Last modified: 2014-07-02 13:27:48 UTC
Around the time of the 1.19 deployments back in February-time, there was some concern on this list about the downwards spike in the parser cache hit rate, as shown on [1]. There was, however, also evidence that it was returning to "normal" (previous) rates and so the concern dissipated. Contrary to expectation, however, we're now two months on and the new "normal" rates are still significantly below the old "normal" rates (or to put it another way, the miss rate is still considerably higher than it used to be). [1] http://bit.ly/JI41CR
Network traffic to db40 is much lower since March 1 (1.19 to enwiki?) : http://ganglia.wikimedia.org/latest/graph.php?r=year&z=xlarge&c=MySQL+pmtpa&h=db40.pmtpa.wmnet&v=49466765&m=mysql_bytes_sent&jr=&js=&vl=bytes&ti=mysql_bytes_sent And the rate of select queries to it is down nearly 50% vs. when it was in service for 1.18. http://ganglia.wikimedia.org/latest/graph_all_periods.php?c=MySQL%20pmtpa&h=db40.pmtpa.wmnet&v=14811&m=mysql_com_select&r=custom&z=default&jr=&js=&st=1336061638&cs=1%2F9%2F2012%2016%3A13&ce=5%2F3%2F2012%202%3A12&vl=stmts&ti=mysql_com_select&z=large (using inspect to get get daily values from the year graph)
Any further thoughts? Looking at the graph, it's almost as though "MISS" is tracking "EXPIRED" but plus a couple of percent. Not sure if that helps diagnose the problem, if there is one.
jarry1250: Is this still worth to investigate (recent similar incidents), or should this report be closed (e.g. WORKSFORME)?
There was an unexplained step-change in February 2012 which has not been fully explained and which suggests to me that some optimisation remains possible. Thus, my instinct is to leave this bug open until the parser cache miss rate is reduced or the change has been explained (e.g. as an artefact of a change in recording methodology). That said, I'm happy to go with Ops' assessment as to the merits of the point. Alas, they tend to live on RT rather than here.
Probably too late on to debug or gain further insight into this, so closing.
