Query Store and Availability Groups—Force Plan on Secondary Replicas
February 17, 2017 Leave a comment
I’m still fighting with some challenges about inconsistent performance between a primary and secondary replica, so I’ve been waste deep in undocumented system views looking at temporary statistics. One of the things I thought about doing was talking advantage of the Force Plan option in the Query Store in SQL Server 2016. If you are not familiar with this feature, it allows you to force a “preferred” execution plan. In this scenario, our query was running in about 20-30 seconds on the primary, and 20-30 minutes on the secondary. The plans were reasonably close, but I wanted to see what would happen if I forced a plan on the primary.
Primer about the Query Store and Availability Groups
Since readable secondary replicas are read-only, the query store on those secondary replicas are also read-only. This means runtime statistics for queries executed on those replicas are not recorded into the query store. All the stats there are from the primary replica. However, I wasn’t sure what would happen if I forced a plan on the primary—would the secondary replica honor that plan?
Let’s Find Out
The first thing I did was to query the query store catalog views to verify that the plan was forced.
I have to copies of the forced plan. If I run an estimated query plan on the primary, I see that plan is forced. You can see this by looked for UsePlan in the XML of the plan.
I did the same thing on the secondary (in the case of the secondary, we are looking at the actual plan, but it doesn’t matter).
You will note that there is no UsePlan. There are extended events and a catalog view that reflect plan forcing failure (Grant Fritchey wrote about this behavior here), While, I wouldn’t expect the catalog view to get updated, I was hoping that the Extended Event might fire. It did not.
The query store, as awesome as it is, doesn’t really do much for you on readable secondary replica. It does not force plans, nor does it record any of your data.
Thanks to Grant Fritchey and Erin Stellato for helping with this post!