Hi @dave-tucker, could you give an example of a practical use case? A test would be nice to have as well. Thanks!

@ti-mo added some assertions to the tests to exercise this API.
Use case is for doing ring buffer tuning work. For example, you may have some logic in your bpf program that decides whether to use BPF_RB_NO_WAKEUP or BPF_RB_FORCE_WAKEUP in bpf_ringbuf_submit

By periodically calling AvailableData() we can see from userspace how far ahead the producer is from the consumer and whether we're at risk of the ringbuf "wrapping around" and overwriting unread records.

Does #1167 accomplish what you need?

@brycekahle I can try it out and compare both.
The difference between the two approaches is that record.Remaining is calculated once every time a record is read.
AvailableData() will always show the difference between the consumer/producer values at any point in time, so it would handle the case of the producer pointer incrementing multiple times since last read, but we haven't been woken up to read an event yet.

@brycekahle I can try it out and compare both.

I tried it out.

• AvailableData was really easy to wire up to a prometheus.GaugeFunc as it easily be queried once every scrape interval.
• Record.Remaining was less easy. I used a prometheus.Gauge and Set the value within my ringbuf reading function after a record was read.

You can see this on the graphs below - Top is AvailableData(), Bottom is Record.Remaining

The gap in the middle was when I restarted my process to only use BPF_RB_FORCE_WAKEUP when there are 10 entries in the ringbuffer - each entry is 80 bytes.

The top graph is more useful for my purposes of evaluating what ☝️ does to the ring buffer.

I find the bottom graph less useful for my purposes right now, but thinking it through that metric might be better suited to creating a histogram.

TL;DR I think both are useful

Thanks for checking, very informative, especially with the graphs! You're right that putting .Remaining in a gauge isn't going to be very helpful. Did you end up trying with a histogram?

I played around with it and it looked promising but I ended up bailing on it because I couldn't find a satisfactory size for the buckets.

What did you decide to change based on the metrics? Increase buffer size? Wake up less frequently?

All the above. I'm still working on it, but I plan to write up what I've learned soon and I'll share it on Slack when I'm done.

The upside of AvailableData is that it's simple to export and graph. The downside is that we can't extend the concept to perf, since that wraps multiple buffers. So far we've always tried to keep the API of both packages very similar, but that is not a deal breaker.

Left some nits, overall I agree with Florian that this is fine to merge.

Addressed! Thanks!

Thanks for the review and suggestions @ti-mo. I've incorporated them and force-pushed an update.

Pushed a small fixup, will merge when CI is green.