Re: REINDEX blocks virtually any queries but some prepared queries.
On Wed, Apr 6, 2022 at 7:49 AM Frédéric Yhuel wrote: > From the documentation > (https://www.postgresql.org/docs/current/sql-reindex.html#id-1.9.3.162.7), > it sounds like REINDEX won't block read queries that don't need the > index. But it seems like the planner wants to take an ACCESS SHARE lock > on every indexes, regardless of the query, and so REINDEX actually > blocks any queries but some prepared queries whose plan have been cached. > > I wonder if it is a bug, or if the documentation should be updated. What > do you think? I've always thought that the docs for REINDEX, while technically accurate, are very misleading in practice. -- Peter Geoghegan
Re: REINDEX blocks virtually any queries but some prepared queries.
On 4/6/22 17:03, Peter Geoghegan wrote: On Wed, Apr 6, 2022 at 7:49 AM Frédéric Yhuel wrote: From the documentation (https://www.postgresql.org/docs/current/sql-reindex.html#id-1.9.3.162.7), it sounds like REINDEX won't block read queries that don't need the index. But it seems like the planner wants to take an ACCESS SHARE lock on every indexes, regardless of the query, and so REINDEX actually blocks any queries but some prepared queries whose plan have been cached. I wonder if it is a bug, or if the documentation should be updated. What do you think? I've always thought that the docs for REINDEX, while technically accurate, are very misleading in practice. Maybe something along this line? (patch attached)From 4930bb8de182b78228d215bce1ab65263baabde7 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Fr=C3=A9d=C3=A9ric=20Yhuel?= Date: Thu, 7 Apr 2022 13:30:59 +0200 Subject: [PATCH] Doc: Elaborate locking considerations for REINDEX --- doc/src/sgml/ref/reindex.sgml | 6 +- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/doc/src/sgml/ref/reindex.sgml b/doc/src/sgml/ref/reindex.sgml index e6b25ee670..06c223d4a3 100644 --- a/doc/src/sgml/ref/reindex.sgml +++ b/doc/src/sgml/ref/reindex.sgml @@ -275,7 +275,11 @@ REINDEX [ ( option [, ...] ) ] { IN considerations are rather different. REINDEX locks out writes but not reads of the index's parent table. It also takes an ACCESS EXCLUSIVE lock on the specific index being processed, - which will block reads that attempt to use that index. In contrast, + which will block reads that attempt to use that index. In particular, + the PostgreSQL query planner wants to take an ACCESS SHARE + lock on every indexes of the table, regardless of the query, and so + REINDEX blocks virtually any queries but some prepared queries + whose plan have been cached and which don't use this very index. In contrast, DROP INDEX momentarily takes an ACCESS EXCLUSIVE lock on the parent table, blocking both writes and reads. The subsequent CREATE INDEX locks out -- 2.30.2
Re: REINDEX blocks virtually any queries but some prepared queries.
On Thu, Apr 07, 2022 at 01:37:57PM +0200, Frédéric Yhuel wrote: > Maybe something along this line? (patch attached) Some language fixes. I didn't verify the behavior, but +1 to document the practical consequences. I guess this is why someone invented REINDEX CONCURRENTLY. > From 4930bb8de182b78228d215bce1ab65263baabde7 Mon Sep 17 00:00:00 2001 > From: =?UTF-8?q?Fr=C3=A9d=C3=A9ric=20Yhuel?= > Date: Thu, 7 Apr 2022 13:30:59 +0200 > Subject: [PATCH] Doc: Elaborate locking considerations for REINDEX > > --- > doc/src/sgml/ref/reindex.sgml | 6 +- > 1 file changed, 5 insertions(+), 1 deletion(-) > > diff --git a/doc/src/sgml/ref/reindex.sgml b/doc/src/sgml/ref/reindex.sgml > index e6b25ee670..06c223d4a3 100644 > --- a/doc/src/sgml/ref/reindex.sgml > +++ b/doc/src/sgml/ref/reindex.sgml > @@ -275,7 +275,11 @@ REINDEX [ ( class="parameter">option [, ...] ) ] { IN > considerations are rather different. REINDEX locks > out writes > but not reads of the index's parent table. It also takes an > ACCESS EXCLUSIVE lock on the specific index being > processed, > - which will block reads that attempt to use that index. In contrast, > + which will block reads that attempt to use that index. In particular, > + the PostgreSQL query planner wants to take an ACCESS > SHARE s/wants/tries/ > + lock on every indexes of the table, regardless of the query, and so every index > + REINDEX blocks virtually any queries but some prepared > queries any query except for > + whose plan have been cached and which don't use this very index. In > contrast, plan has > DROP INDEX momentarily takes an > ACCESS EXCLUSIVE lock on the parent table, blocking > both > writes and reads. The subsequent CREATE INDEX locks > out > -- > 2.30.2 >
Re: REINDEX blocks virtually any queries but some prepared queries.
