Bryan Beaudreault created HBASE-28005:
-----------------------------------------
Summary: Tunable hdfs readahead for STREAM reads
Key: HBASE-28005
URL: https://issues.apache.org/jira/browse/HBASE-28005
Project: HBase
Issue Type: Improvement
Reporter: Bryan Beaudreault
In HBASE-27896 we disabled HDFS readahead for ReadType.PREAD. In our
environment this had a huge impact on disk IO (and network IO if low locality).
The default HDFS readahead is 4mb, which is way larger than many scans might
actually need. In our case, we have clusters serving millions of scans and
those are almost all < 1mb.
I can think of two options, but open to others:
# Set the readahead to whatever the scan.setMaxResultSize is, or some multiple
of block size.
# Simply make it configurable, and let operators configure it on a per table
basis.
It may make sense to do both.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
