The logs from an LDIF import indicated that gathering non-leaf IDs for creating the ancestorid index took an enormous amount of time over 10hrs. The root cause is that the parentid index btree ordering is lexical, but the IDList being built up from it is sorted numerically. In the existing code, the IDList is maintained in constantly sorted order by idl_insert().
Fix Description:
ldbm_get_nonleaf_ids() switches to idl_append_extend() instead idl_insert() for building up the IDList and then sorts the result only once, using qsort with idl_sort_cmp, after the entire list has been gathered.
"The improvement on identical hardware is for the operation to take 10 seconds rather than 10 hours" - Thomas
Bug Description:
The logs from an LDIF import indicated that gathering non-leaf IDs for creating the ancestorid index took an enormous amount of time over 10hrs. The root cause is that the parentid index btree ordering is lexical, but the IDList being built up from it is sorted numerically. In the existing code, the IDList is maintained in constantly sorted order by idl_insert().
Fix Description:
ldbm_get_nonleaf_ids() switches to idl_append_extend() instead idl_insert() for building up the IDList and then sorts the result only once, using qsort with idl_sort_cmp, after the entire list has been gathered.
"The improvement on identical hardware is for the operation to take 10 seconds rather than 10 hours" - Thomas
Patch Author: Thomas Lackey telackey@bozemanpass.com Thanks for the great contribution!!!
relates: https://pagure.io/389-ds-base/issue/49850