Gateway#

Every Flow provides an API Gateway to receive requests over the network. Supported protocols are gRPC, HTTP and WebSocket with TLS.

There are two ways of defining a Gateway, either directly from the Python or using YAML. The full YAML specification of Gateway can be found here.

Supported protocols#

You can use three different protocols to serve the Flow: gRPC, HTTP and Websocket.

from docarray import Document, DocumentArray
from jina import Client, Executor, Flow, requests


class FooExecutor(Executor):
    @requests
    def foo(self, docs: DocumentArray, **kwargs):
        docs.append(Document(text='foo was called'))


f = Flow(protocol='grpc', port=12345).add(uses=FooExecutor)
with f:
    client = Client(port=12345)
    docs = client.post(on='/')
    print(docs.texts)

['foo was called']

from docarray import Document, DocumentArray
from jina import Client, Executor, Flow, requests


class FooExecutor(Executor):
    @requests
    def foo(self, docs: DocumentArray, **kwargs):
        docs.append(Document(text='foo was called'))


f = Flow(protocol='http', port=12345).add(uses=FooExecutor)
with f:
    client = Client(port=12345, protocol='http')
    docs = client.post(on='/')
    print(docs.texts)

['foo was called']

from docarray import Document, DocumentArray
from jina import Client, Executor, Flow, requests


class FooExecutor(Executor):
    @requests
    def foo(self, docs: DocumentArray, **kwargs):
        docs.append(Document(text='foo was called'))


f = Flow(protocol='websocket', port=12345).add(uses=FooExecutor)
with f:
    client = Client(port=12345, protocol='websocket')
    docs = client.post(on='/')
    print(docs.texts)

['foo was called']

To configure the protocol using a YAML file, just do

Note that gRPC is the default protocol, so you can just omit it.

jtype: Flow
with:
  protocol: 'grpc'

jtype: Flow
with:
  protocol: 'http'

jtype: Flow
with:
  protocol: 'websocket'

Customize HTTP interface#

Not every Executor endpoint will automatically be exposed through the external HTTP interface. By default, any Flow exposes the following CRUD and debug HTTP endpoints: /status, /post, /index, /search, /update, and /delete.

Executors that provide additional endpoints (e.g. /foo) will be exposed only after manual configuration. These custom endpoints can be added to the HTTP interface using Flow.expose_endpoint.

../../_images/expose-endpoints.svg
from jina import Executor, requests, Flow


class MyExec(Executor):
    @requests(on='/foo')
    def foo(self, docs, **kwargs):
        pass


f = Flow(protocol='http').add(uses=MyExec)
f.expose_endpoint('/foo', summary='my endpoint')
with f:
    f.block()

You can enable custom endpoints in a Flow using yaml syntax as well.

jtype: Flow
with:
  protocol: http
  expose_endpoints:
    /foo:
      summary: my endpoint

Now, sending an HTTP data request to the /foo endpoint is equivalent to calling f.post('/foo', ...) using the Python Client.

You can add more kwargs to build richer semantics on your HTTP endpoint. Those meta information will be rendered by Swagger UI and be forwarded to the OpenAPI schema.

f.expose_endpoint('/bar', summary='my endpoint', tags=['fine-tuning'], methods=['PUT'])

jtype: Flow
with:
  protocol: http
  expose_endpoints:
    /bar:
      methods: ["PUT"]
      summary: my endpoint
      tags:
        - fine-tuning

Hide default endpoints#

It is possible to hide the default CRUD and debug endpoints in production. This might be useful when the context is not applicable. For example, in the code snippet below, we didn’t implement any CRUD endpoints for the executor, hence it does not make sense to expose them to public.

from jina import Flow

f = Flow(protocol='http', no_debug_endpoints=True, no_crud_endpoints=True)

jtype: Flow
with:
  protocol: 'http'
  no_debug_endpoints: True, 
  no_crud_endpoints: True

After setting up a Flow in this way, the default HTTP endpoints will not be exposed.

Enable cross-origin resource sharing#

To make a Flow accessible from a website with a different domain, you need to enable Cross-Origin Resource Sharing (CORS). Among other things, CORS is necessary to provide a Swagger UI interface for your Flow.

