[core-http] node-fetch takes above 2 seconds to send the request on some environments

[sadasant]

While we stopped using node-fetch on core-rest-pipeline, core-http does use it. Node fetch takes above 2 seconds to begin sending network requests in some environments. We’re able to reproduce this issue in Kubernetes pods, but it could be present elsewhere.

It is possible to test this by running the following code on a Kubernetes pod (asume the latest versions of the dependencies):

const { ManagedIdentityCredential } = require("@azure/identity");
const nock = require("nock");

const credential = new ManagedIdentityCredential();

async function main() {
    nock.recorder.rec();
    token = await credential.getToken("https://vault.azure.net/.default", { requestOptions: { timeout: 2000 } });
    console.log(token);
}

main().catch((err) => {
    console.log("error code: ", err.code);
    console.log("error message: ", err.message);
    console.log("error stack: ", err.stack);
});

The code above won’t reach to the IMDS MSI endpoint before two seconds, causing the pipeline to abort the request.

nock here is used to log the responses. We have another issue which makes it hard to debug network requests: #15936

For the case of Identity, we should be able to solve this as soon as we move Identity to core-rest-pipeline: #15595 We’re planning to do this during July. However, that won’t solve the issue in core-http.

[ramya-rao-a]

#15937 fixes this issue in the context of requests made as part of the ManagedIdentityCredential where we previously had a very low default timeout of 500ms

In the next few months we will be moving all packages that use @azure/core-http to the newer @azure/core-rest-pipeline and lose the dependency on node-fetch, see #15594

Therefore, no plans to investigate the current issue here in core-http

cc @chradek @xirzec @jeremymeng