Percona Testing: Failover with no App Errors
These results were the third in a series with Percona. The first sysbench performance testingshowed how ScaleArc increased response time, even without caching enabled, and introduced no measurable latency. The second test highlighted performance gain on WordPress – limitations of the Percona environment limited gains to 3.5x vs. the 20x in a more loaded environment.
This third test measured failover times in various MySQL cluster environments. In the scenarios, the tests calculated the length of time for the failover, which the database needs to complete, and the impact of the application in the interim.
The testing included a graceful failover, in which they manually switched the master and informed ScaleArc via an API call, and a hard stop failover, in which they killed the master and let the replication software and ScaleArc software both discover the failure.
In the graceful scenario, the tests consistently report sub-second failover time and – most importantly – no application errors. These methods are what our customers routinely use to perform zero-downtime maintenance – enabling 100% patching success with no application interruption.
In the hard failure scenario, failover took 4 to 18 seconds, depending on the replication technique – MHA vs. Percona XtraDB, with faster failover on the second system. A handful of queries failed. These were the queries that were in flight past the ScaleArc software but not completed on the database failed in these tests, as they should – for ACID compliance reasons, you wouldn’t want ScaleArc committing writes on behalf of the database.
The majority of the queries did not fail – those still coming into the ScaleArc software while the cluster was promoting a new master. ScaleArc successfully queued those queries and then directed them on to the new master.
ScaleArc’s role in delivering zero downtime for applications is the primary driver for most of customers’ deployments, and it’s exciting to have these third-party results to highlight ScaleArc’s impressive performance.
comments powered by Disqus