wrap_pyfunction!(llm::kv::compute_block_hash_for_seq_py, m)?)?;
Ryan Olson's avatar
Ryan Olson committed
97
    m.add_function(wrap_pyfunction!(log_message, m)?)?;
98
    m.add_function(wrap_pyfunction!(register_llm, m)?)?;
99
100
    m.add_function(wrap_pyfunction!(llm::entrypoint::make_engine, m)?)?;
    m.add_function(wrap_pyfunction!(llm::entrypoint::run_input, m)?)?;
101

102
103
104
105
106
107
108
    m.add_class::<DistributedRuntime>()?;
    m.add_class::<CancellationToken>()?;
    m.add_class::<Namespace>()?;
    m.add_class::<Component>()?;
    m.add_class::<Endpoint>()?;
    m.add_class::<Client>()?;
    m.add_class::<AsyncResponseStream>()?;
109
    m.add_class::<llm::disagg_router::DisaggregatedRouter>()?;
110
111
112
    m.add_class::<llm::entrypoint::EntrypointArgs>()?;
    m.add_class::<llm::entrypoint::EngineConfig>()?;
    m.add_class::<llm::entrypoint::EngineType>()?;
113
114
    m.add_class::<llm::entrypoint::RouterConfig>()?;
    m.add_class::<llm::entrypoint::KvRouterConfig>()?;
115
    m.add_class::<llm::kv::WorkerMetricsPublisher>()?;
116
    m.add_class::<llm::model_card::ModelDeploymentCard>()?;
117
    m.add_class::<llm::local_model::ModelRuntimeConfig>()?;
118
119
    m.add_class::<llm::preprocessor::OAIChatPreprocessor>()?;
    m.add_class::<llm::backend::Backend>()?;
120
121
    m.add_class::<llm::kv::OverlapScores>()?;
    m.add_class::<llm::kv::KvIndexer>()?;
122
    m.add_class::<llm::kv::ApproxKvIndexer>()?;
123
124
125
    m.add_class::<llm::kv::EndpointKvMetrics>()?;
    m.add_class::<llm::kv::AggregatedMetrics>()?;
    m.add_class::<llm::kv::KvMetricsAggregator>()?;
126
    m.add_class::<llm::kv::KvEventPublisher>()?;
Yan Ru Pei's avatar
Yan Ru Pei committed
127
128
    m.add_class::<llm::kv::RadixTree>()?;
    m.add_class::<llm::kv::ZmqKvEventListener>()?;
129
130
    m.add_class::<llm::kv::ZmqKvEventPublisher>()?;
    m.add_class::<llm::kv::ZmqKvEventPublisherConfig>()?;
131
    m.add_class::<llm::kv::KvRecorder>()?;
132
133
    m.add_class::<http::HttpService>()?;
    m.add_class::<http::HttpAsyncEngine>()?;
134
    m.add_class::<context::Context>()?;
135
    m.add_class::<ModelType>()?;
136
    m.add_class::<ModelInput>()?;
137
138
139
140
    m.add_class::<llm::kv::ForwardPassMetrics>()?;
    m.add_class::<llm::kv::WorkerStats>()?;
    m.add_class::<llm::kv::KvStats>()?;
    m.add_class::<llm::kv::SpecDecodeStats>()?;
141
142
    m.add_class::<llm::kv::KvPushRouter>()?;
    m.add_class::<llm::kv::KvPushRouterStream>()?;
143
    m.add_class::<RouterMode>()?;
144
    m.add("__version__", env!("CARGO_PKG_VERSION"))?;
145
146
147
    m.add_class::<planner::VirtualConnectorCoordinator>()?;
    m.add_class::<planner::VirtualConnectorClient>()?;
    m.add_class::<planner::PlannerDecision>()?;
148
149

    engine::add_to_module(m)?;
150
    parsers::add_to_module(m)?;
151

152
    m.add_class::<prometheus_metrics::RuntimeMetrics>()?;
153
154
155
156
    let prometheus_metrics = PyModule::new(m.py(), "prometheus_metrics")?;
    prometheus_metrics::add_to_module(&prometheus_metrics)?;
    m.add_submodule(&prometheus_metrics)?;

