capo-sagemaker-runtime-http2 0.11.0__tar.gz → 0.13.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (56) hide show
  1. {capo_sagemaker_runtime_http2-0.11.0/src/capo_sagemaker_runtime_http2.egg-info → capo_sagemaker_runtime_http2-0.13.0}/PKG-INFO +3 -1
  2. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/README.md +2 -0
  3. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/pyproject.toml +1 -1
  4. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_services/_aws_config.py +15 -0
  5. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_services/_pipeline.py +207 -0
  6. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_services/async_sage_maker_runtime_http2.py +14 -1
  7. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_services/sage_maker_runtime_http2.py +14 -1
  8. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/request_stream_event.py +2 -2
  9. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/response_stream_event.py +6 -6
  10. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0/src/capo_sagemaker_runtime_http2.egg-info}/PKG-INFO +3 -1
  11. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/LICENSE +0 -0
  12. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/setup.cfg +0 -0
  13. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/__init__.py +0 -0
  14. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_async.py +0 -0
  15. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_auth/_identity.py +0 -0
  16. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_auth/_providers.py +0 -0
  17. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_auth/_signers.py +0 -0
  18. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_auth/_sigv4.py +0 -0
  19. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_auth/_zapros_handler.py +0 -0
  20. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_aws.py +0 -0
  21. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_body.py +0 -0
  22. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_iter.py +0 -0
  23. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_operations/amazon_sage_maker_runtime_http2/invoke_endpoint_with_bidirectional_stream.py +0 -0
  24. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_pagination.py +0 -0
  25. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_protocol/__init__.py +0 -0
  26. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_protocol/errors.py +0 -0
  27. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_protocol/eventstream.py +0 -0
  28. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_protocol/serialize.py +0 -0
  29. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_protocol/xml.py +0 -0
  30. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_rule_engine/__init__.py +0 -0
  31. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_rule_engine/_aws_partition.py +0 -0
  32. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_rule_engine/_endpoint_rule_set.py +0 -0
  33. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/_rule_engine/_endpoint_runtime.py +0 -0
  34. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/errors/__init__.py +0 -0
  35. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/errors/_base.py +0 -0
  36. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/errors/input_validation_error.py +0 -0
  37. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/errors/internal_server_error.py +0 -0
  38. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/errors/internal_stream_failure.py +0 -0
  39. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/errors/model_error.py +0 -0
  40. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/errors/model_stream_error.py +0 -0
  41. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/errors/service_unavailable_error.py +0 -0
  42. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/py.typed +0 -0
  43. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/_prelude/blob.py +0 -0
  44. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/_prelude/timestamp.py +0 -0
  45. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/_unknown.py +0 -0
  46. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/invoke_endpoint_with_bidirectional_stream_input.py +0 -0
  47. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/invoke_endpoint_with_bidirectional_stream_output.py +0 -0
  48. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/request_payload_part.py +0 -0
  49. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/response_payload_part.py +0 -0
  50. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2/types/sensitive_blob.py +0 -0
  51. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2.egg-info/SOURCES.txt +0 -0
  52. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2.egg-info/dependency_links.txt +0 -0
  53. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2.egg-info/requires.txt +0 -0
  54. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/src/capo_sagemaker_runtime_http2.egg-info/top_level.txt +0 -0
  55. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/tests/test_body.py +0 -0
  56. {capo_sagemaker_runtime_http2-0.11.0 → capo_sagemaker_runtime_http2-0.13.0}/tests/test_sigv4.py +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: capo-sagemaker-runtime-http2
3
- Version: 0.11.0
3
+ Version: 0.13.0
4
4
  Summary: Typed, async Python SDK for AWS Sagemaker Runtime Http2.
5
5
  Project-URL: Homepage, https://capo-sdk.dev/
6
6
  Project-URL: Documentation, https://capo-sdk.dev/services/sagemaker-runtime-http2/
@@ -108,6 +108,8 @@ The SDK retries failed operations automatically. Retry behaviour follows the Smi
108
108
 