Note that CORS is disabled by default, for security reasons. To enable CORS, configure your Flow in the following way:

from jina import Flow

f = Flow(cors=True, protocol='http')

Advanced options#

HTTP support in Jina is powered by Uvicorn. You can configure the Flow’s internal Uvicorn sever to your heart’s content by passing uvicorn_kwargs to the Flow:

from jina import Flow

f = Flow(protocol='http', uvicorn_kwargs={'loop': 'asyncio', 'http': 'httptools'})

These arguments will be directly passed to the Uvicorn server.

See Also

For more details about the arguments that are used here, and about other available settings for the Uvicorn server, see their website.

Get status#

Tip

Though you can run Executors, Gateway and Client in different Jina versions, it is recommended to work with the same Jina version.

Gateway provides an endpoint that exposes relevant information about the environment where it runs.

This information exposes information in a dict-like structure with the following keys:

  • jina: A dictionary containing information about the system and the versions of several packages including jina package itself

  • envs: A dictionary containing all the values if set of the environment variables used in Jina

Use gRPC#

To see how this works, first instantiate a Flow with an Executor exposed to a specific port and block it for serving:

from jina import Flow

PROTOCOL = 'grpc'  # it could also be http or websocket

with Flow(protocol=PROTOCOL, port=12345).add() as f:
    f.block()

Then, you can use grpcurl sending status check request to the Gateway.

docker pull fullstorydev/grpcurl:latest
docker run --network='host' fullstorydev/grpcurl -plaintext 127.0.0.1:12345 jina.JinaInfoRPC/_status

The error-free output below signifies a correctly running Gateway:

{
  "jina": {
    "architecture": "######",
    "ci-vendor": "######",
    "docarray": "######",
    "grpcio": "######",
    "jina": "######",
    "jina-proto": "######",
    "jina-vcs-tag": "######",
    "platform": "######",
    "platform-release": "######",
    "platform-version": "######",
    "processor": "######",
    "proto-backend": "######",
    "protobuf": "######",
    "python": "######", 
    "pyyaml": "######",
    "session-id": "######",
    "uid": "######",
    "uptime": "######"
  },
  "envs": {
    "JINA_AUTH_TOKEN": "(unset)",
    "JINA_DEFAULT_HOST": "(unset)",
    "JINA_DEFAULT_TIMEOUT_CTRL": "(unset)",
    "JINA_DEPLOYMENT_NAME": "(unset)",
    "JINA_DISABLE_HEALTHCHECK_LOGS": "(unset)",
    "JINA_DISABLE_UVLOOP": "(unset)",
    "JINA_EARLY_STOP": "(unset)",
    "JINA_FULL_CLI": "(unset)",
    "JINA_GATEWAY_IMAGE": "(unset)",
    "JINA_GRPC_RECV_BYTES": "(unset)",
    "JINA_GRPC_SEND_BYTES": "(unset)",
    "JINA_HUBBLE_REGISTRY": "(unset)",
    "JINA_HUB_NO_IMAGE_REBUILD": "(unset)",
    "JINA_LOCKS_ROOT": "(unset)",
    "JINA_LOG_CONFIG": "(unset)",
    "JINA_LOG_LEVEL": "(unset)",
    "JINA_LOG_NO_COLOR": "(unset)",
    "JINA_MP_START_METHOD": "(unset)",
    "JINA_RANDOM_PORT_MAX": "(unset)",
    "JINA_RANDOM_PORT_MIN": "(unset)"
  }
}

Tip

You can also use it to check Executor status, as Executor’s communication protocol is gRPC.

Use HTTP/Websocket#

When using HTTP or Websocket as the Gateway protocol, you can use curl to target the /status endpoint and get the Jina info.