157
158
159
    #[cfg(feature = "block-manager")]
    llm::block_manager::add_to_module(m)?;

160
161
162
163
164
165
166
167
168
169
    Ok(())
}

pub fn to_pyerr<E>(err: E) -> PyErr
where
    E: Display,
{
    PyException::new_err(format!("{}", err))
}

Ryan Olson's avatar
Ryan Olson committed
170
171
172
173
174
175
176
/// Log a message from Python with file and line info
#[pyfunction]
#[pyo3(text_signature = "(level, message, module, file, line)")]
fn log_message(level: &str, message: &str, module: &str, file: &str, line: u32) {
    logging::log_message(level, message, module, file, line);
}

177
#[pyfunction]
178
#[pyo3(signature = (model_input, model_type, endpoint, model_path, model_name=None, context_length=None, kv_cache_block_size=None, router_mode=None, migration_limit=0, runtime_config=None, user_data=None, custom_template_path=None))]
179
#[allow(clippy::too_many_arguments)]
180
181
fn register_llm<'p>(
    py: Python<'p>,
182
    model_input: ModelInput,
183
    model_type: ModelType,
184
185
186
    endpoint: Endpoint,
    model_path: &str,
    model_name: Option<&str>,
187
188
    context_length: Option<u32>,
    kv_cache_block_size: Option<u32>,
189
    router_mode: Option<RouterMode>,
190
    migration_limit: u32,
191
    runtime_config: Option<ModelRuntimeConfig>,
192
    user_data: Option<&Bound<'p, PyDict>>,
193
    custom_template_path: Option<&str>,
194
) -> PyResult<Bound<'p, PyAny>> {
195
196
197
    let model_input = match model_input {
        ModelInput::Text => llm_rs::model_type::ModelInput::Text,
        ModelInput::Tokens => llm_rs::model_type::ModelInput::Tokens,
198
        ModelInput::Tensor => llm_rs::model_type::ModelInput::Tensor,
199
200
    };

201
202
    let model_type_obj = model_type.inner;

203
204
    let inner_path = model_path.to_string();
    let model_name = model_name.map(|n| n.to_string());
205
206
207
    let router_mode = router_mode.unwrap_or(RouterMode::RoundRobin);
    let router_config = RouterConfig::new(router_mode.into(), KvRouterConfig::default());

208
209
210
211
212
213
214
215
216
217
218
219
220
    // Early validation of custom template path
    let custom_template_path_owned = custom_template_path
        .map(|s| {
            let path = PathBuf::from(s);
            if !path.exists() {
                return Err(PyErr::new::<pyo3::exceptions::PyFileNotFoundError, _>(
                    format!("Custom template file does not exist: {}", path.display()),
                ));
            }
            Ok(path)
        })
        .transpose()?;

221
222
223
224
225
226
227
    let user_data_json = user_data
        .map(|dict| pythonize::depythonize(dict))
        .transpose()
        .map_err(|err| {
            PyErr::new::<PyException, _>(format!("Failed to convert user_data: {}", err))
        })?;

228
    pyo3_async_runtimes::tokio::future_into_py(py, async move {
229
230
231
232
233
        let mut builder = dynamo_llm::local_model::LocalModelBuilder::default();
        builder
            .model_path(Some(PathBuf::from(inner_path)))
            .model_name(model_name)
            .context_length(context_length)
234
            .kv_cache_block_size(kv_cache_block_size)
235
            .router_config(Some(router_config))
236
            .migration_limit(Some(migration_limit))
237
            .runtime_config(runtime_config.unwrap_or_default().inner)
238
239
            .user_data(user_data_json)
            .custom_template_path(custom_template_path_owned);
240
        // Download from HF, load the ModelDeploymentCard
241
        let mut local_model = builder.build().await.map_err(to_pyerr)?;
242
243
        // Advertise ourself on etcd so ingress can find us
        local_model
244
            .attach(&endpoint.inner, model_type_obj, model_input)
245
246
247
248
249
250
251
            .await
            .map_err(to_pyerr)?;