109
109
  The number of attempts defaults to 3 and can be changed at the client level via `retry_max_attempts`, or per call via `config_overrides`.
110
110
 
111
+ `retry_mode="adaptive"` (or `AWS_RETRY_MODE=adaptive`, or `retry_mode = adaptive` in the AWS profile) also slows the whole client down once the service starts throttling it, then speeds back up as requests succeed. The default is `standard`.
112
+
111
113
  ```python
112
114
  from capo_sagemaker_runtime_http2 import AsyncSageMakerRuntimeHTTP2Client
113
115
 
@@ -76,6 +76,8 @@ The SDK retries failed operations automatically. Retry behaviour follows the Smi
76
76
 
77
77
  The number of attempts defaults to 3 and can be changed at the client level via `retry_max_attempts`, or per call via `config_overrides`.
78
78
 
79
+ `retry_mode="adaptive"` (or `AWS_RETRY_MODE=adaptive`, or `retry_mode = adaptive` in the AWS profile) also slows the whole client down once the service starts throttling it, then speeds back up as requests succeed. The default is `standard`.
80
+
79
81
  ```python
80
82
  from capo_sagemaker_runtime_http2 import AsyncSageMakerRuntimeHTTP2Client
81
83
 
@@ -1,6 +1,6 @@
1
1
  [project]
2
2
  name = "capo-sagemaker-runtime-http2"
3
- version = "0.11.0"
3
+ version = "0.13.0"
4
4
  description = "Typed, async Python SDK for AWS Sagemaker Runtime Http2."
5
5
  readme = "README.md"
6
6
  requires-python = ">=3.10"
@@ -31,6 +31,7 @@ class AwsSettings:
31
31
  use_fips: bool
32
32
  use_dual_stack: bool
33
33
  max_attempts: int
34
+ retry_mode: str
34
35
 
35
36
 
36
37
  def _env_bool(name: str) -> bool | None:
@@ -109,12 +110,22 @@ def load_aws_settings() -> AwsSettings:
109
110
  use_dual_stack = False
110
111
  max_attempts_raw = os.environ.get("AWS_MAX_ATTEMPTS") or profile.get("max_attempts")
111
112
  max_attempts = int(max_attempts_raw) if max_attempts_raw else 3
113
+ retry_mode = (
114
+ (os.environ.get("AWS_RETRY_MODE") or profile.get("retry_mode") or "standard")
115
+ .strip()
116
+ .lower()
117
+ )
118
+ if retry_mode not in ("legacy", "standard", "adaptive"):
119
+ raise ValueError(
120
+ f"invalid retry_mode {retry_mode!r}: expected 'legacy', 'standard' or 'adaptive'"
121
+ )
112
122
  return AwsSettings(
113
123
  region=region,
114
124
  endpoint=endpoint,
115
125
  use_fips=use_fips,
116
126
  use_dual_stack=use_dual_stack,
117
127
  max_attempts=max_attempts,
128
+ retry_mode=retry_mode,
118
129
  )
119
130
 
120
131
 
@@ -134,6 +145,8 @@ def aws_config() -> Interceptor[TInput, TOutput]:
134
145
  options.use_dual_stack = settings.use_dual_stack
135
146
  if options.retry_max_attempts is None:
136
147
  options.retry_max_attempts = settings.max_attempts
148
+ if options.retry_mode is None:
149
+ options.retry_mode = settings.retry_mode
137
150
  return next(request)
138
151
 
139
152
  return interceptor
@@ -155,6 +168,8 @@ def aaws_config() -> AsyncInterceptor[TInput, TOutput]:
155
168
  options.use_dual_stack = settings.use_dual_stack
156
169
  if options.retry_max_attempts is None:
157
170
  options.retry_max_attempts = settings.max_attempts
171
+ if options.retry_mode is None:
172
+ options.retry_mode = settings.retry_mode
158
173
  return await next(request)
159
174
 
160
175
  return interceptor
@@ -1,6 +1,7 @@
1
1
  from __future__ import annotations
2
2
 
