Database Engine Updates 2016-11-10
Version: 1.9.0, 1.9.1
Improved index build – The implementation for building secondary indexes now operates by building the index in a bottom-up fashion, which eliminates unnecessary page splits. This can reduce the time needed to create an index or rebuild a table by up to 75% (based on an
db.r3.8xlargeDB instance class). This feature was in lab mode in Aurora version 1.7 and is enabled by default in Aurora version 1.9 and later. For information, see Aurora Lab Mode.
Lock compression (lab mode) – This implementation significantly reduces the amount of memory that lock manager consumes by up to 66%. Lock manager can acquire more row locks without encountering an out-of-memory exception. This feature is disabled by default and can be activated by enabling Aurora lab mode. For information, see Aurora Lab Mode.
Performance schema – Amazon Aurora now includes support for performance schema with minimal impact on performance. In our testing using SysBench, enabling performance schema could degrade MySQL performance by up to 60%.
SysBench testing of an Aurora DB cluster showed an impact on performance that is 4x less than MySQL. Running the
db.r3.8xlargeDB instance class resulted in 100K SQL writes/sec and over 550K SQL reads/sec, even with performance schema enabled.
Hot row contention improvement – This feature reduces CPU utilization and increases throughput when a small number of hot rows are accessed by a large number of connections. This feature also eliminates
error 188when there is hot row contention.
Improved out-of-memory handling – When non-essential, locking SQL statements are executed and the reserved memory pool is breached, Aurora forces rollback of those SQL statements. This feature frees memory and prevents engine crashes due to out-of-memory exceptions.
Smart read selector – This implementation improves read latency by choosing the optimal storage segment among different segments for every read, resulting in improved read throughput. SysBench testing has shown up to a 27% performance increase for write workloads .
Fixed an issue where an Aurora Replica encounters a shared lock during engine start up.
Fixed a potential crash on an Aurora Replica when the read view pointer in the purge system is NULL.