        Ok(())
    })
}

252
253
#[pyclass]
#[derive(Clone)]
Ryan Olson's avatar
Ryan Olson committed
254
pub struct DistributedRuntime {
255
256
257
258
    inner: rs::DistributedRuntime,
    event_loop: PyObject,
}

Ryan Olson's avatar
Ryan Olson committed
259
260
impl DistributedRuntime {
    #[allow(dead_code)]
261
    pub(crate) fn inner(&self) -> &rs::DistributedRuntime {
Ryan Olson's avatar
Ryan Olson committed
262
263
264
265
        &self.inner
    }
}

266
#[pyclass]
267
#[derive(Clone)]
268
269
270
271
272
struct CancellationToken {
    inner: rs::CancellationToken,
}

#[pyclass]
273
#[derive(Clone)]
274
275
276
277
278
279
struct Namespace {
    inner: rs::component::Namespace,
    event_loop: PyObject,
}

#[pyclass]
280
#[derive(Clone)]
281
282
283
284
285
286
struct Component {
    inner: rs::component::Component,
    event_loop: PyObject,
}

#[pyclass]
287
#[derive(Clone)]
288
289
290
291
292
293
struct Endpoint {
    inner: rs::component::Endpoint,
    event_loop: PyObject,
}

#[pyclass]
294
#[derive(Clone)]
295
struct Client {
296
    router: rs::pipeline::PushRouter<serde_json::Value, RsAnnotated<serde_json::Value>>,
297
298
}

299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
#[pyclass]
#[derive(Clone, PartialEq)]
struct ModelType {
    inner: llm_rs::model_type::ModelType,
}

#[pymethods]
#[allow(non_upper_case_globals)]
impl ModelType {
    #[classattr]
    const Chat: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Chat,
    };
    #[classattr]
    const Completions: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Completions,
    };
    #[classattr]
    const Embedding: Self = ModelType {
        inner: llm_rs::model_type::ModelType::Embedding,
    };
320
321
322
323
    #[classattr]
    const TensorBased: Self = ModelType {
        inner: llm_rs::model_type::ModelType::TensorBased,
    };
324
325
326
327
328
329
330
331
332
333
334
335

    fn __or__(&self, other: &Self) -> Self {
        ModelType {
            inner: self.inner | other.inner,
        }
    }

    fn __str__(&self) -> String {
        self.inner.to_string()
    }
}

336
337
#[pyclass(eq, eq_int)]
#[derive(Clone, PartialEq)]
338
339
340
enum ModelInput {
    Text = 1,
    Tokens = 2,
341
    Tensor = 3,
342
343
}

344
345
346
#[pymethods]
impl DistributedRuntime {
    #[new]
347
    fn new(event_loop: PyObject, is_static: bool) -> PyResult<Self> {
348
        let worker = rs::Worker::from_settings().map_err(to_pyerr)?;
349
        INIT.get_or_try_init(|| {
350
            let primary = worker.tokio_runtime()?;
351
352
353
354
355
356
            pyo3_async_runtimes::tokio::init_with_runtime(primary)
                .map_err(|e| rs::error!("failed to initialize pyo3 static runtime: {:?}", e))?;
            rs::OK(())
        })
        .map_err(to_pyerr)?;

357
        let runtime = worker.runtime().clone();
358

359
360
361
362
363
364
        // Initialize logging in context where tokio runtime is available
        // otel exporter requires it
        runtime.secondary().block_on(async {
            rs::logging::init();
        });

365
366
367
368
369
370
371
372
373
374
375
        let inner =
            if is_static {
                runtime.secondary().block_on(
                    rs::DistributedRuntime::from_settings_without_discovery(runtime),
                )
            } else {
                runtime
                    .secondary()
                    .block_on(rs::DistributedRuntime::from_settings(runtime))
            };
        let inner = inner.map_err(to_pyerr)?;
376
377
378
379