3
3
  import logging
4
+ import math
4
5
  import random
5
6
  import threading
6
7
  import time
@@ -88,15 +89,134 @@ class ThreadSafeRetryQuota(RetryQuota):
88
89
  super().release(last_retry_cost)
89
90
 
90
91
 
92
+ class RateLimiter:
93
+ """Send tokens shared by all operations of one client in adaptive retry mode: the rate drops on throttling and regrows on success."""
94
+
95
+ MIN_FILL_RATE: float = 0.5
96
+ MIN_CAPACITY: float = 1.0
97
+ BETA: float = 0.7
98
+ SCALE_CONSTANT: float = 0.4
99
+ SMOOTHING: float = 0.8
100
+ MAX_RATE_SCALE: float = 2.0
101
+
102
+ def __init__(self, clock: Callable[[], float] = time.monotonic) -> None:
103
+ self._clock = clock
104
+ now = clock()
105
+ self._enabled: bool = False
106
+ self._fill_rate: float = 1.0
107
+ self._max_capacity: float = 1.0
108
+ self._capacity: float = 0.0
109
+ self._last_refill: float = now
110
+ self._last_max_rate: float = 0.0
111
+ self._last_throttle: float = now
112
+ self._measured_rate: float = 0.0
113
+ self._rate_bucket: float = math.floor(now)
114
+ self._request_count: int = 0
115
+
116
+ def acquire(self) -> float:
117
+ """Take a send token: return 0, or the seconds to wait before asking again."""
118
+ if not self._enabled:
119
+ return 0.0
120
+ self._refill(self._clock())
121
+ if self._capacity >= 1:
122
+ self._capacity -= 1
123
+ return 0.0
124
+ return (1 - self._capacity) / self._fill_rate
125
+
126
+ def update(self, is_throttling: bool) -> None:
127
+ """Adjust the send rate after an attempt: cut it on throttling, regrow it otherwise."""
128
+ now = self._clock()
129
+ self._measure(now)
130
+ self._adjust(is_throttling, now)
131
+
132
+ def throttled_mid_stream(self) -> None:
133
+ """Cut the send rate for throttling reported inside an open event stream. No new request was sent, so the measured rate is left alone."""
134
+ self._adjust(True, self._clock())
135
+
136
+ def _adjust(self, is_throttling: bool, now: float) -> None:
137
+ if is_throttling:
138
+ rate = (
139
+ min(self._measured_rate, self._fill_rate)
140
+ if self._enabled
141
+ else self._measured_rate
142
+ )
143
+ self._last_max_rate = rate
144
+ self._last_throttle = now
145
+ new_rate = rate * self.BETA
146
+ self._enabled = True
147
+ else:
148
+ window = (self._last_max_rate * (1 - self.BETA) / self.SCALE_CONSTANT) ** (
149
+ 1 / 3
150
+ )
151
+ new_rate = (
152
+ self.SCALE_CONSTANT * (now - self._last_throttle - window) ** 3