curl http://localhost:12345/status

{"jina":{"jina":"######","docarray":"######","jina-proto":"######","jina-vcs-tag":"(unset)","protobuf":"######","proto-backend":"######","grpcio":"######","pyyaml":"######","python":"######","platform":"######","platform-release":"######","platform-version":"######","architecture":"######","processor":"######","uid":"######","session-id":"######","uptime":"######","ci-vendor":"(unset)"},"envs":{"JINA_AUTH_TOKEN":"(unset)","JINA_DEFAULT_HOST":"(unset)","JINA_DEFAULT_TIMEOUT_CTRL":"(unset)","JINA_DEPLOYMENT_NAME":"(unset)","JINA_DISABLE_UVLOOP":"(unset)","JINA_EARLY_STOP":"(unset)","JINA_FULL_CLI":"(unset)","JINA_GATEWAY_IMAGE":"(unset)","JINA_GRPC_RECV_BYTES":"(unset)","JINA_GRPC_SEND_BYTES":"(unset)","JINA_HUBBLE_REGISTRY":"(unset)","JINA_HUB_NO_IMAGE_REBUILD":"(unset)","JINA_LOG_CONFIG":"(unset)","JINA_LOG_LEVEL":"(unset)","JINA_LOG_NO_COLOR":"(unset)","JINA_MP_START_METHOD":"(unset)","JINA_RANDOM_PORT_MAX":"(unset)","JINA_RANDOM_PORT_MIN":"(unset)","JINA_DISABLE_HEALTHCHECK_LOGS":"(unset)","JINA_LOCKS_ROOT":"(unset)"}}

Add gRPC compression#

The communication between Executors inside a Flow is done via gRPC. To optimize the performance and the bandwidth of these connections, Jina allows the users to specify their compression by specifying compression argument to the Flow.

The supported methods are: none, gzip and deflate.

from jina import Flow

f = Flow(compression='gzip').add(...)

Note that this setting is only effective the internal communication of the Flow. One can also specify the compression between client and gateway as described here.

Enable TLS#

You can enable TLS encryption between your Flow’s Gateway and a Client, for any of the protocols supported by Jina (HTTP, gRPC, and Websocket).

Caution

Enabling TLS will encrypt the data that is transferred between the Flow and the Client. Data that is passed between the microservices configured by the Flow, such as Executors, will not be encrypted.

To enable TLS encryption, you need to pass a valid keyfile and certfile to the Flow, using the ssl_keyfile ssl_certfile parameters:

PORT = ...

f = Flow(
    port=PORT,
    ssl_certfile='path/to/certfile.crt',
    ssl_keyfile='path/to/keyfile.crt',
)

If both of these are provided, the Flow will automatically configure itself to use TLS encryption for its communication with any Client.

Limit outstanding requests#

By default, Jina’s Client sends requests to the Flow as fast as possible without any delay. If a client sends their request faster than the Flow can process them, this can put a high load on the Flow. Typically, this is most likely to happen for Flows with expensive indexing.

You can control the number of in flight requests per Client with the prefetch argument. E.g. setting prefetch=2 lets the API accept only 2 requests per client in parallel, hence limiting the load. By default, prefetch is set to 1000. To disable it you can set it to 0.

def requests_generator():
    while True:
        yield Document(...)

class MyExecutor(Executor):
    @requests
    def foo(self, **kwargs):
        slow_operation()

# Makes sure only 2 requests reach the Executor at a time.
with Flow(prefetch=2).add(uses=MyExecutor) as f:
    f.post(on='/', inputs=requests_generator)

Danger

When working with very slow executors and a big amount of data, you must set prefetch to some small number to prevent out of memory problems. If you are unsure, always set prefetch=1.

from jina import Flow

f = Flow(protocol='http', cors=True)

jtype: Flow
with:
  protocol: 'http'
  cors: True,

Set timeouts#

You can set timeouts for sending requests to the Executors within a Flow by passing the timeout_send parameter. The timeout is specified in milliseconds. By default, it is None and the timeout is disabled.

If you use timeouts, you may also need to set the prefetch option in the Flow. Otherwise, requests may queue up at an Executor and eventually time out.

with Flow(timeout_send=1000) as f:
    f.post(on='/', inputs=[Document()])

The example above limits every request to the Executors in the Flow to a timeout of 1 second.

GraphQL support#

Caution

GraphQL support is an optional feature that requires optional dependencies. To install these, run pip install jina[graphql] or pip install jina[all].

Unfortunately, these dependencies are not available through Conda. You will have to use pip to be able to use GraphQL feature.

A Flow can optionally expose a GraphQL endpoint, located at /graphql. To enable this endpoint, all you need to do is set expose_graphql_endpoint=True on your HTTP Flow:

from jina import Flow

f = Flow(protocol='http', expose_graphql_endpont=True)

jtype: Flow
with:
  protocol: 'http'
  expose_graphql_endpont: True,

See Also

For more details about the Jina GraphQL enpoint, see here.

See further#