On 4/7/22 14:40, Justin Pryzby wrote: On Thu, Apr 07, 2022 at 01:37:57PM +0200, Frédéric Yhuel wrote: Maybe something along this line? (patch attached) Some language fixes. Thank you Justin! I applied your fixes in the v2 patch (attached). I didn't verify the behavior, but +1 to document the practical consequences. I guess this is why someone invented REINDEX CONCURRENTLY. Indeed ;) That being said, REINDEX CONCURRENTLY could give you an invalid index, so sometimes you may be tempted to go for a simpler REINDEX, especially if you believe that the SELECTs won't be blocked.From 0b6c7d6e466fabc0233b6960b8a33141d512652f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Fr=C3=A9d=C3=A9ric=20Yhuel?= Date: Thu, 7 Apr 2022 13:30:59 +0200 Subject: [PATCH] Doc: Elaborate locking considerations for REINDEX --- doc/src/sgml/ref/reindex.sgml | 7 ++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/doc/src/sgml/ref/reindex.sgml b/doc/src/sgml/ref/reindex.sgml index e6b25ee670..d3c63c4deb 100644 --- a/doc/src/sgml/ref/reindex.sgml +++ b/doc/src/sgml/ref/reindex.sgml @@ -275,7 +275,12 @@ REINDEX [ ( option [, ...] ) ] { IN considerations are rather different. REINDEX locks out writes but not reads of the index's parent table. It also takes an ACCESS EXCLUSIVE lock on the specific index being processed, - which will block reads that attempt to use that index. In contrast, + which will block reads that attempt to use that index. In particular, + the PostgreSQL query planner tries to take an ACCESS SHARE + lock on every index of the table, regardless of the query, and so + REINDEX blocks virtually any queries except for some prepared + queries whose plan has been cached and which don't use this very index. + In contrast, DROP INDEX momentarily takes an ACCESS EXCLUSIVE lock on the parent table, blocking both writes and reads. The subsequent CREATE INDEX locks out -- 2.30.2
Re: REINDEX blocks virtually any queries but some prepared queries.
Le jeu. 7 avr. 2022 à 15:44, Frédéric Yhuel a écrit : > > > On 4/7/22 14:40, Justin Pryzby wrote: > > On Thu, Apr 07, 2022 at 01:37:57PM +0200, Frédéric Yhuel wrote: > >> Maybe something along this line? (patch attached) > > Some language fixes. > > Thank you Justin! I applied your fixes in the v2 patch (attached). > > v2 patch sounds good. > > I didn't verify the behavior, but +1 to document the practical > consequences. > > I guess this is why someone invented REINDEX CONCURRENTLY. > > > > Indeed ;) That being said, REINDEX CONCURRENTLY could give you an > invalid index, so sometimes you may be tempted to go for a simpler > REINDEX, especially if you believe that the SELECTs won't be blocked. Agreed. -- Guillaume.
Re: REINDEX blocks virtually any queries but some prepared queries.
On Thu, Apr 07, 2022 at 05:29:36PM +0200, Guillaume Lelarge a écrit : > Le jeu. 7 avr. 2022 à 15:44, Frédéric Yhuel a > écrit : >> On 4/7/22 14:40, Justin Pryzby wrote: >> Thank you Justin! I applied your fixes in the v2 patch (attached). > > v2 patch sounds good. The location of the new sentence and its wording seem fine to me. So no objections from me to add what's suggested, as suggested. I'll wait for a couple of days first. >> Indeed ;) That being said, REINDEX CONCURRENTLY could give you an >> invalid index, so sometimes you may be tempted to go for a simpler >> REINDEX, especially if you believe that the SELECTs won't be blocked. > > Agreed. There are many factors to take into account, one is more expensive than the other in terms of resources and has downsides, downsides compensated by the reduction in the lock requirements. There are cases where REINDEX is a must-have, as CONCURRENTLY does not support catalog indexes, and these tend to be easily noticed when corruption spreads around. -- Michael signature.asc Description: PGP signature
Re: REINDEX blocks virtually any queries but some prepared queries.
On 4/8/22 02:22, Michael Paquier wrote: On Thu, Apr 07, 2022 at 05:29:36PM +0200, Guillaume Lelarge a écrit : Le jeu. 7 avr. 2022 à 15:44, Frédéric Yhuel a écrit : On 4/7/22 14:40, Justin Pryzby wrote: Thank you Justin! I applied your fixes in the v2 patch (attached). v2 patch sounds good. The location of the new sentence and its wording seem fine to me. So no objections from me to add what's suggested, as suggested. I'll wait for a couple of days first. Thank you Michael. Indeed ;) That being said, REINDEX CONCURRENTLY could give you an invalid index, so sometimes you may be tempted to go for a simpler REINDEX, especially if you believe that the SELECTs won't be blocked. Agreed. There are many factors to take into account, one is more expensive than the other in terms of resources and has downsides, downsides compensated by the reduction in the lock requirements. There are cases where REINDEX is a must-have, as CONCURRENTLY does not support catalog indexes, and these tend to be easily noticed when corruption spreads around. Indeed! Best regards, Frédéric
Re: REINDEX blocks virtually any queries but some prepared queries.
On Fri, Apr 08, 2022 at 04:23:48PM +0200, Frédéric Yhuel wrote: > Thank you Michael. And done as of 8ac700a. -- Michael signature.asc Description: PGP signature
Re: REINDEX blocks virtually any queries but some prepared queries.
On 4/11/22 02:57, Michael Paquier wrote: On Fri, Apr 08, 2022 at 04:23:48PM +0200, Frédéric Yhuel wrote: Thank you Michael. And done as of 8ac700a. -- Thank you Micheal! For reference purposes, we can see in the code of get_relation_info(), in plancat.c, that indeed every index of the table are opened, and therefore locked, regardless of the query. Best regards, Frédéric