This commit is a follow up to cockroachdb#46170. That PR dealt with the table resolution
self-deadlock and additionally ensures that lease acquisition is not blocked.
That commit did not deal with the fact that schema changes which mutate
databases can deadlock themselves on retry because a separate transaction
is used to resolve database descriptors during planning.

This patch makes reading the system.namespace table and reading database
descriptors in high-priority transactions. The transactions will push
locks owned by schema changes out of their way. The idea is that regular
SQL transactions reading schema don't want to wait for the transactions
performing DDL statements. Instead, those DDLs will be pushed and forced
to refresh.

Besides the benefit to regular transactions, this patch also prevents
deadlocks for the transactions performing DDL. Before this patch, the
final select in the following sequence would deadlock:

begin; savepoint s; create database d; rollback to savepoint s;
select * from d.t;

The select is reading the namespace table, using a different txn from
its own. That read would block on the intent laid down by the prior
create. With this patch, the transaction effectively pushes itself, but
gets to otherwise run.

Fixes cockroachdb#46224

Release justification: Fix for an old bug that became more preminent
when we introduced SAVEPOINTs recently.

Release note (bug fix): A rare bug causing transactions that have performed
schema changes to deadlock after they restart has been fixed.

46384: sql: resolve databases in high-priority transactions r=andreimatei a=ajwerner

This commit is a follow up to #46170. That PR dealt with the table resolution
self-deadlock and additionally ensures that lease acquisition is not blocked.
That commit did not deal with the fact that schema changes which mutate
databases can deadlock themselves on retry because a separate transaction
is used to resolve database descriptors during planning.

This patch makes reading the system.namespace table and reading database
descriptors in high-priority transactions. The transactions will push
locks owned by schema changes out of their way. The idea is that regular
SQL transactions reading schema don't want to wait for the transactions
performing DDL statements. Instead, those DDLs will be pushed and forced
to refresh.

Besides the benefit to regular transactions, this patch also prevents
deadlocks for the transactions performing DDL. Before this patch, the
final select in the following sequence would deadlock:

begin; savepoint s; create database d; rollback to savepoint s;
select * from d.t;

The select is reading the namespace table, using a different txn from
its own. That read would block on the intent laid down by the prior
create. With this patch, the transaction effectively pushes itself, but
gets to otherwise run.

Fixes #46224

Release justification: Fix for an old bug that became more preminent
when we introduced SAVEPOINTs recently.

Release note (bug fix): A rare bug causing transactions that have performed
schema changes to deadlock after they restart has been fixed.

Co-authored-by: Andrew Werner <ajwerner@cockroachlabs.com>

This commit is a follow up to cockroachdb#46170. That PR dealt with the table resolution
self-deadlock and additionally ensures that lease acquisition is not blocked.
That commit did not deal with the fact that schema changes which mutate
databases can deadlock themselves on retry because a separate transaction
is used to resolve database descriptors during planning.

This patch makes reading the system.namespace table and reading database
descriptors in high-priority transactions. The transactions will push
locks owned by schema changes out of their way. The idea is that regular
SQL transactions reading schema don't want to wait for the transactions
performing DDL statements. Instead, those DDLs will be pushed and forced
to refresh.

Besides the benefit to regular transactions, this patch also prevents
deadlocks for the transactions performing DDL. Before this patch, the
final select in the following sequence would deadlock:

begin; savepoint s; create database d; rollback to savepoint s;
select * from d.t;

The select is reading the namespace table, using a different txn from
its own. That read would block on the intent laid down by the prior
create. With this patch, the transaction effectively pushes itself, but
gets to otherwise run.

Fixes cockroachdb#46224

Release justification: Fix for an old bug that became more preminent
when we introduced SAVEPOINTs recently.

Release note (bug fix): A rare bug causing transactions that have performed
schema changes to deadlock after they restart has been fixed.

Prior to this change, the StatsRefresher was being notified of a new table
during the execution of the createTable node, before the creating transaction
had committed. Prior to cockroachdb#46170, the StatsRefresher was likely to block on the
intent of the creating transaction. After that change, the StatsRefresher might
not discover the new table because the creating transaction is much more likely
to get pushed further into the future, past the read of the stats refresher.

Release note (bug fix): Fix rare bug where stats were not automatically
generated for a new table.

Prior to this change, the StatsRefresher was being notified of a new table
during the execution of the createTable node, before the creating transaction
had committed. Prior to cockroachdb#46170, the StatsRefresher was likely to block on the
intent of the creating transaction. After that change, the StatsRefresher might
not discover the new table because the creating transaction is much more likely
to get pushed further into the future, past the read of the stats refresher.

Release note (bug fix): Fix rare bug where stats were not automatically
generated for a new table.

47718: sql: move notifying StatsRefresh of new table to post-commit hook r=andreimatei a=ajwerner

Prior to this change, the StatsRefresher was being notified of a new table
during the execution of the createTable node, before the creating transaction
had committed. Prior to #46170, the StatsRefresher was likely to block on the
intent of the creating transaction. After that change, the StatsRefresher might
not discover the new table because the creating transaction is much more likely
to get pushed further into the future, past the read of the stats refresher.

Release note (bug fix): Fix rare bug where stats were not automatically
generated for a new table.

Co-authored-by: Andrew Werner <ajwerner@cockroachlabs.com>

Prior to this change, the StatsRefresher was being notified of a new table
during the execution of the createTable node, before the creating transaction
had committed. Prior to cockroachdb#46170, the StatsRefresher was likely to block on the
intent of the creating transaction. After that change, the StatsRefresher might
not discover the new table because the creating transaction is much more likely
to get pushed further into the future, past the read of the stats refresher.

Release note (bug fix): Fix rare bug where stats were not automatically
generated for a new table.

Prior to this change, the StatsRefresher was being notified of a new table
during the execution of the createTable node, before the creating transaction
had committed. Prior to cockroachdb#46170, the StatsRefresher was likely to block on the
intent of the creating transaction. After that change, the StatsRefresher might
not discover the new table because the creating transaction is much more likely
to get pushed further into the future, past the read of the stats refresher.

Release note (bug fix): Fix rare bug where stats were not automatically
generated for a new table.