My system has a write intensive workload largely consisting of deletes and inserts. The work is applied by drawing from queues such that I can control the concurrency. The transactions touch about 100 tables with a handful being much larger than the rest by row count (several hundred million rows) . There are foreign keys and indexes present.
I'm finding that my transaction throughput is at best the same as I could do with a RDS multi-AZ cluster that was half the size (4xlarge vs 8xlarge). The bottleneck I identified with Cloudwatch and Performance Insights are from DBLoadNonCPU / io:xactsync wait events. If helpful, it is possible to drop the foreign keys, but we are not able to change the composition of our transactions. Would PostgreSQL table partitions help with this?
AWS UFFICIALEAggiornata 4 mesi fa