153
+ + self._last_max_rate
154
+ )
155
+ self._set_rate(min(new_rate, self.MAX_RATE_SCALE * self._measured_rate), now)
156
+ if is_throttling:
157
+ logger.debug(
158
+ "throttled, send rate lowered to %.2f requests/s",
159
+ self._fill_rate,
160
+ extra={"send_rate": self._fill_rate},
161
+ )
162
+
163
+ def _refill(self, now: float) -> None:
164
+ self._capacity = min(
165
+ self._max_capacity,
166
+ self._capacity + (now - self._last_refill) * self._fill_rate,
167
+ )
168
+ self._last_refill = now
169
+
170
+ def _measure(self, now: float) -> None:
171
+ # Requests per second, counted in half-second buckets and smoothed.
172
+ bucket = math.floor(now * 2) / 2
173
+ self._request_count += 1
174
+ if bucket > self._rate_bucket:
175
+ current = self._request_count / (bucket - self._rate_bucket)
176
+ self._measured_rate = current * self.SMOOTHING + self._measured_rate * (
177
+ 1 - self.SMOOTHING
178
+ )
179
+ self._request_count = 0
180
+ self._rate_bucket = bucket
181
+
182
+ def _set_rate(self, rate: float, now: float) -> None:
183
+ self._refill(now)
184
+ self._fill_rate = max(rate, self.MIN_FILL_RATE)
185
+ self._max_capacity = max(rate, self.MIN_CAPACITY)
186
+ self._capacity = min(self._capacity, self._max_capacity)
187
+
188
+
189
+ class ThreadSafeRateLimiter(RateLimiter):
190
+ """A ``RateLimiter`` that sync clients can share across threads."""
191
+
192
+ def __init__(self, clock: Callable[[], float] = time.monotonic) -> None:
193
+ super().__init__(clock)
194
+ self._lock: threading.Lock = threading.Lock()
195
+
196
+ def acquire(self) -> float:
197
+ with self._lock:
198
+ return super().acquire()
199
+
200
+ def update(self, is_throttling: bool) -> None:
201
+ with self._lock:
202
+ super().update(is_throttling)
203
+
204
+ def throttled_mid_stream(self) -> None:
205
+ with self._lock:
206
+ super().throttled_mid_stream()
207
+
208
+
91
209
  @dataclass
92
210
  class OperationOptions:
93
211
  client: Client
94
212
  retry_quota: RetryQuota
213
+ rate_limiter: RateLimiter
95
214
  use_dual_stack: bool | None = None
96
215
  use_fips: bool | None = None
97
216
  endpoint: str | None = None
98
217
  region: str | None = None
99
218
  retry_max_attempts: int | None = None
219
+ retry_mode: str | None = None
100
220
  credentials_provider: IdentityProvider[Credentials] | None = None
101
221
  anonymous: bool | None = None
102
222
 