        Ok(DistributedRuntime { inner, event_loop })
    }

Ryan Olson's avatar
Ryan Olson committed
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
    #[staticmethod]
    fn detached(py: Python) -> PyResult<Self> {
        let rt = rs::Worker::runtime_from_existing().map_err(to_pyerr)?;
        let handle = rt.primary();

        let inner = handle
            .block_on(rs::DistributedRuntime::from_settings(rt))
            .map_err(to_pyerr)?;

        Ok(DistributedRuntime {
            inner,
            event_loop: py.None(),
        })
    }

395
396
397
398
399
400
401
    fn namespace(&self, name: String) -> PyResult<Namespace> {
        Ok(Namespace {
            inner: self.inner.namespace(name).map_err(to_pyerr)?,
            event_loop: self.event_loop.clone(),
        })
    }

402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
    /// Allocate a contiguous block of ports from the specified range and atomically reserve them.
    /// Returns a list of all allocated ports in order.
    #[pyo3(signature = (namespace, port_min, port_max, block_size, context=None))]
    fn allocate_port_block<'p>(
        &self,
        py: Python<'p>,
        namespace: &str,
        port_min: u16,
        port_max: u16,
        block_size: u16,
        context: Option<String>, // Optional info to store alongside the reservation
    ) -> PyResult<Bound<'p, PyAny>> {
        const MAX_ALLOCATE_ATTEMPTS: usize = 100;
        if block_size == 0 {
            return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(
                "Block size must be at least 1",
            ));
        }

        let Some(etcd_client) = self.inner.etcd_client() else {
            return Err(PyErr::new::<PyException, _>(
                "Static workers should not need to reserve ports",
            ));
        };

        let min = port_min;
        let max = port_max;

        // Compute maximum valid starting port (inclusive)
        let max_start_port = max.saturating_sub(block_size.saturating_sub(1));
        if max_start_port < min {
            return Err(PyErr::new::<PyException, _>(format!(
                "Port range {min}-{max} is too small for block size {block_size}",
            )));
        }

        // Randomize candidate starting ports to reduce contention/races
        let candidate_count =
            (max_start_port - port_min + 1).min(MAX_ALLOCATE_ATTEMPTS as u16) as usize;
        let mut rng = rand::rng();
        let candidate_ports: Vec<u16> =
            (port_min..=max_start_port).choose_multiple(&mut rng, candidate_count);

        let local_ip = match local_ip() {
            Ok(ip) => ip,
            Err(err) => {
                return Err(PyErr::new::<PyException, _>(format!(
                    "Failed fetching local IP address: {err}"
                )));
            }
        };

        let context_bytes = context.map(|s| s.as_bytes().to_vec()).unwrap_or_default();
        let namespace = namespace.to_owned();

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            for (attempt_idx, start_port) in candidate_ports.into_iter().enumerate() {
                let end_port_exclusive = start_port + block_size;
                let ports_to_reserve: Vec<u16> = (start_port..end_port_exclusive).collect();

                // Hold/bind all ports in the block
                let mut sockets = Vec::with_capacity(ports_to_reserve.len());
                let mut bind_failed = false;

                for &port in &ports_to_reserve {
                    match bind_tcp_port(port) {
                        Ok(sock) => sockets.push(sock),
                        Err(e) => {
                            tracing::error!(
471
472
473
                                "Failed to bind to port block starting at {start_port} (attempt {}): {e}",
                                attempt_idx + 1,
                            );
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
                            bind_failed = true;
                            break;
                        }
                    }
                }

                if bind_failed {
                    // Let previously bound sockets drop here
                    if attempt_idx < candidate_count - 1 {
                        tokio::time::sleep(Duration::from_millis(10)).await;
                    }
                    continue;
                }

