forked from bentoml/BentoML
/
runner_app.py
333 lines (279 loc) · 11.8 KB
/
runner_app.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
from __future__ import annotations
import json
import pickle
import typing as t
import logging
import functools
from typing import TYPE_CHECKING
from functools import partial
from simple_di import inject
from simple_di import Provide
from bentoml._internal.types import LazyType
from ..context import trace_context
from ..context import component_context
from ..runner.utils import Params
from ..runner.utils import PAYLOAD_META_HEADER
from ..runner.utils import payload_paramss_to_batch_params
from ..utils.metrics import metric_name
from ..utils.metrics import exponential_buckets
from ..server.base_app import BaseAppFactory
from ..runner.container import Payload
from ..runner.container import AutoContainer
from ..marshal.dispatcher import CorkDispatcher
from ..configuration.containers import BentoMLContainer
feedback_logger = logging.getLogger("bentoml.feedback")
logger = logging.getLogger(__name__)
if TYPE_CHECKING:
from starlette.routing import BaseRoute
from starlette.requests import Request
from starlette.responses import Response
from starlette.middleware import Middleware
from opentelemetry.sdk.trace import Span
from ..runner.runner import Runner
from ..runner.runner import RunnerMethod
class RunnerAppFactory(BaseAppFactory):
@inject
def __init__(
self,
runner: Runner,
worker_index: int = 0,
enable_metrics: bool = Provide[BentoMLContainer.runners_config.metrics.enabled],
) -> None:
self.runner = runner
self.worker_index = worker_index
self.enable_metrics = enable_metrics
from starlette.responses import Response
TooManyRequests = partial(Response, status_code=429)
self.dispatchers: dict[str, CorkDispatcher] = {}
for method in runner.runner_methods:
if not method.config.batchable:
continue
self.dispatchers[method.name] = CorkDispatcher(
max_latency_in_ms=method.max_latency_ms,
max_batch_size=method.max_batch_size,
fallback=TooManyRequests,
)
@property
def name(self) -> str:
return self.runner.name
def _init_metrics_wrappers(self):
metrics_client = BentoMLContainer.metrics_client.get()
self.legacy_adaptive_batch_size_hist_map = {
method.name: metrics_client.Histogram(
name=metric_name(
self.runner.name,
self.worker_index,
method.name,
"adaptive_batch_size",
),
documentation="Legacy runner adaptive batch size",
labelnames=[],
buckets=exponential_buckets(1, 2, method.max_batch_size),
)
for method in self.runner.runner_methods
}
max_max_batch_size = max(
method.max_batch_size for method in self.runner.runner_methods
)
self.adaptive_batch_size_hist = metrics_client.Histogram(
namespace="bentoml_runner",
name="adaptive_batch_size",
documentation="Runner adaptive batch size",
labelnames=[
"runner_name",
"worker_index",
"method_name",
"service_version",
"service_name",
],
buckets=exponential_buckets(1, 2, max_max_batch_size),
)
@property
def on_startup(self) -> t.List[t.Callable[[], None]]:
on_startup = super().on_startup
on_startup.insert(0, functools.partial(self.runner.init_local, quiet=True))
on_startup.insert(0, self._init_metrics_wrappers)
return on_startup
@property
def on_shutdown(self) -> t.List[t.Callable[[], None]]:
on_shutdown = [self.runner.destroy]
for dispatcher in self.dispatchers.values():
on_shutdown.append(dispatcher.shutdown)
on_shutdown.extend(super().on_shutdown)
return on_shutdown
@property
def routes(self) -> t.List[BaseRoute]:
"""
Setup routes for Runner server, including:
/healthz liveness probe endpoint
/readyz Readiness probe endpoint
/metrics Prometheus metrics endpoint
For method in self.runner.runner_methods:
/{method.name} Run corresponding runnable method
/ Run the runnable method "__call__" if presented
"""
from starlette.routing import Route
routes = super().routes
for method in self.runner.runner_methods:
path = "/" if method.name == "__call__" else "/" + method.name
if method.config.batchable:
routes.append(
Route(
path=path,
endpoint=self._mk_request_handler(runner_method=method),
methods=["POST"],
)
)
else:
routes.append(
Route(
path=path,
endpoint=self.async_run(runner_method=method),
methods=["POST"],
)
)
return routes
@property
def middlewares(self) -> list[Middleware]:
middlewares = super().middlewares
# otel middleware
import opentelemetry.instrumentation.asgi as otel_asgi # type: ignore[import]
from starlette.middleware import Middleware
