Containers

Handling requests

In order to receive batches in your Batch API, one of your containers must run a web server which is listening for HTTP requests on the port which is configured in the pod.port field of your API configuration (default: 8080).

Batches will be sent to your web server via HTTP POST requests to the root path (/). The payload will be a JSON-encoded array representing one batch, and the Content-Type header will be set to "application/json". In addition, the job's ID will be passed in via the "X-Cortex-Job-ID" header.

Your web server must respond with status code 200 for the batch to be marked as succeeded (the response body will be ignored).

Once all batches have been processed, one of your workers will receive an HTTP POST request to /on-job-complete. It is not necessary for your web server to handle requests to /on-job-complete (404 errors will be ignored).

Job specification

If you need access to any parameters in the job submission (e.g. config), the entire job specification is available at /cortex/spec/job.json in your API containers' filesystems.

Readiness checks

It is often important to implement a readiness check for your API. By default, as soon as your web server has bound to the port, it will start receiving batches. In some cases, the web server may start listening on the port before its workers are ready to handle traffic (e.g. tiangolo/uvicorn-gunicorn-fastapi behaves this way). Readiness checks ensure that traffic is not sent into your web server before it's ready to handle them.

There are two types of readiness checks which are supported: http_get and tcp_socket (see API configuration for usage instructions). A simple and often effective approach is to add a route to your web server (e.g. /healthz) which responds with status code 200, and configure your readiness probe accordingly:

readiness_probe:
  http_get:
    port: 8080
    path: /healthz

Multiple containers

Your API pod can contain multiple containers, only one of which can be listening for requests on the target port (it can be any of the containers).

The /mnt directory is mounted to each container's filesystem, and is shared across all containers.

Resource requests

Each container in the pod requests its own amount of CPU, memory, GPU, and Inferentia resources. In addition, Cortex's dequeuer sidecar container (which is automatically added to the pod) requests 100m CPU and 100Mi memory.

Observability

See docs for logging, metrics, and alerting.

Using the Cortex CLI or client

It is possible to use the Cortex CLI or client to interact with your cluster's APIs from within your API containers. All containers will have a CLI configuration file present at /cortex/client/cli.yaml, which is configured to connect to the cluster. In addition, the CORTEX_CLI_CONFIG_DIR environment variable is set to /cortex/client by default. Therefore, no additional configuration is required to use the CLI or Python client (which can be instantiated via cortex.client()).

Note: your Cortex CLI or client must match the version of your cluster (available in the CORTEX_VERSION environment variable).

Chaining APIs

It is possible to submit Batch jobs from any Cortex API within a Cortex cluster. Jobs can be submitted to http://ingressgateway-operator.istio-system.svc.cluster.local/batch/<api_name>, where <api_name> is the name of the Batch API you are making a request to.

For example, if there is a Batch API named hello-world running in the cluster, you can make a request to it from a different API in Python by using:

import requests

job_spec = {
    "workers": 1,
    "item_list": {"items": [...], "batch_size": 10},
    "config": {"my_key": "my_value"},
}

response = requests.post(
    "http://ingressgateway-operator.istio-system.svc.cluster.local/batch/hello-world",
    json=job_spec,
)

To make requests from your Batch API to a Realtime, Task, or Async API running within the cluster, see the "Chaining APIs" docs associated with the target workload type.

Last updated