                // With sockets held, reserve in ETCD
                let mut reserved_keys = Vec::with_capacity(ports_to_reserve.len());
                let mut reservation_failed = false;
                for port in &ports_to_reserve {
                    let key = make_port_key(&namespace, local_ip, *port).map_err(to_pyerr)?;

                    if let Err(e) = etcd_client
                        .kv_create(&key, context_bytes.clone(), None)
                        .await
                    {
                        tracing::error!(
                            "Failed to reserve port block starting at {start_port} (attempt {}): {e}",
                            attempt_idx + 1,
                        );
                        reservation_failed = true;
                        break;
                    }
                    reserved_keys.push(key);
                }

                if reservation_failed {
                    // Cleanup partial reservations
                    for key in reserved_keys {
                        if let Err(e) = etcd_client.kv_delete(key.as_str(), None).await {
                            tracing::warn!("Failed to cleanup reserved port {key}: {e}");
                        }
                    }

                    // Sockets automatically released via RAII
                    if attempt_idx < candidate_count - 1 {
                        tokio::time::sleep(Duration::from_millis(10)).await;
                    }
                    continue;
                }

                // Success - sockets will be released automatically
                tracing::debug!("Reserved port block {ports_to_reserve:?}");
                return Ok(ports_to_reserve);
            }

            Err(PyErr::new::<PyException, _>(format!(
529
530
                "Failed to allocate and reserve a port block of size {block_size} from range {min}-{max} after {candidate_count} attempts"
            )))
531
532
533
        })
    }

534
535
536
537
538
539
540
    fn shutdown(&self) {
        self.inner.runtime().shutdown();
    }

    fn event_loop(&self) -> PyObject {
        self.event_loop.clone()
    }
541
542
543
544
545

    fn child_token(&self) -> CancellationToken {
        let inner = self.inner.runtime().child_token();
        CancellationToken { inner }
    }
546
547
}

548
549
550
551
552
553
554
555
556
557
558
559
560
561
// Bind a TCP port and return a socket held until dropped.
fn bind_tcp_port(port: u16) -> std::io::Result<socket2::Socket> {
    let sock = socket2::Socket::new(
        socket2::Domain::IPV4,
        socket2::Type::STREAM,
        Some(socket2::Protocol::TCP),
    )?;
    sock.set_reuse_address(true)?;
    let addr = SocketAddr::V4(SocketAddrV4::new(Ipv4Addr::UNSPECIFIED, port));
    sock.bind(&addr.into())?;
    Ok(sock)
}

fn make_port_key(namespace: &str, node_ip: IpAddr, port: u16) -> anyhow::Result<String> {
562
    Ok(format!("v1/{namespace}/ports/{node_ip}/{port}"))
563
564
565
566
567
568
569
570
571
572
573
574
}

fn local_ip() -> Result<IpAddr, local_ip_address::Error> {
    local_ip_address::local_ip().or_else(|err| match err {
        local_ip_address::Error::LocalIpAddressNotFound => {
            // Fall back to IPv6 if no IPv4 addresses are found
            local_ip_address::local_ipv6()
        }
        _ => Err(err),
    })
}

575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
#[pymethods]
impl CancellationToken {
    fn cancel(&self) {
        self.inner.cancel();
    }

    fn cancelled<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let token = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            token.cancelled().await;
            Ok(())
        })
    }
}

#[pymethods]
impl Component {
    fn endpoint(&self, name: String) -> PyResult<Endpoint> {
        let inner = self.inner.endpoint(name);
        Ok(Endpoint {
            inner,
            event_loop: self.event_loop.clone(),
        })
    }

    fn create_service<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let builder = self.inner.service_builder();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            let _ = builder.create().await.map_err(to_pyerr)?;
            Ok(())
        })
    }
607
608
609

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
610
611
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_component(self.inner.clone())
612
    }
613
614
615
616
}

#[pymethods]
impl Endpoint {
617
    #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
618
619
620
621
    fn serve_endpoint<'p>(
        &self,
        py: Python<'p>,
        generator: PyObject,
622
        graceful_shutdown: Option<bool>,
623
        metrics_labels: Option<Vec<(String, String)>>,
624
        health_check_payload: Option<&Bound<'p, PyDict>>,
625
626
627
628
629
630
    ) -> PyResult<Bound<'p, PyAny>> {
        let engine = Arc::new(engine::PythonAsyncEngine::new(
            generator,
            self.event_loop.clone(),
        )?);
        let ingress = JsonServerStreamingIngress::for_engine(engine).map_err(to_pyerr)?;
631
632
633
634
635
636
637
638
639
640
641
642
643