@@ -105,11 +225,13 @@ class OperationOptions:
105
225
  class AsyncOperationOptions:
106
226
  client: AsyncClient
107
227
  retry_quota: RetryQuota
228
+ rate_limiter: RateLimiter
108
229
  use_dual_stack: bool | None = None
109
230
  use_fips: bool | None = None
110
231
  endpoint: str | None = None
111
232
  region: str | None = None
112
233
  retry_max_attempts: int | None = None
234
+ retry_mode: str | None = None
113
235
  credentials_provider: IdentityProvider[Credentials] | None = None
114
236
  anonymous: bool | None = None
115
237
 
@@ -428,6 +550,91 @@ def aretry() -> AsyncInterceptor[TInput, TOutput]:
428
550
  return interceptor
429
551
 
430
552
 
553
+ def rate_limit() -> Interceptor[TInput, TOutput]:
554
+ def interceptor(
555
+ request: OperationRequest[TInput], next: NextFn[TInput, TOutput]
556
+ ) -> OperationResponse[TOutput]:
557
+ if request.options.retry_mode != "adaptive":
558
+ return next(request)
559
+ limiter = request.options.rate_limiter
560
+ while (wait := limiter.acquire()) > 0:
561
+ time.sleep(wait)
562
+ try:
563
+ response = next(request)
564
+ except Exception as exc:
565
+ limiter.update(isinstance(exc, ServiceError) and exc.is_throttling_error)
566
+ raise
567
+ limiter.update(False)
568
+ return response
569
+
570
+ return interceptor
571
+
572
+
573
+ def arate_limit() -> AsyncInterceptor[TInput, TOutput]:
574
+ async def interceptor(
575
+ request: AsyncOperationRequest[TInput], next: AsyncNextFn[TInput, TOutput]
576
+ ) -> AsyncOperationResponse[TOutput]:
577
+ if request.options.retry_mode != "adaptive":
578
+ return await next(request)
579
+ limiter = request.options.rate_limiter
580
+ while (wait := limiter.acquire()) > 0:
581
+ await anysleep(wait)
582
+ try:
583
+ response = await next(request)
584
+ except Exception as exc:
585
+ limiter.update(isinstance(exc, ServiceError) and exc.is_throttling_error)
586
+ raise
587
+ limiter.update(False)
588
+ return response
589
+
590
+ return interceptor
591
+
592
+
593
+ def _on_event_stream_error(
594
+ exc: ServiceError, options: OperationOptions | AsyncOperationOptions
595
+ ) -> None:
596
+ # An error event arrives after a 200 response, so raise_error never saw it:
597
+ # mark it by its code here. It is not retried, but adaptive mode slows down.
598
+ _apply_retry_rules(exc, exc.response)
599
+ if exc.is_throttling_error and options.retry_mode == "adaptive":
600
+ options.rate_limiter.throttled_mid_stream()
601
+
602
+
603
+ def watch_event_stream(output: Any, member: str, options: OperationOptions) -> None:
604
+ """Wrap the event stream ``output[member]`` so the errors it raises get their retry flags and reach the rate limiter."""
605
+ if member not in output:
606
+ return
607
+ events = output[member]
608
+
609
+ def watched() -> Iterator[Any]:
610
+ try:
611
+ yield from events
612
+ except ServiceError as exc:
613
+ _on_event_stream_error(exc, options)
614
+ raise
615
+
616
+ output[member] = watched()
617
+
618
+
619
+ def awatch_event_stream(
620
+ output: Any, member: str, options: AsyncOperationOptions
621
+ ) -> None:
622
+ """Wrap the event stream ``output[member]`` so the errors it raises get their retry flags and reach the rate limiter."""
623
+ if member not in output:
624
+ return
625
+ events = output[member]
626
+
627
+ async def watched() -> AsyncIterator[Any]:
628
+ try:
629
+ async for event in events:
630
+ yield event
631
+ except ServiceError as exc:
632
+ _on_event_stream_error(exc, options)
633
+ raise
634
+
635
+ output[member] = watched()
636
+
637
+
431
638
  def log_operation(operation_logging: OperationLogging) -> Interceptor[TInput, TOutput]:
432
639
  operation = operation_logging.operation
433
640
  redact_input = operation_logging.redact_input
@@ -3,7 +3,7 @@
3
3
  import warnings
4
4
  from collections.abc import AsyncGenerator, AsyncIterator
5
5
  from contextlib import asynccontextmanager
6
- from typing import TYPE_CHECKING, Any, Iterable, Optional
6
+ from typing import TYPE_CHECKING, Any, Iterable, Literal, Optional
7
7
 
8
8
  from typing_extensions import Self, TypedDict
9
9
  from zapros import AsyncBaseHandler, AsyncClient
@@ -28,10 +28,13 @@ from capo_sagemaker_runtime_http2._services._pipeline import (
28
28
  AsyncOperationRequest,
29
29
  AsyncOperationResponse,
30
30
  OperationLogging,
31
+ RateLimiter,
31
32
  RetryQuota,
32
33
  aexecute_pipeline,
33
34
  alog_operation,
35
+ arate_limit,
34
36
  aretry,
37
+ awatch_event_stream,
35
38
  )
36
39
 
