The omegam-ml REST API¶
The REST API provides a direct interface to models and datasets from any connected client. Unlike the Python API, the client does not need access to either MongoDB or RabbitMQ to make use of omegaml, nor does the client need to use the Python language. Use the REST API to interface from any third-party system to omegaml.
Starting the REST API¶
The REST API is started by the following command. It starts a `Flask`_ application that provides endpoints for datasets, models, scripts and jobs.
$ om runtime serve
Access to endpoints can be selectively restricted by specifing rules as regular expressions:
# allow all datasets, models, scripts, jobs
$ om runtime serve
$ om runtime serve ".*/.*/.*"
# allow all models, no datasets, scripts and jobs
$ om runtime serve "model/.*/.*"
# allow only model foo
$ om runtime serve "model/foo/.*"
# allow only model foo, prediction
$ om runtime serve "model/foo/predict"
Rules can be specified in a file, one rule per line:
# rules.txt
model/.*/.*
scripts/.*/.*
$ om runtime serve --rules rules.txt
Rules can be specified in the OMEGA_RESTAPI_FILTER env variable,
with rules separated by ;
$ export OMEGA_RESTAPI_FILTER="model/.*/.*"
$ om runtime serve
API Reference¶
The API reference is accessible online from your omega-ml instance at:
/api/doc/v1- Swagger UI/api/doc/v1/specs/swagger.json- the Swagger specs (JSON)/api/redoc- ReDoc UI, based on Swagger specs
API Semantics¶
The omega-ml REST API resources are all of the form
/api/version/resource-name/resource-key/?param=value.
Resources¶
The valid resource names are:
dataset - access datasets
model - run model tasks like fit, predict
job - runs jobs (notebooks)
script - runs scripts
config - access the user-specific omega-ml configuration Commercial Edition
The resource-key and query parameters are optional. If a resource-key is not provided, a list of existing resources is returned. If a resource-key is provided the API will look up the respective specific resource for this key and return its content.
Note that the dataset and job resources will return dataset and job contents, respectively. The model resource will only provide meta data, but not the actual contents of the model.
Methods and Responses¶
All resources support a set of HTTP GET, PUT, POST or DELETE methods.
successful GET => HTTP 200 OK
successful POST => HTTP 201 created
successful PUT => HTTP 202 accepted
successful DELETE => HTTP 202 accepted
Commercial Edition
error due to bad input parameters => HTTP 400 Bad Request
error due to authentication => HTTP 401 Unauthorized
error due to wrong authorization => HTTP 403 Forbidden
error due to non existing resource => HTTP 404 Not found
error due to not allowed method => HTTP 405 Method not allowed
severe server errors => HTTP 500 Internal Server error
Asynchronous API execution¶
The omega-ml REST API by default is synchronous, however model, script
and notebook tasks can be run asynchronously. To do this, add async=1
as query parameter:
GET /api/v1/model/mymodel/?datax=sample[x]&datay=sample[y]&async=1
=>
HTTP 202/ACCEPTED
Location: /api/v1/task/b552e887-9ced-4e87-a203-0a4183e4f461/result
Next the most recent task status result can be queried at the URI given by Location:
GET /api/v1/task/b552e887-9ced-4e87-a203-0a4183e4f461/result?resource_uri=/api/v1/model/mymodel/
=>
{"response": {"model": "mymodel", "result": ...},
"status": "SUCCESS",
"task_id": "b552e887-9ced-4e87-a203-0a4183e4f461"
}
Working with data (REST API)¶
Listing datasets¶
resp = requests.get('http://host:port/api/v1/dataset/', auth=auth)
resp.json()
=>
{'meta': {'limit': 20,
'next': None,
'offset': 0,
'previous': None,
'total_count': 3},
'objects': [{'data': {'kind': 'pandas.dfrows', 'name': 'sample'},
'dtypes': None,
'index': None,
'name': None,
'orient': None,
'resource_uri': '/api/v1/dataset/sample/'},
{'data': {'kind': 'pandas.dfrows', 'name': 'sample2'},
'dtypes': None,
'index': None,
'name': None,
'orient': None,
'resource_uri': '/api/v1/dataset/sample2/'},
{'data': {'kind': 'pandas.dfrows', 'name': 'sample99'},
'dtypes': None,
'index': None,
'name': None,
'orient': None,
'resource_uri': '/api/v1/dataset/sample99/'}]}
Reading data¶
resp = requests.get('http://host:port/api/v1/dataset/sample', auth=auth)
resp.json()
=>
{'data': {'x': {'0': 0,
'1': 1,
'10': 0,
'11': 1,
'12': 2,
'13': 3,
'14': 4,
'15': 5,
'16': 6,
'17': 7,
'18': 8,
'19': 9,
'2': 2,
'20': 0,
'21': 1,
'22': 2,
'23': 3,
'24': 4,
'25': 5,
'26': 6,
'27': 7,
'28': 8,
'29': 9,
'3': 3,
'4': 4,
'5': 5,
'6': 6,
'7': 7,
'8': 8,
'9': 9}},
'dtypes': {'x': 'int64'},
'index': {'type': 'Int64Index',
'values': [0,
1,
2,
3,
4,
5,
6,
7,
8,
9,
0,
1,
2,
3,
4,
5,
6,
7,
8,
9,
0,
1,
2,
3,
4,
5,
6,
7,
8,
9]},
'name': 'sample',
'orient': 'dict',
'resource_uri': '/api/v1/dataset/None/'}
Note
To get a valid dataframe back do as follows.
import pandas as pd
df = pd.DataFrame.from_dict(resp.json().get('data'))
df.index = index=resp.json().get('index').get('values')
It is important to set the index to restore the correct row order. This
is due to Python’s arbitrary order of keys in the data dict.