        // Convert Python dict to serde_json::Value if provided and validate it's an object
        let health_payload_json = health_check_payload
            .map(|dict| pythonize::depythonize::<serde_json::Value>(dict))
            .transpose()
            .map_err(|err| {
                pyo3::exceptions::PyTypeError::new_err(format!(
                    "Failed to convert health_check_payload: {}",
                    err
                ))
            })?;

        // Require an object/dict
644
645
646
647
648
649
        if let Some(ref payload) = health_payload_json
            && !payload.is_object()
        {
            return Err(pyo3::exceptions::PyTypeError::new_err(
                "health_check_payload must be a JSON object (dict)",
            ));
650
651
652
        }

        let mut builder = self
653
654
655
656
            .inner
            .endpoint_builder()
            .metrics_labels(metrics_labels)
            .handler(ingress);
657
658
659
660
661

        if let Some(payload) = health_payload_json {
            builder = builder.health_check_payload(payload);
        }

662
        let graceful_shutdown = graceful_shutdown.unwrap_or(true);
663
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
664
665
666
667
668
            builder
                .graceful_shutdown(graceful_shutdown)
                .start()
                .await
                .map_err(to_pyerr)?;
669
670
671
672
            Ok(())
        })
    }

673
674
675
    fn client<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let inner = self.inner.clone();
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
676
            let client = inner.client().await.map_err(to_pyerr)?;
677
678
679
680
681
682
            let push_router = rs::pipeline::PushRouter::<
                serde_json::Value,
                RsAnnotated<serde_json::Value>,
            >::from_client(client, Default::default())
            .await
            .map_err(to_pyerr)?;
683
684
685
            Ok(Client {
                router: push_router,
            })
686
687
        })
    }
688
689

    fn lease_id(&self) -> i64 {
690
691
692
693
694
        self.inner
            .drt()
            .primary_lease()
            .map(|l| l.id())
            .unwrap_or(0)
695
    }
696
697
698

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
699
700
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_endpoint(self.inner.clone())
701
    }
702
703
704
705
706
707
708
709
710
711
712
}

#[pymethods]
impl Namespace {
    fn component(&self, name: String) -> PyResult<Component> {
        let inner = self.inner.component(name).map_err(to_pyerr)?;
        Ok(Component {
            inner,
            event_loop: self.event_loop.clone(),
        })
    }
713
714
715

    /// Get a RuntimeMetrics helper for creating Prometheus metrics
    #[getter]
716
717
    fn metrics(&self) -> prometheus_metrics::RuntimeMetrics {
        prometheus_metrics::RuntimeMetrics::from_namespace(self.inner.clone())
718
    }
719
720
721
722
}

#[pymethods]
impl Client {
723
724
725
726
    /// Get list of current instances.
    /// Replaces endpoint_ids.
    fn instance_ids(&self) -> Vec<i64> {
        self.router.client.instance_ids()
727
728
    }

729
730
731
    /// Wait for an instance to be available for work.
    /// Replaces wait_for_endpoints.
    fn wait_for_instances<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
732
        let inner = self.router.client.clone();
733
        pyo3_async_runtimes::tokio::future_into_py(py, async move {
734
            inner
735
                .wait_for_instances()
736
737
738
                .await
                .map(|v| v.into_iter().map(|cei| cei.id()).collect::<Vec<i64>>())
                .map_err(to_pyerr)
739
740
741
742
        })
    }

    /// Issue a request to the endpoint using the default routing strategy.
743
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
744
745
746
747
748
    fn generate<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
749
        context: Option<context::Context>,
750
    ) -> PyResult<Bound<'p, PyAny>> {
751
        if self.router.client.is_static() {
752
            self.r#static(py, request, annotated, context)
753
        } else {
754
            self.random(py, request, annotated, context)
755
        }
756
757
758
    }

