GitLab

It looks like our solution to #33666 (closed) has solved the connectivity issues for clients behind restricted NATs, but we still have a relatively small number of unrestricted proxies available for restricted clients.

See the counts of proxy by type:

And a close-up on the counts of unrestricted proxies:

So, we're making progress! But, I'm still seeing a lot of restricted clients getting turned away at the broker because there were no available proxies. Here's a plot of the times a restricted client failed to be matched with a snowflake:

For reference, November (when we start seeing the spikes) was around the time we rolled out our remote probe test and started correctly classifying browser-based proxies. This both increased the amount of unrestricted proxies we had, but also decreased the number of unknown proxies we had. The end result for restricted clients (since they are handed either unknown or unrestricted proxies), is that while each proxy they get is more likely to work for them, the pool of proxies they can pull from is greatly reduced.

The spikes themselves are interesting. Some days we don't get any denied counts and some days we get way more than I'd expect. We almost surely need more unrestricted proxies, but it would also be useful to get some more metrics on what proportion of successful matches are for unrestricted vs restricted clients, and how many times a restricted client has to poll to get a working proxy.

So this ticket is both for measurements, but I also think we should try to run some reliable proxy-go instances that have unrestricted NATs.