def client_request_hook(span: Span, _scope: t.Dict[str, t.Any]) -> None:
if span is not None:
span_id: int = span.context.span_id
trace_context.request_id = span_id
middlewares.append(
Middleware(
otel_asgi.OpenTelemetryMiddleware,
excluded_urls=BentoMLContainer.tracing_excluded_urls.get(),
default_span_details=None,
server_request_hook=None,
client_request_hook=client_request_hook,
tracer_provider=BentoMLContainer.tracer_provider.get(),
)
)
if self.enable_metrics:
from .instruments import RunnerTrafficMetricsMiddleware
middlewares.append(Middleware(RunnerTrafficMetricsMiddleware))
access_log_config = BentoMLContainer.runners_config.logging.access
if access_log_config.enabled.get():
from .access import AccessLogMiddleware
access_logger = logging.getLogger("bentoml.access")
if access_logger.getEffectiveLevel() <= logging.INFO:
middlewares.append(
Middleware(
AccessLogMiddleware,
has_request_content_length=access_log_config.request_content_length.get(),
has_request_content_type=access_log_config.request_content_type.get(),
has_response_content_length=access_log_config.response_content_length.get(),
has_response_content_type=access_log_config.response_content_type.get(),
)
)
return middlewares
def _mk_request_handler(
self,
runner_method: RunnerMethod[t.Any, t.Any, t.Any],
) -> t.Callable[[Request], t.Coroutine[None, None, Response]]:
from starlette.responses import Response
server_str = f"BentoML-Runner/{self.runner.name}/{runner_method.name}/{self.worker_index}"
async def infer_batch(
params_list: t.Sequence[Params[t.Any]],
) -> list[Payload] | list[tuple[Payload, ...]]:
self.legacy_adaptive_batch_size_hist_map[runner_method.name].observe( # type: ignore
len(params_list)
)
self.adaptive_batch_size_hist.labels( # type: ignore
runner_name=self.runner.name,
worker_index=self.worker_index,
method_name=runner_method.name,
service_version=component_context.bento_version,
service_name=component_context.bento_name,
).observe(len(params_list))
if not params_list:
return []
input_batch_dim, output_batch_dim = runner_method.config.batch_dim
batched_params, indices = payload_paramss_to_batch_params(
params_list, input_batch_dim
)
batch_ret = await runner_method.async_run(
*batched_params.args, **batched_params.kwargs
)
# multiple output branch
if LazyType["tuple[t.Any, ...]"](tuple).isinstance(batch_ret):
output_num = len(batch_ret)
payloadss = tuple(
AutoContainer.batch_to_payloads(
batch_ret[idx], indices, batch_dim=output_batch_dim
)
for idx in range(output_num)
)
ret = list(zip(*payloadss))
return ret
# single output branch
payloads = AutoContainer.batch_to_payloads(
batch_ret,
indices,
batch_dim=output_batch_dim,
)
return payloads
infer = self.dispatchers[runner_method.name](infer_batch)
async def _request_handler(request: Request) -> Response:
assert self._is_ready
r_: bytes = await request.body()
params: Params[t.Any] = pickle.loads(r_)
payload = await infer(params)
if not isinstance(
payload, Payload
): # a tuple, which means user runnable has multiple outputs
return Response(
pickle.dumps(payload),
headers={
PAYLOAD_META_HEADER: json.dumps({}),
"Content-Type": "application/vnd.bentoml.multiple_outputs",
"Server": server_str,
},
)
return Response(
payload.data,
headers={
PAYLOAD_META_HEADER: json.dumps(payload.meta),
"Content-Type": f"application/vnd.bentoml.{payload.container}",
"Server": server_str,
},
)
return _request_handler
def async_run(
self,
runner_method: RunnerMethod[t.Any, t.Any, t.Any],
) -> t.Callable[[Request], t.Coroutine[None, None, Response]]:
from starlette.responses import Response
async def _run(request: Request) -> Response:
assert self._is_ready
params = pickle.loads(await request.body())
params = params.map(AutoContainer.from_payload)
try:
ret = await runner_method.async_run(*params.args, **params.kwargs)
except Exception as exc: # pylint: disable=broad-except
logger.error(
f"Exception on runner '{runner_method.runner.name}' method '{runner_method.name}'",
exc_info=exc,
)
return Response(
status_code=500,
headers={
"Content-Type": "text/plain",
"Server": f"BentoML-Runner/{self.runner.name}/{runner_method.name}/{self.worker_index}",
},
)
else:
payload = AutoContainer.to_payload(ret, 0)
return Response(
payload.data,
headers={
PAYLOAD_META_HEADER: json.dumps(payload.meta),
"Content-Type": f"application/vnd.bentoml.{payload.container}",
"Server": f"BentoML-Runner/{self.runner.name}/{runner_method.name}/{self.worker_index}",
},
)
return _run