    /// Send a request to the next endpoint in a round-robin fashion.
759
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
760
761
762
763
764
    fn round_robin<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
765
        context: Option<context::Context>,
766
767
768
769
770
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
771
        let client = self.router.clone();
772
773

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
774
775
            let stream = match context {
                Some(context) => {
776
777
778
779
780
781
782
                    let request_ctx = RsContext::with_id(request, context.inner().id().to_string());

                    // Always instrument with appropriate span (none if no trace context)
                    let span = get_span_for_context(&context, "round_robin");
                    let stream_future = client.round_robin(request_ctx).instrument(span);

                    let stream = stream_future.await.map_err(to_pyerr)?;
783
784
785
786
787
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.round_robin(request.into()).await.map_err(to_pyerr)?,
            };
788
789
790
791
792
793
794
795
796
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Send a request to a random endpoint.
797
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
798
799
800
801
802
    fn random<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
803
        context: Option<context::Context>,
804
805
806
807
808
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
809
        let client = self.router.clone();
810
811

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
812
813
            let stream = match context {
                Some(context) => {
814
815
816
817
818
819
                    let request_ctx = RsContext::with_id(request, context.inner().id().to_string());

                    let span = get_span_for_context(&context, "random");
                    let stream_future = client.random(request_ctx).instrument(span);

                    let stream = stream_future.await.map_err(to_pyerr)?;
820
821
822
823
824
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.random(request.into()).await.map_err(to_pyerr)?,
            };
825
826
827
828
829
830
831
832
833
            tokio::spawn(process_stream(stream, tx));
            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }

    /// Directly send a request to a specific endpoint.
834
    #[pyo3(signature = (request, instance_id, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
835
836
837
838
    fn direct<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
839
        instance_id: i64,
840
        annotated: Option<bool>,
841
        context: Option<context::Context>,
842
843
844
845
846
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
847
        let client = self.router.clone();
848
849

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
850
851
            let stream = match context {
                Some(context) => {
852
853
854
855
856
857
858
                    let request_ctx = RsContext::with_id(request, context.inner().id().to_string());

                    let span =
                        get_span_for_direct_context(&context, "direct", &instance_id.to_string());
                    let stream_future = client.direct(request_ctx, instance_id).instrument(span);

                    let stream = stream_future.await.map_err(to_pyerr)?;
859
860
861
862
863
864
865
866
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client
                    .direct(request.into(), instance_id)
                    .await
                    .map_err(to_pyerr)?,
            };
867
868
869
870
871
872
873
874
875

            tokio::spawn(process_stream(stream, tx));

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
876
877

    /// Directly send a request to a pre-defined static worker
878
    #[pyo3(signature = (request, annotated=DEFAULT_ANNOTATED_SETTING, context=None))]
879
880
881
882
883
    fn r#static<'p>(
        &self,
        py: Python<'p>,
        request: PyObject,
        annotated: Option<bool>,
884
        context: Option<context::Context>,
885
886
887
888
889
    ) -> PyResult<Bound<'p, PyAny>> {
        let request: serde_json::Value = pythonize::depythonize(&request.into_bound(py))?;
        let annotated = annotated.unwrap_or(false);

        let (tx, rx) = tokio::sync::mpsc::channel(32);
890
        let client = self.router.clone();
891
892

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
893
894
            let stream = match context {
                Some(context) => {
895
896
897
898
899
900
                    let request_ctx = RsContext::with_id(request, context.inner().id().to_string());

                    let span = get_span_for_context(&context, "static");
                    let stream_future = client.r#static(request_ctx).instrument(span);

                    let stream = stream_future.await.map_err(to_pyerr)?;
901
902
903
904
905
                    context.inner().link_child(stream.context());
                    stream
                }
                _ => client.r#static(request.into()).await.map_err(to_pyerr)?,
            };
906
907
908
909
910
911
912
913
914

            tokio::spawn(process_stream(stream, tx));

