On 1/8/16, Simon Riggs <simon@2ndquadrant.com> wrote: > On 8 January 2016 at 12:49, Vitaly Burovoy <vitaly.burovoy@gmail.com> > wrote: > > >> In Postgres9.1 a new feature was implemented [1] for adding PK and >> UNIQUE constraints using indexes created concurrently, but constraints >> NOT NULL and CHECK still require full seqscan of a table. New CHECK >> constraint allows "NOT VALID" option but VALIDATE CONSTRAINT still >> does seqscan (with RowExclusiveLock, but for big and constantly >> updatable table it is still awful). >> >> It is possible to find wrong rows in a table without seqscan if there >> is an index with a predicate allows to find such rows. There is no >> sense what columns it has since it is enough to check whether >> index_getnext for it returns NULL (table is OK) or any tuple (table >> has wrong rows). >> > > You avoid a full seqscan by creating an index which also does a full seq > scan. > > How does this help? The lock and scan times are the same. I avoid not a full seqscan, but a time when table is under ExclusiveLock: index can be build concurrently without locking table.
That is exactly what ADD ...NOT VALID and VALIDATE already does, as of 9.4.
--
Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services