Also, a major compaction doesn't flush the memtable. If the memtable is still full, reads may take slightly longer as they may have to be merged with any on-disk data before being served.
On 10 February 2014 21:18, Tupshin Harper <tups...@tupshin.com> wrote: > You don't mention disks and RAM, but I would assume that the additional > data meant that you could now cache a lower percentage and that you have to > seek on disk more often. > > -Tupshin > On Feb 10, 2014 4:14 PM, "Jiaan Zeng" <l.alle...@gmail.com> wrote: > >> Hi All, >> >> I am using Cassandra 1.2.4. I wonder if update operation has >> *permanent* impacts on read operation. Below is the scenario. >> >> Previously, a read only workload runs against one column family and >> has 4000 qps. Later, a read-update mixed workload runs against the >> same column family. After that, the read only workload runs again but >> it could not get 4000 qps, only get 3500 qps. After a manual major >> compaction is issued through command line, the read only workload gets >> 3600 qps which seems to tell major compaction does not help much. >> >> Did anyone have similar experiences? Any idea why this is happening? >> Thanks. >> >> -- >> Regards, >> Jiaan >> >