37
40
  if TYPE_CHECKING:
@@ -43,6 +46,7 @@ if TYPE_CHECKING:
43
46
  class AsyncSageMakerRuntimeHTTP2ClientConfig(TypedDict, total=False, closed=True):
44
47
  operation_interceptors: Iterable[AsyncInterceptor[Any, Any]]
45
48
  retry_max_attempts: int | None
49
+ retry_mode: Literal["standard", "adaptive"] | None
46
50
  use_dual_stack: bool | None
47
51
  use_fips: bool | None
48
52
  endpoint: str | None
@@ -58,6 +62,7 @@ class AsyncSageMakerRuntimeHTTP2Client:
58
62
  http_handler: HTTP handler for sending requests. If not provided, creates a default handler.
59
63
  operation_interceptors: Interceptors that wrap every operation call. If not provided, defaults to an empty list.
60
64
  retry_max_attempts: Maximum number of times to retry a failed operation. Defaults to 3.
65
+ retry_mode: ``"standard"`` (the default) or ``"adaptive"``, which also slows the whole client down when the service throttles it.
61
66
  use_dual_stack: The value of the ``AWS::UseDualStack`` endpoint parameter.
62
67
  use_fips: The value of the ``AWS::UseFIPS`` endpoint parameter.
63
68
  endpoint: The value of the ``SDK::Endpoint`` endpoint parameter.
@@ -72,6 +77,7 @@ class AsyncSageMakerRuntimeHTTP2Client:
72
77
  http_handler: AsyncBaseHandler | None = None,
73
78
  operation_interceptors: Iterable[AsyncInterceptor[Any, Any]] | None = None,
74
79
  retry_max_attempts: int | None = None,
80
+ retry_mode: Literal["standard", "adaptive"] | None = None,
75
81
  use_dual_stack: bool | None = None,
76
82
  use_fips: bool | None = None,
77
83
  endpoint: str | None = None,
@@ -88,6 +94,8 @@ class AsyncSageMakerRuntimeHTTP2Client:
88
94
  self._client = client.wrap_with_middleware(lambda next: AuthMiddleware(next))
89
95
  # One retry quota for all of this client's operations.
90
96
  self._retry_quota = RetryQuota()
97
+ # One send-rate limiter too, used when retry_mode is adaptive.
98
+ self._rate_limiter = RateLimiter()
91
99
  if credentials is not None and credentials_provider is not None:
92
100
  warnings.warn(
93
101
  "Both credentials and credentials_provider given; provider takes precedence"
@@ -105,6 +113,7 @@ class AsyncSageMakerRuntimeHTTP2Client:
105
113
  {
106
114
  "operation_interceptors": operation_interceptors or [],
107
115
  "retry_max_attempts": retry_max_attempts,
116
+ "retry_mode": retry_mode,
108
117
  "use_dual_stack": use_dual_stack,
109
118
  "use_fips": use_fips,
110
119
  "endpoint": endpoint,
@@ -129,12 +138,15 @@ class AsyncSageMakerRuntimeHTTP2Client:
129
138
  if operation_logging is not None:
130
139
  interceptors_.append(alog_operation(operation_logging))
131
140
  interceptors_.append(aretry())
141
+ interceptors_.append(arate_limit())
132
142
  options_: AsyncOperationOptions = AsyncOperationOptions(
133
143
  client=self._client,
134
144
  retry_quota=self._retry_quota,
145
+ rate_limiter=self._rate_limiter,
135
146
  retry_max_attempts=overrides.get(
136
147
  "retry_max_attempts", self._config.get("retry_max_attempts")
137
148
  ),
149
+ retry_mode=overrides.get("retry_mode", self._config.get("retry_mode")),
138
150
  use_dual_stack=overrides.get(
139
151
  "use_dual_stack", self._config.get("use_dual_stack")
140
152
  ),
@@ -191,6 +203,7 @@ class AsyncSageMakerRuntimeHTTP2Client:
191
203
  ) = await capo_sagemaker_runtime_http2._operations.amazon_sage_maker_runtime_http2.invoke_endpoint_with_bidirectional_stream.async_invoke_endpoint_with_bidirectional_stream(
192
204
  req.options, req.input
193
205
  )
206
+ awatch_event_stream(output, "body", req.options)
194
207
  return AsyncOperationResponse(output=output, response=http_response)
195
208
 
196
209
  def _redact_input(
@@ -3,7 +3,7 @@
3
3
  import warnings
4
4
  from collections.abc import Generator, Iterator
5
5
  from contextlib import contextmanager
6
- from typing import TYPE_CHECKING, Any, Iterable, Optional
6
+ from typing import TYPE_CHECKING, Any, Iterable, Literal, Optional
7
7
 
8
8
  from typing_extensions import Self, TypedDict
9
9
  from zapros import BaseHandler, Client
@@ -28,10 +28,13 @@ from capo_sagemaker_runtime_http2._services._pipeline import (
28
28
  OperationOptions,
29
29
  OperationRequest,
30
30
  OperationResponse,
31
+ ThreadSafeRateLimiter,
31
32
  ThreadSafeRetryQuota,
32
33
  execute_pipeline,
33
34
  log_operation,
35
+ rate_limit,
34
36
  retry,
37
+ watch_event_stream,
35
38
  )
36
39
 
