Document that setup data is copied per VU on it's first iteration

[mstoykov]

Also, explain a trick to delete the setup data once it has been used in cases that a VU doesn't need all the data.

import {sleep} from "k6";
export function setup () {
    var result ={
        "key1": [],
        "key2": [],
    }
    for (let i = 0; i < 10000; i++ ) {
        result.key1.push({"something": "slse", "i": i})
        result.key2.push({"something": "slse", "i": i})
    }
    return result
}
export let options = {
    stages: [
        {target: 1000, duration: "50s"},
        {target: 1000, duration: "50s"},
    ],
}
var myKeys= [];
export default function(data) {
    if (__ITER ==0) {
        for (let i = 0; i < 5; i++) {
            myKeys.push(data["key1"][Math.floor(Math.random() * data["key1"].length)])
        }
        // console.log(JSON.stringify(myKeys));
        //delete data["key1"];
        //delete data["key2"];
    }
    sleep(1);
}

this will use up 2GB with 150 VUs, but work perfectly fine if the deletes are uncommented as it will free up the memory.

It is important to note that the stages here helps as well as not all VU start at the same time letting the earlier ones having time to get the data they need from setup data and drop the ones it doesn't need.

[na--]

I'd rather we don't document that, or even officially suggest it all that much... Once we document the current implementation details, it will tie our hands in making improvements to them in the future.

W don't copy the setup() data to every VU just because we feel like it, we do it to prevent multiple VUs modifying it and causing a data race panic... But as SharedArray has demonstrated, these days we have better options for sharing immutable data between VUs, so we should probably make the setup() result into something similar in the future... That will use even less memory.

[MattDodsonEnglish]

The setup data being a copy was documented in #638