@dalias @ojensen @trysdyn yes, the various specs (there is no definite one lol) allow it. however, almost none of the implementations support it in a meaningful way. some will use the first value, others the last. a good implementation would throw an error or ignore the key, as there is obviously something wrong with the input.

it's a fun thing to abuse when you have two different implementations in a system.

@dahukanna @trysdyn : ☝️ wise.

For anyone who doesn't see the value of this format: this keeps the comment adjacent to the thing you're commenting, even if the structure goes through a decode/encode cycle (provided the keys get alphabetized, which is pretty common). Few things are more frustrating than picking up a line comment and moving it arbitrarily away from the thing it refers to.

@davew @ojensen @trysdyn I prefer multiline comments using arrays, and using //topic as a naming scheme:

```json
{
“//dependencies”: [
“Make sure if you update Next.JS here that”,
“you update the script that handles installing it in”,
“Docker because of dependency weirdness.”
],
“dependencies”: {
}
}
```

@tomrittervg @fluidlogic @andrew_shadura @trysdyn as a hypothetical, how would you adjust the telemetry so an adversary who could read your data couldn’t fingerprint your users with it?

I’d want to simulate it first, but I suspect from my work with OpenTelemetry it’d hurt less than one’d at first suspect to send a subset of signals with each ping? Send the random bitmask used for selection along with the data in case some queries should only consider samples with all metrics of interest present.

That’d augment basic sampling. You’re no doubt familiar but, for those in the back: we can draw operationally useful and statistically defensible conclusions as long as we have enough samples per interval. If N=100K is strong enough, collecting more embiggens our cloud bills without any improvement in our results. Instead, we can throw out 90%… or 99%, or more. Fun!

If you have 100M users, it might suffice to have each user send a ping only on 0.1% of the intervals. Or, to make fingerprinting harder, only 10% of the signals on 1% of the intervals? Maybe some, you simply don’t ever send in combination with others? Might be a paper or two to be had in that.