37
40
  if TYPE_CHECKING:
@@ -43,6 +46,7 @@ if TYPE_CHECKING:
43
46
  class SageMakerRuntimeHTTP2ClientConfig(TypedDict, total=False, closed=True):
44
47
  operation_interceptors: Iterable[Interceptor[Any, Any]]
45
48
  retry_max_attempts: int | None
49
+ retry_mode: Literal["standard", "adaptive"] | None
46
50
  use_dual_stack: bool | None
47
51
  use_fips: bool | None
48
52
  endpoint: str | None
@@ -58,6 +62,7 @@ class SageMakerRuntimeHTTP2Client:
58
62
  http_handler: HTTP handler for sending requests. If not provided, creates a default handler.
59
63
  operation_interceptors: Interceptors that wrap every operation call. If not provided, defaults to an empty list.
60
64
  retry_max_attempts: Maximum number of times to retry a failed operation. Defaults to 3.
65
+ retry_mode: ``"standard"`` (the default) or ``"adaptive"``, which also slows the whole client down when the service throttles it.
61
66
  use_dual_stack: The value of the ``AWS::UseDualStack`` endpoint parameter.
62
67
  use_fips: The value of the ``AWS::UseFIPS`` endpoint parameter.
63
68
  endpoint: The value of the ``SDK::Endpoint`` endpoint parameter.
@@ -72,6 +77,7 @@ class SageMakerRuntimeHTTP2Client:
72
77
  http_handler: BaseHandler | None = None,
73
78
  operation_interceptors: Iterable[Interceptor[Any, Any]] | None = None,
74
79
  retry_max_attempts: int | None = None,
80
+ retry_mode: Literal["standard", "adaptive"] | None = None,
75
81
  use_dual_stack: bool | None = None,
76
82
  use_fips: bool | None = None,
77
83
  endpoint: str | None = None,
@@ -88,6 +94,8 @@ class SageMakerRuntimeHTTP2Client:
88
94
  self._client = client.wrap_with_middleware(lambda next: AuthMiddleware(next))
89
95
  # One retry quota for all of this client's operations.
90
96
  self._retry_quota = ThreadSafeRetryQuota()
97
+ # One send-rate limiter too, used when retry_mode is adaptive.
98
+ self._rate_limiter = ThreadSafeRateLimiter()
91
99
  if credentials is not None and credentials_provider is not None:
92
100
  warnings.warn(
93
101
  "Both credentials and credentials_provider given; provider takes precedence"
@@ -105,6 +113,7 @@ class SageMakerRuntimeHTTP2Client:
105
113
  {
106
114
  "operation_interceptors": operation_interceptors or [],
107
115
  "retry_max_attempts": retry_max_attempts,
116
+ "retry_mode": retry_mode,
108
117
  "use_dual_stack": use_dual_stack,
109
118
  "use_fips": use_fips,
110
119
  "endpoint": endpoint,
@@ -129,12 +138,15 @@ class SageMakerRuntimeHTTP2Client:
129
138
  if operation_logging is not None:
130
139
  interceptors_.append(log_operation(operation_logging))
131
140
  interceptors_.append(retry())
141
+ interceptors_.append(rate_limit())
132
142
  options_: OperationOptions = OperationOptions(
133
143
  client=self._client,
134
144
  retry_quota=self._retry_quota,
145
+ rate_limiter=self._rate_limiter,
135
146
  retry_max_attempts=overrides.get(
136
147
  "retry_max_attempts", self._config.get("retry_max_attempts")
137
148
  ),
149
+ retry_mode=overrides.get("retry_mode", self._config.get("retry_mode")),
138
150
  use_dual_stack=overrides.get(
139
151
  "use_dual_stack", self._config.get("use_dual_stack")
140
152
  ),
@@ -190,6 +202,7 @@ class SageMakerRuntimeHTTP2Client:
190
202
  req.options, req.input
191
203
  )
192
204
  )
205
+ watch_event_stream(output, "body", req.options)
193
206
  return OperationResponse(output=output, response=http_response)
194
207
 
