Re:Running 10M PostgreSQL Indexes in Production
by barrkel on Tuesday June 13, 2017 @06:09AM
(#14543821)
It's Citus Data and CitusDB, not Citrus.
I had a look at Citus for analytics workload (typically using full table scan queries), comparing it with Greenplum and Impala (which uses HDFS and Hive metastore, but not much more from Hadoop). I found Citus to not be very competitive with Greenplum; Greenplum executed table scans quite a bit faster, but more importantly Citus didn't have column stores without adding the cstore extension, and column stores massively increase table scan performance (like, multiple orders of magnitude); and cstore_fdw isn't in the box, you need to add it manually. Greenplum in turn didn't perform as well as Impala as data scaled up; also, Impala could deal with thousands of columns, whereas Greenplum and Citus inherit Postgres limits. Greenplum is also tied together with ssh and scripting duct tape, operationally it's a lot less flexible than HDFS.
Postgres on its own isn't very good at analytics queries, it's far too slow for similar reasons to why Citus without cstore was too slow for us, only you get even less I/O parallelization.
I had a look at Citus for analytics workload (typically using full table scan queries), comparing it with Greenplum and Impala (which uses HDFS and Hive metastore, but not much more from Hadoop). I found Citus to not be very competitive with Greenplum; Greenplum executed table scans quite a bit faster, but more importantly Citus didn't have column stores without adding the cstore extension, and column stores massively increase table scan performance (like, multiple orders of magnitude); and cstore_fdw isn't in the box, you need to add it manually. Greenplum in turn didn't perform as well as Impala as data scaled up; also, Impala could deal with thousands of columns, whereas Greenplum and Citus inherit Postgres limits. Greenplum is also tied together with ssh and scripting duct tape, operationally it's a lot less flexible than HDFS.
Postgres on its own isn't very good at analytics queries, it's far too slow for similar reasons to why Citus without cstore was too slow for us, only you get even less I/O parallelization.