Stream bandwidth, consensus weight and ratio increase in longclaw when enabling uploads
To re-cap: back then when we worked on Sponsor 61 we tried to adapt sbws for Congestion Control (CC). While we did so for the download case it did not give us the desired results and we struggled to get the sbws optimizations working, in particular uploading data instead of downloading measurement data. After a bunch of bug fixes and some debugging efforts we resorted to land a reduced patch but left the upload feature disabled.
Fast forward to now. While working on bandwidth inflation the location-dependence of measurements did bite me quite a bit and I was reviewing what actually kept us from enabling upload to see whether that changes things, in particular that we might be able to debug failures better with DB and Grafana setup. So, we enabled it again with a command line option. This happened on 2024-09-03 around 1700UTC. Now see what happened to longlaw
's measurements since then:
Note longclaw
is still ramping up but we already see a massive improvement in bandwidth measurements. But there is more when looking at the relays' CDF ration. Here is how that look like for the week before 2024-09-02:
and here for the last 7 days:
There are other improvements visible behind those Grafana links, e.g. for MAX(stream_bandwidth) and AVG(stream_bandwidth), too.
Now, I am filing this ticket because for longclaw
we get an error when showing the last 30d on the relays' CDF ratio pane. That does not happen for a period before we deployed the latest changes as can be seen on Grafana as well.
One idea @juga brought up was that maybe we have a bunch of relays doing bandwidth lying which now got caught off-guard due to upload. The second option as to what is happening could be that this is related to relays being always measured in the first position now (EDIT: but see #80 (comment 3076449) for us running that code actually since June already). The third option for this SQL error is that we now use live server descriptors during the measurements not old cached ones being available (EDIT: but see #80 (comment 3076449) for us running that code actually since June already as well).
/cc @mikeperry