8.14.Parallel Operations and Bulk Data Transfer with Remote Tables
Especially when bulk copying data from a remote database into Virtuoso, parallel operation is essential for performance.
From version 7.00.3206, automatic query parallelization applies to selected remote table operations.
If read committed isolation is sufficient for remote table access, the remote table access is automatically parallelized if there is an integer key in the remote table that can be used for partitioning.
When the table is attached or when its statistics are updated with sys_stat_analyze, the remote schema is read and if there is a numeric first part of primary key, its minimum and maximum values are read. Samples of rows are read from the table at different ranges between the minimum and maximum and the values are used for filling in column value statistics in sys_col_stat.
If a remote table operation has no condition on the first part of primary key, a range condition is automatically added if the scan of the remote table is the first in a sequence of nested joins. This is the case for example when scanning the table inside a group by, order by, hash join build side or other operation that does not immediately return rows to the client. The number of parallel ranges is given by ThreadsPerQuery in the ini file, accessed as enable_qp with sys_stat and __dbf_set. In a cluster setting, the same DSN is expected to be defined on all processes. The cluster splitting will scan a range per each slice of the elastic cluster.
The lowest range does not specify a low bound and the highest does not specify an upper bound, thus the min and max of the range column do not have to be exactly up to date.
It may be that the primary key of the remote table does not start with a numeric column. If this is the case but there still exists a column usable for splitting a scan, one can create the table in Virtuoso and declare the splitting column as first in the primary key. One can then use vd_remote_table to declare the modified table as remote and use sys_stat_analyze for getting the statistics. After this, parallel scans with the selected column used for setting ranges can be done.
Use the explain function to verify that a scan is actually parallel. One expects to see the text of the remote select followed by "split by range of <column name>" below the text of the remote statement.
In a parallel scan each range is scanned over a different connection. The transaction contexts are independent and the isolation cannot be higher than read committed. Setting a higher isolation level or selecting for update disables the parallel scan feature.
The below snippet copies a table from another Virtuoso.
use R1; ATTACH TABLE lineitem FROM '1208' USER 'dba' PASSWORD 'dba';
This makes a a remote table R1.1208.LINEITEM. This is the lineitem table from the DSN "1208".
Now make a local table:
use DB; CREATE TABLE LINEITEM ( L_ORDERKEY INTEGER NOT NULL, L_PARTKEY INTEGER NOT NULL, L_SUPPKEY INTEGER NOT NULL, L_LINENUMBER INTEGER NOT NULL, L_QUANTITY double precision NOT NULL, L_EXTENDEDPRICE double precision NOT NULL, L_DISCOUNT double precision NOT NULL, L_TAX double precision NOT NULL, L_RETURNFLAG CHAR(1) NOT NULL, L_LINESTATUS CHAR(1) NOT NULL, L_SHIPDATE DATE NOT NULL, L_COMMITDATE DATE NOT NULL, L_RECEIPTDATE DATE NOT NULL, L_SHIPINSTRUCT CHAR(25) NOT NULL, L_SHIPMODE CHAR(10) NOT NULL, L_COMMENT VARCHAR(44) NOT NULL, PRIMARY KEY (L_ORDERKEY, L_LINENUMBER) column );
Note the column modifier after the primary key. This ensures that the table is created in the column store mode. When copying large tables over VDB, this is nearly always done for warehousing, not for OLTP, so the column store mode is essentially always best. Insert is faster and the table takes much less space.
To transfer the data from the remote table into the local one:
log_enable (2); INSERT INTO lineietm SELECT * FROM r1..llineiten; checkpoint;
log_enabel (2) turns off logging an transactions and allows auto committing insert. The explicit checkpoint makes the changes durable. Killing the database server during the transfer would start without any of the effects of the transfer since there is no logging. The data reading and insertion is automatically parallelized.
Most such data transfer operations are in fact network bound since the local insert rate is well in excess of 100MB/s on a commodity server and a 1gbE connection may only transfer around 80MB/s.
The database is normally online during the transfer and the progress may be tracked by periodically counting the target table.