I've been working on a Monitor to SQL Server Availability Groups, and I'd like to share it with you.
Where do we begin? In the end
I guess diving into all the details, problems, and conclusions is not the main goal of everyone. Maybe some of you just want the solution, so here it is, with a Read Me file explaining what to do, a Design (Excel) file explaining every object and relation, an "All Together script" for simple creation of all the objects, and a "Job Creation" file to create all the jobs. Please, before doing anything, read the Read Me File for a correct installation and the right-for-you adjustments.
It is also a great opportunity to credit a great Post that helped me a lot along with developing this solution, written by Derik Hammer.
Now let us dig deeper with some background -
When talking about monitoring the synchronization process of two (or more) replicas of Availability Group, we break down the aspects of what should be tracked:
Syncing Status - Knowing what the syncing status is now and having this information documented historically. Knowing about changes in these statuses and alerting when an AG is not synching. As it is the most critical thing to monitor, short changes (due to short network disconnections) happen frequently and raise alerts. For that, I implemented a 3 samples-based monitor. When a status changes between one sample (1st) to another (2nd), the code will enter an inner loop. A third sample will be taken 10 seconds after the first one, and if it is still different than the first one, only then will it alert. I also filtered out “synchronizing” statuses, a decision taken in my organization.
Syncing Rate - To measure the syncing process flow = its rate. How much data has been copied from one replica to another in a measured constant time delta? I kept it as a "background" monitor without any alert. It is a "nice to have" metric for network/syncing issues investigations.
Lags - This is the most complex thing to measure, starting with what we call a "lag," what is a lag to the organization we work at, what counts as "high lag," and more. All these questions are also explained in the blog I’ve mentioned, but there were several more difficulties I ran into when creating this solution.
Doing the Opposite - Pays Off
The most common and popular way to measure a lag is based on the Redo Queue and Send Queue, and there are great queries ready to use throughout the web. But, when data movement is suspended/instance is down, these metrics return NULL. Microsoft DOCs also bothered to mention that any information regarding Secondary replicas from DMV’s/Performance Counters, read from the Primary replica, is inaccurate, from the base fact that it takes time to update the Primary, about that the Secondary has already committed. Another way would be measuring Send Queue at Primary and Redo Queue at the Secondary, dealing with them separately. I avoided it since my goal was to keep it simple and easy for the next person, and to have one table in one place with everything I need. After testing every script I could find and think of, I decided to base my monitor on Last Commit End Time. It does have edge cases needed to be taken care of, like – database not being used at all for a while, then a transaction happens, and it shows a huge lag because it has been a while since the last commit happened. Another issue was selecting data regarding the Secondary from the Primary when the commit has not been updated yet (even though it actually did commit). It resulted in a fake lag. All these issues (and more), together with the fact you always have a short lag of milliseconds (depending on your network bandwidth), resulted in a “price” of up to 2 seconds standard deviation. Rest assure you are safe if it keeps like that. In my alerting procedure, I check what happens along with the 3 following samples and alert only when there is a potential risk, not for every short lag that is gone a second later.
I’d like to sum up by saying all the “tricks” and logic implemented in this solution were made to match the needs of the organization I worked with. I believe it will serve others well, but you might find things that will be better different for your purposes.
I tried to keep the code as simple and documented as possible, and more technical explanations can be found in the AG Monitoring Design excel sheet. I will be happy to answer any questions as well.
Komentarze