Writing data¶
Writing data is equally straight forward. Note this works for both new and existing datasets. By default data is appended to an existing dataset.
data = {'data': {'x': {'0': 0,
'1': 1,
'2': 2,
'3': 3,
'4': 4,
'5': 5,
'6': 6,
'7': 7,
'8': 8,
'9': 9}},
'dtypes': {'x': 'int64'},
'orient': 'dict',
'index': {'type': 'Int64Index', 'values': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]},
'name': 'sample'}
requests.put('http://host:port/api/v1/dataset/sample/', auth=auth,
json=data)
=>
<Response [204]>
To overwrite an existing data set, use append: false
data = {'data': {'x': {'0': 0,
'1': 1,
'2': 2,
'3': 3,
'4': 4,
'5': 5,
'6': 6,
'7': 7,
'8': 8,
'9': 9}},
'dtypes': {'x': 'int64'},
'append': False,
'orient': 'dict',
'index': {'type': 'Int64Index', 'values': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]},
'name': 'sample'}
requests.put('http://localhost:8001/api/v1/dataset/sample/', auth=auth,
json=data)
=>
<Response [204]>
Transform a DataFrame to API format¶
To transform a Pandas DataFrame into the format expected by the API, use the following code snippet.
def pandas_to_apidata(df, append=False):
# TODO put logic for this into client lib
data = {
'append': append,
'data': json.loads(df.to_json()),
'dtypes': {k: str(v)
for k, v in df.dtypes.to_dict().items()},
'orient': 'columns',
'index': {
'type': type(df.index).__name__,
# ensure type conversion to object for Py3 tastypie does
# not recognize numpy.int64
'values': list(df.index.astype('O').values),
}
}
return data
Working with models (REST API)¶
Create a model¶
data = {'name': 'mymodel',
'pipeline': [
# step name, model class, kwargs
['LinearRegression', dict()],
]}
requests.post('http://localhost:8001/api/v1/model/',
json=data,
auth=auth)
=>
<Response [201]>
{'model': {'bucket': 'store',
'created': '2016-01-16 22:05:06.192000',
'kind': 'sklearn.joblib',
'name': 'mymodel'}}
Fit a model¶
Create some data first:
# a simple data frame to learn
df = pd.DataFrame({'x': range(10)})
df['y'] = df['x'] * 2
datax = pandas_to_apidata(df[['x']])
datay = pandas_to_apidata(df[['y']])
# store data
requests.put('http://localhost:8001/api/v1/dataset/datax/', auth=auth,
data=json.dumps(datax))
requests.put('http://localhost:8001/api/v1/dataset/datay/', auth=auth,
json=datay)
=>
<Response [204]>
Then we can fit the model:
resp = requests.put('http://localhost:8001/api/v1/model/mymodel/fit/?datax=datax&datay=datay', auth=auth, data={})
resp.json()
=>
{'datax': 'datax', 'datay': 'datay', 'result': 'ok'}
Subsequently, the model is ready for prediction:
resp = requests.get('http://localhost:8001/api/v1/model/mymodel/predict/?datax=datax', auth=auth, data={})
resp.json()
=>
{'datax': 'datax',
'datay': None,
'result': [0.0, 2.0, 4.0, 6.0, 8.0, 10.0, 12.0, 14.0, 16.0, 18.0]}
Working with jobs (REST API)¶
Listing jobs¶
resp = requests.get('http://localhost:8001/api/v1/job/',
auth=auth)
resp.json()
=>
{u'meta': {u'previous': None, u'total_count': 1,
u'offset': 0, u'limit': 20, u'next': None},
u'objects': [{u'job_runs': [],
u'job_results': {}, u'name': u'testjob.ipynb',
u'created': u'2016-02-06T21:33:49.833000'}]}
Running a job¶
resp = requests.post('http://localhost:8001/api/v1/job/testjob/run/',
auth=auth)
resp.json()
=>
{u'job_runs': {u'1517953074': u'OK'},
u'job_results': [u'results/testjob_1517953074.ipynb'],
u'name': u'testjob.ipynb',
u'created': u'2016-02-06T21:37:54.014000'}
Getting job results¶
To get job results in iPython notebook format, use
resp = requests.get('http://localhost:8001/api/v1/job/results/testjob_1517953074.ipynb/',
auth=auth)
resp.json()
=>
{u'source_job': u'testjob', u'job_results': {},
u'created': u'2016-02-06T21:36:06.704000',
u'content': {u'nbformat_minor': 0, u'nbformat': 4,
u'cells': [{u'execution_count': 1, u'cell_type': u'code',
u'source': u"print('hello')",
u'outputs': [{u'output_type':
u'stream', u'name': u'stdout',
u'text': u'hello\n'}],
u'metadata': {}}],
u'metadata': {}},
u'job_runs': [],
u'name': u'results/testjob_1517952965.ipynb'}
Getting a job report¶
To get job results in HTML format, use
resp = requests.get('http://localhost:8001/api/v1/job/export/testjob_1517953074.ipynb/',
auth=auth)
resp.json()
=>
{u'content': "<html> ... </html>",
u'name': 'testjob_1517953074.ipynb'}
Getting information on a job¶
resp = requests.get('http://localhost:8001/api/v1/job/testjob/',
json=data, auth=auth)
resp.json()
=>
{u'content': {u'nbformat_minor': 0, u'nbformat': 4,
u'cells': [{u'execution_count': None, u'cell_type':
u'code', u'source': u"print('hello')",
u'outputs': [], u'metadata': {}}],
u'metadata': {}}, u'job_runs': [],
u'job_results': {},
u'name': u'testjob.ipynb',
u'created': u'2016-02-06T21:44:59.290000'}