I have been using the Hbase for a while. In the latest project, I have to use HBase for low latency access. I am talking about having average latency of 5 msec and reducing the variability of the long tail latency( mainly 99.99% ) within 250 msec. And having throughtput of 5B requests for 1K response per second. Here are top 10 things to tune in HBase to make it happen.
1. Schema design is starting point to achieve the low latency Salt the key – To avoid the hotspot
2. Pre-split the region
3. Data Locality – Data required to access by Region Server has to local HDFS file.
4. Short Circuit Read – Bypass the data nodes.
5. RPC Handler
6. JVM and GC Setting
7. TCP No delay
8. Enable Bucket Cache
9. MTTR Setting
10. Monitor everything
Big Data Weblog 