--- a/entities/__init__.py Fri Oct 02 17:28:33 2015 +0200
+++ b/entities/__init__.py Thu Mar 06 15:55:33 2014 +0100
@@ -26,6 +26,12 @@
from cubicweb.entity import Entity
+def chunks(seq, step):
+ """See http://stackoverflow.com/questions/312443/how-do-you-split-a-list-into-evenly-sized-chunks-in-python)"""
+ return (seq[i:i+step]
+ for i in xrange(0, len(seq), step))
+
+
class AnyEntity(Entity):
"""an entity instance has e_schema automagically set on the class and
instances have access to their issuing cursor
@@ -44,21 +50,25 @@
@classmethod
def cw_fti_index_rql_queries(cls, req):
- """return the list of rql queries to fetch entities to FT-index
+ """return an iterator on rql queries to fetch entities to FT-index
- The default is to fetch all entities at once and to prefetch
- indexable attributes but one could imagine iterating over
+ The default is to fetch entities 1000 per 1000 and to prefetch
+ indexable attributes, but one could imagine iterating over
"smaller" resultsets if the table is very big or returning
a subset of entities that match some business-logic condition.
"""
- restrictions = ['X is %s' % cls.__regid__]
+ restrictions = []
selected = ['X']
for attrschema in sorted(cls.e_schema.indexable_attributes()):
varname = attrschema.type.upper()
restrictions.append('X %s %s' % (attrschema, varname))
selected.append(varname)
- return ['Any %s WHERE %s' % (', '.join(selected),
- ', '.join(restrictions))]
+ rset = req.execute('Any EID WHERE X eid EID, X is %s' % cls.__regid__)
+ for rows in chunks(rset.rows, 1000):
+ q_restrictions = restrictions + [
+ 'X eid IN (%s)' % ', '.join(str(r[0]) for r in rows)]
+ yield 'Any %s WHERE %s' % (', '.join(selected),
+ ', '.join(q_restrictions))
# meta data api ###########################################################