After updating to ATS 9.1.2 we have detected an increase on self-reported metrics by ATS. On both 8.x and 9.x instances TTFB is gathered with %<{TS_MILESTONE_UA_BEGIN_WRITE-TS_MILESTONE_SM_START}msdms>. As it can be seen in this dashboard:
TTFB as reported by ATS has increased ~25% while ATS reported by the varnish instances that feed traffic to ATS remained basically the same.
We've also detected some differences on reported plugin time, we log total plugin time on both 8.x and 9.x instances with %<{TS_MILESTONE_PLUGIN_TOTAL-TS_MILESTONE_SM_START}msdms>. ATS 8.x instances almost always report a total plugin time of 0ms:
vgutierrez@cp6015:~$ curl -s http://127.0.0.1:3904/metrics |egrep "total_plugin_time_(sum|count)" |grep appservers-rw
trafficserver_backend_total_plugin_time_sum{backend="appservers-rw.discovery.wmnet",prog="atsbackend.mtail"} 0.33
trafficserver_backend_total_plugin_time_count{backend="appservers-rw.discovery.wmnet",prog="atsbackend.mtail"} 8.036859e+06While ATS 9.x instances report >0ms more often (bucket measures time in seconds not ms):
trafficserver_backend_total_plugin_time_sum{backend="appservers-rw.discovery.wmnet",prog="atsbackend.mtail"} 1486.0159999757298
trafficserver_backend_total_plugin_time_count{backend="appservers-rw.discovery.wmnet",prog="atsbackend.mtail"} 3.439146e+06As expected, the described behavior hasn't changed after updating to 9.1.3
Issue has been reported to upstream in: https://github.com/apache/trafficserver/issues/9028