195
208
  def _redact_input(
@@ -16,13 +16,13 @@ if TYPE_CHECKING:
16
16
  import capo_sagemaker_runtime_http2.types.request_payload_part
17
17
 
18
18
 
19
- class _RequestStreamEvent_PayloadPart(TypedDict, closed=True):
19
+ class RequestStreamEventPayloadPartMember(TypedDict, closed=True):
20
20
  PayloadPart: (
21
21
  "capo_sagemaker_runtime_http2.types.request_payload_part.RequestPayloadPart"
22
22
  )
23
23
 
24
24
 
25
- _RequestStreamEvent: TypeAlias = _RequestStreamEvent_PayloadPart | UnknownEvent
25
+ _RequestStreamEvent: TypeAlias = RequestStreamEventPayloadPartMember | UnknownEvent
26
26
  RequestStreamEvent: TypeAlias = AnyIterator[_RequestStreamEvent]
27
27
 
28
28
 
@@ -18,26 +18,26 @@ if TYPE_CHECKING:
18
18
  import capo_sagemaker_runtime_http2.types.response_payload_part
19
19
 
20
20
 
21
- class _ResponseStreamEvent_PayloadPart(TypedDict, closed=True):
21
+ class ResponseStreamEventPayloadPartMember(TypedDict, closed=True):
22
22
  PayloadPart: (
23
23
  "capo_sagemaker_runtime_http2.types.response_payload_part.ResponsePayloadPart"
24
24
  )
25
25
 
26
26
 
27
- class _ResponseStreamEvent_ModelStreamError(TypedDict, closed=True):
27
+ class ResponseStreamEventModelStreamErrorMember(TypedDict, closed=True):
28
28
  ModelStreamError: (
29
29
  "capo_sagemaker_runtime_http2.errors.model_stream_error.ModelStreamError_"
30
30
  )
31
31
 
32
32
 
33
- class _ResponseStreamEvent_InternalStreamFailure(TypedDict, closed=True):
33
+ class ResponseStreamEventInternalStreamFailureMember(TypedDict, closed=True):
34
34
  InternalStreamFailure: "capo_sagemaker_runtime_http2.errors.internal_stream_failure.InternalStreamFailure_"
35
35
 
36
36
 
37
37
  _ResponseStreamEvent: TypeAlias = (
38
- _ResponseStreamEvent_PayloadPart
39
- | _ResponseStreamEvent_ModelStreamError
40
- | _ResponseStreamEvent_InternalStreamFailure
38
+ ResponseStreamEventPayloadPartMember
39
+ | ResponseStreamEventModelStreamErrorMember
40
+ | ResponseStreamEventInternalStreamFailureMember
41
41
  | UnknownEvent
42
42
  )
43
43
  ResponseStreamEvent: TypeAlias = AnyIterator[_ResponseStreamEvent]
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: capo-sagemaker-runtime-http2
3
- Version: 0.11.0
3
+ Version: 0.13.0
4
4
  Summary: Typed, async Python SDK for AWS Sagemaker Runtime Http2.
5
5
  Project-URL: Homepage, https://capo-sdk.dev/
6
6
  Project-URL: Documentation, https://capo-sdk.dev/services/sagemaker-runtime-http2/
@@ -108,6 +108,8 @@ The SDK retries failed operations automatically. Retry behaviour follows the Smi
108
108
 
109
109
  The number of attempts defaults to 3 and can be changed at the client level via `retry_max_attempts`, or per call via `config_overrides`.
110
110
 
111
+ `retry_mode="adaptive"` (or `AWS_RETRY_MODE=adaptive`, or `retry_mode = adaptive` in the AWS profile) also slows the whole client down once the service starts throttling it, then speeds back up as requests succeed. The default is `standard`.
112
+
111
113
  ```python
112
114
  from capo_sagemaker_runtime_http2 import AsyncSageMakerRuntimeHTTP2Client
113
115