/
mlflow.py
315 lines (287 loc) · 11.3 KB
/
mlflow.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
from dataclasses import dataclass
from typing import Any, Callable, Iterable, Optional, TypeVar, Union
from mlflow import MlflowClient
from mlflow.entities import Run
from mlflow.entities.model_registry import ModelVersion, RegisteredModel
from mlflow.store.entities import PagedList
from pydantic.fields import Field
import datahub.emitter.mce_builder as builder
from datahub.configuration.source_common import EnvConfigMixin
from datahub.emitter.mcp import MetadataChangeProposalWrapper
from datahub.ingestion.api.common import PipelineContext
from datahub.ingestion.api.decorators import (
SupportStatus,
capability,
config_class,
platform_name,
support_status,
)
from datahub.ingestion.api.source import Source, SourceCapability, SourceReport
from datahub.ingestion.api.workunit import MetadataWorkUnit
from datahub.metadata.schema_classes import (
GlobalTagsClass,
MLHyperParamClass,
MLMetricClass,
MLModelGroupPropertiesClass,
MLModelPropertiesClass,
TagAssociationClass,
TagPropertiesClass,
VersionTagClass,
_Aspect,
)
T = TypeVar("T")
class MLflowConfig(EnvConfigMixin):
tracking_uri: Optional[str] = Field(
default=None,
description="Tracking server URI. If not set, an MLflow default tracking_uri is used (local `mlruns/` directory or `MLFLOW_TRACKING_URI` environment variable)",
)
registry_uri: Optional[str] = Field(
default=None,
description="Registry server URI. If not set, an MLflow default registry_uri is used (value of tracking_uri or `MLFLOW_REGISTRY_URI` environment variable)",
)
model_name_separator: str = Field(
default="_",
description="A string which separates model name from its version (e.g. model_1 or model-1)",
)
@dataclass
class MLflowRegisteredModelStageInfo:
name: str
description: str
color_hex: str
@platform_name("MLflow")
@config_class(MLflowConfig)
@support_status(SupportStatus.TESTING)
@capability(
SourceCapability.DESCRIPTIONS,
"Extract descriptions for MLflow Registered Models and Model Versions",
)
@capability(SourceCapability.TAGS, "Extract tags for MLflow Registered Model Stages")
class MLflowSource(Source):
platform = "mlflow"
registered_model_stages_info = (
MLflowRegisteredModelStageInfo(
name="Production",
description="Production Stage for an ML model in MLflow Model Registry",
color_hex="#308613",
),
MLflowRegisteredModelStageInfo(
name="Staging",
description="Staging Stage for an ML model in MLflow Model Registry",
color_hex="#FACB66",
),
MLflowRegisteredModelStageInfo(
name="Archived",
description="Archived Stage for an ML model in MLflow Model Registry",
color_hex="#5D7283",
),
MLflowRegisteredModelStageInfo(
name="None",
description="None Stage for an ML model in MLflow Model Registry",
color_hex="#F2F4F5",
),
)
def __init__(self, ctx: PipelineContext, config: MLflowConfig):
super().__init__(ctx)
self.config = config
self.report = SourceReport()
self.client = MlflowClient(
tracking_uri=self.config.tracking_uri,
registry_uri=self.config.registry_uri,
)
def get_report(self) -> SourceReport:
return self.report
def get_workunits_internal(self) -> Iterable[MetadataWorkUnit]:
yield from self._get_tags_workunits()
yield from self._get_ml_model_workunits()
def _get_tags_workunits(self) -> Iterable[MetadataWorkUnit]:
"""
Create tags for each Stage in MLflow Model Registry.
"""
for stage_info in self.registered_model_stages_info:
tag_urn = self._make_stage_tag_urn(stage_info.name)
tag_properties = TagPropertiesClass(
name=self._make_stage_tag_name(stage_info.name),
description=stage_info.description,
colorHex=stage_info.color_hex,
)
wu = self._create_workunit(urn=tag_urn, aspect=tag_properties)
yield wu
def _make_stage_tag_urn(self, stage_name: str) -> str:
tag_name = self._make_stage_tag_name(stage_name)
tag_urn = builder.make_tag_urn(tag_name)
return tag_urn
def _make_stage_tag_name(self, stage_name: str) -> str:
return f"{self.platform}_{stage_name.lower()}"
def _create_workunit(self, urn: str, aspect: _Aspect) -> MetadataWorkUnit:
"""
Utility to create an MCP workunit.
"""
return MetadataChangeProposalWrapper(
entityUrn=urn,
aspect=aspect,
).as_workunit()
def _get_ml_model_workunits(self) -> Iterable[MetadataWorkUnit]:
"""
Traverse each Registered Model in Model Registry and generate a corresponding workunit.
"""
registered_models = self._get_mlflow_registered_models()
for registered_model in registered_models:
yield self._get_ml_group_workunit(registered_model)
model_versions = self._get_mlflow_model_versions(registered_model)
for model_version in model_versions:
run = self._get_mlflow_run(model_version)
yield self._get_ml_model_properties_workunit(
registered_model=registered_model,
model_version=model_version,
run=run,
)
yield self._get_global_tags_workunit(model_version=model_version)
def _get_mlflow_registered_models(self) -> Iterable[RegisteredModel]:
"""
Get all Registered Models in MLflow Model Registry.