            Ok(AsyncResponseStream {
                rx: Arc::new(Mutex::new(rx)),
                annotated,
            })
        })
    }
915
916
917
}

async fn process_stream(
918
    stream: EngineStream<RsAnnotated<serde_json::Value>>,
919
920
921
922
923
    tx: tokio::sync::mpsc::Sender<RsAnnotated<PyObject>>,
) {
    let mut stream = stream;
    while let Some(response) = stream.next().await {
        // Convert the response to a PyObject using Python's GIL
924
        let annotated: RsAnnotated<serde_json::Value> = response;
925
        let annotated: RsAnnotated<PyObject> = annotated.map_data(|data| {
926
            Python::with_gil(|py| match pythonize::pythonize(py, &data) {
927
928
                Ok(pyobj) => Ok(pyobj.into()),
                Err(e) => Err(e.to_string()),
929
            })
930
931
932
933
934
935
        });

        let is_error = annotated.is_error();

        // Send the PyObject through the channel or log an error
        if let Err(e) = tx.send(annotated).await {
936
            tracing::error!("Failed to send response: {:?}", e);
937
            break;
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
        }

        if is_error {
            break;
        }
    }
}

#[pyclass]
struct AsyncResponseStream {
    rx: Arc<Mutex<tokio::sync::mpsc::Receiver<RsAnnotated<PyObject>>>>,
    annotated: bool,
}

#[pymethods]
impl AsyncResponseStream {
    /// This method is required to implement the `AsyncIterator` protocol.
    #[pyo3(name = "__aiter__")]
    fn aiter(slf: PyRef<Self>, py: Python) -> PyResult<Py<PyAny>> {
        slf.into_py_any(py)
    }
    /// This method is required to implement the `AsyncIterator` protocol.
    #[pyo3(name = "__anext__")]
    fn next<'p>(&self, py: Python<'p>) -> PyResult<Bound<'p, PyAny>> {
        let rx = self.rx.clone();
        let annotated = self.annotated;

        pyo3_async_runtimes::tokio::future_into_py(py, async move {
            loop {
                let value = rx.lock().await.recv().await;
                match value {
                    Some(pyobj) => {
                        let pyobj = match pyobj.ok() {
                            Ok(pyobj) => pyobj,
                            Err(e) => {
                                return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(e));
                            }
                        };

                        if annotated {
                            let object = Annotated { inner: pyobj };
                            #[allow(deprecated)]
                            let object = Python::with_gil(|py| object.into_py(py));
                            return Ok(object);
                        } else {
                            match pyobj.data {
                                Some(data) => return Ok(data),
                                None => continue,
                            }
                        }
                    }
                    None => return Err(PyStopAsyncIteration::new_err("Stream exhausted")),
                }
            }
        })
    }
}

#[pyclass]
struct Annotated {
    inner: RsAnnotated<PyObject>,
}

#[pymethods]
impl Annotated {
    #[new]
    fn new(data: PyObject) -> Self {
        Annotated {
            inner: RsAnnotated::from_data(data),
        }
    }

    fn is_error(&self) -> bool {
        self.inner.is_error()
    }

    fn data(&self) -> Option<PyObject> {
        self.inner.data.clone()
    }

    fn event(&self) -> Option<String> {
        self.inner.event.clone()
    }

    fn comments(&self) -> Option<Vec<String>> {
        self.inner.comment.clone()
    }

    fn id(&self) -> Option<String> {
        self.inner.id.clone()
    }

    #[pyo3(name = "__repr__")]
    fn _repr(&self, py: Python) -> String {
        let data = self.inner.data.clone().map(|obj| {
            obj.call_method0(py, "__repr__")
                .and_then(|repr_obj| repr_obj.extract::<Py<PyString>>(py))
                .map(|py_str| py_str.to_string_lossy(py).into_owned())
                .unwrap_or_else(|_| "<failed_repr>".to_string())
        });

        format!(
            "Annotated(data={}, event={}, comment={:?}, id={})",
            data.unwrap_or_else(|| "<no_data>".to_string()),
            self.inner.event.as_deref().unwrap_or("None"),
            self.inner.comment.as_deref().unwrap_or(&[]),
            self.inner.id.as_deref().unwrap_or("None")
        )
    }
}