We have been using TrafficServer 5.3.2 for quite some time now, for forward proxying of a number of different HTML5 applications, one of the most important ones being YouTube's TV interface, e.g. https://youtube.com/tv. This is all hosted on CentOS 7.2 x86_64 machines.
We recently upgraded to 6.2.0, and then started having problems with the CONNECT requests for port 443 which are generated by the YouTube app. It seems like these connections are "stalled" somehow, sometimes for >10 seconds. Meanwhile, diags.log is getting spammed lots of the following:
Requests that seem to stall are most likely all of the CONNECT kind, e.g.:
As part of figuring out how to diagnose this, I tried a downgrade to TrafficServer 6.1.1, and this made all the stalling and problems disappear. Afterwards, I did a git bisect on master, from the branch point of 6.1 to the branch point of 6.2, and I ended up at commit af76977:
Author: Susan Hinrichs <firstname.lastname@example.org>
Date: Wed Apr 13 19:57:39 2016 +0000
TS-3612: Restructure client session and transaction processing. This closes #570.
Unfortunately, this is a quite big refactoring commit, so it is not possible to revert it individually to see whether it improves things.
TS-3612 and #570, and I saw there were also a number of follow-up commits to fix various problems with it, but this particular problem of stalled SSL connections is still occurring with master as of today, 2016-09-09.
I realize that this report is still missing reproduction details, since it is tricky to analyze what the YouTube app is doing, and simple curl https:// tests appear to go fast, and don't seem to trigger any stalling. But YouTube itself is pretty easy to try out, I think.