"""
registered_models: Iterable[
RegisteredModel
] = self._traverse_mlflow_search_func(
search_func=self.client.search_registered_models,
)
return registered_models
@staticmethod
def _traverse_mlflow_search_func(
search_func: Callable[..., PagedList[T]],
**kwargs: Any,
) -> Iterable[T]:
"""
Utility to traverse an MLflow search_* functions which return PagedList.
"""
next_page_token = None
while True:
paged_list = search_func(page_token=next_page_token, **kwargs)
yield from paged_list.to_list()
next_page_token = paged_list.token
if not next_page_token:
return
def _get_ml_group_workunit(
self,
registered_model: RegisteredModel,
) -> MetadataWorkUnit:
"""
Generate an MLModelGroup workunit for an MLflow Registered Model.
"""
ml_model_group_urn = self._make_ml_model_group_urn(registered_model)
ml_model_group_properties = MLModelGroupPropertiesClass(
customProperties=registered_model.tags,
description=registered_model.description,
createdAt=registered_model.creation_timestamp,
)
wu = self._create_workunit(
urn=ml_model_group_urn,
aspect=ml_model_group_properties,
)
return wu
def _make_ml_model_group_urn(self, registered_model: RegisteredModel) -> str:
urn = builder.make_ml_model_group_urn(
platform=self.platform,
group_name=registered_model.name,
env=self.config.env,
)
return urn
def _get_mlflow_model_versions(
self,
registered_model: RegisteredModel,
) -> Iterable[ModelVersion]:
"""
Get all Model Versions for each Registered Model.
"""
filter_string = f"name = '{registered_model.name}'"
model_versions: Iterable[ModelVersion] = self._traverse_mlflow_search_func(
search_func=self.client.search_model_versions,
filter_string=filter_string,
)
return model_versions
def _get_mlflow_run(self, model_version: ModelVersion) -> Union[None, Run]:
"""
Get a Run associated with a Model Version. Some MVs may exist without Run.
"""
if model_version.run_id:
run = self.client.get_run(model_version.run_id)
return run
else:
return None
def _get_ml_model_properties_workunit(
self,
registered_model: RegisteredModel,
model_version: ModelVersion,
run: Union[None, Run],
) -> MetadataWorkUnit:
"""
Generate an MLModel workunit for an MLflow Model Version.
Every Model Version is a DataHub MLModel entity associated with an MLModelGroup corresponding to a Registered Model.
If a model was registered without an associated Run then hyperparams and metrics are not available.
"""
ml_model_group_urn = self._make_ml_model_group_urn(registered_model)
ml_model_urn = self._make_ml_model_urn(model_version)
if run:
hyperparams = [
MLHyperParamClass(name=k, value=str(v))
for k, v in run.data.params.items()
]
training_metrics = [
MLMetricClass(name=k, value=str(v)) for k, v in run.data.metrics.items()
]
else:
hyperparams = None
training_metrics = None
ml_model_properties = MLModelPropertiesClass(
customProperties=model_version.tags,
externalUrl=self._make_external_url(model_version),
description=model_version.description,
date=model_version.creation_timestamp,
version=VersionTagClass(versionTag=str(model_version.version)),
hyperParams=hyperparams,
trainingMetrics=training_metrics,
# mlflow tags are dicts, but datahub tags are lists. currently use only keys from mlflow tags
tags=list(model_version.tags.keys()),
groups=[ml_model_group_urn],
)
wu = self._create_workunit(urn=ml_model_urn, aspect=ml_model_properties)
return wu
def _make_ml_model_urn(self, model_version: ModelVersion) -> str:
urn = builder.make_ml_model_urn(
platform=self.platform,
model_name=f"{model_version.name}{self.config.model_name_separator}{model_version.version}",
env=self.config.env,
)
return urn
def _make_external_url(self, model_version: ModelVersion) -> Union[None, str]:
"""
Generate URL for a Model Version to MLflow UI.
"""
base_uri = self.client.tracking_uri
if base_uri.startswith("http"):
return f"{base_uri.rstrip('/')}/#/models/{model_version.name}/versions/{model_version.version}"
else:
return None
def _get_global_tags_workunit(
self,
model_version: ModelVersion,
) -> MetadataWorkUnit:
"""
Associate a Model Version Stage with a corresponding tag.
"""
global_tags = GlobalTagsClass(
tags=[
TagAssociationClass(
tag=self._make_stage_tag_urn(model_version.current_stage),
),
]
)
wu = self._create_workunit(
urn=self._make_ml_model_urn(model_version),
aspect=global_tags,
)
return wu
@classmethod
def create(cls, config_dict: dict, ctx: PipelineContext) -> Source:
config = MLflowConfig.parse_obj(config_dict)
return cls(ctx, config)