documented that external table names should start with x_

git-svn-id: https://svn.code.sf.net/p/xcat/code/xcat-core/trunk@2640 8638fb3e-16cb-4fca-ae20-7b5d299a9bcd
This commit is contained in:
linggao 2009-01-22 00:55:55 +00:00
parent 8fcbfe7cc8
commit 6a21dc9e46

View File

@ -7,8 +7,9 @@ package xCAT_schema::Sample;
# 1 copy this file to /opt/xcat/lib/perl/xCAT_schema directory, rename it.
# 2 change the word Sample above to be the same as the file name.
# 3 Do NOT change the variable name "%tabspec".
# 4 lljob and llnode are the table names.
# jobid, status, node, jobstatus are the column names. Change them to your like.
# 4 x_lljob and x_llnode are the table names.
# jobid, status, node, jobstatus are the column names.
# Change them to your like. Please make sure all table names start with "x_".
# Each table must have a 'disable' column.
# Please do not use SQL reserved words for your table names and column names.
# Use this site to check the reserved words:
@ -25,7 +26,7 @@ package xCAT_schema::Sample;
#
###############################################################################
%tabspec = (
lljob => {
x_lljob => { #your table name should start with "x_".
cols => [qw(jobid status comments disable)], #do not change 'disable' and 'comments', it is required by xCAT
keys => [qw(jobid)],
required => [qw(jobid)],
@ -40,7 +41,7 @@ package xCAT_schema::Sample;
disable => "Set to 'yes' or '1' to comment out this row.",
},
},
llnode => {
x_llnode => {
cols => [qw(node jobid jobstatus cpu_usage comments disable)],
keys => [qw(node)],
required => [qw(node jobid)],
@ -74,40 +75,43 @@ package xCAT_schema::Sample;
# Note: The xCAT database accessting commands such as
# tabdump, chtab, gettab, nodels, nodeadd, nodech, etc.
# still work without it.
#
# Please make sure that any new object name and attribute name
# should start with "x_".
##################################################################
%defspec = (
job => { attrs => [], attrhash => {}, objkey => 'jobid' }, #create a new object called 'job',
x_job => { attrs => [], attrhash => {}, objkey => 'x_jobid' }, #create a new object called 'x_job',
);
#define the attribtues in the 'job' object using the lljob talbe columns.
@{$defspec{job}->{'attrs'}} =
#define the attribtues in the 'x_job' object using the x_lljob talbe columns.
@{$defspec{x_job}->{'attrs'}} =
(
{ attr_name => 'jobid',
tabentry => 'lljob.jobid',
access_tabentry => 'lljob.jobid=attr:jobid',
{ attr_name => 'x_jobid',
tabentry => 'x_lljob.jobid',
access_tabentry => 'x_lljob.jobid=attr:x_jobid',
},
{ attr_name => 'status',
tabentry => 'lljob.status',
access_tabentry => 'lljob.jobid=attr:jobid',
{ attr_name => 'x_status',
tabentry => 'x_lljob.status',
access_tabentry => 'x_lljob.jobid=attr:x_jobid',
},
);
#object 'node' already defined in /opt/xcat/lib/perl/xCAT/Schema.pm.
#Here we just add jobid and jobstatus attributes to the node object
#Here we just add x_jobid and x_jobstatus attributes to the node object
@{$defspec{node}->{'attrs'}} =
(
{ attr_name => 'jobid',
tabentry => 'llnode.jobid',
access_tabentry => 'llnode.node=attr:node',
{ attr_name => 'x_jobid',
tabentry => 'x_llnode.jobid',
access_tabentry => 'x_llnode.node=attr:node',
},
{ attr_name => 'jobstatus',
tabentry => 'llnode.jobstatus',
access_tabentry => 'llnode.node=attr:node',
{ attr_name => 'x_jobstatus',
tabentry => 'x_llnode.jobstatus',
access_tabentry => 'x_llnode.node=attr:node',
},
{ attr_name => 'cpu',
tabentry => 'llnode.cpu_usage',
access_tabentry => 'llnode.node=attr:node',
{ attr_name => 'x_cpu',
tabentry => 'x_llnode.cpu_usage',
access_tabentry => 'x_llnode.node=attr:node',
},
);
1;