[YSQL] TEMP TABLE changes system catalog on master

[d-uspenskiy]

Jira Link: DB-1716 Due to the fact that system catalog is changed on master in case of temp table creation/deletion the following scenario is possible:

• each time pgsql table is deleted YsqlCatalogVersion gets updated several times (10 or 20 times)
• the same is valid for temp table now
• when user closes the connection with lots of temp tables system starts to drop all temp tables at same time. This generates storm of YsqlCatalogVersion updates
• at this time all read requests of system catalog will be failed with Catalog Version Mismatch error

[d-uspenskiy]

Some other symptoms also described in #1469

[jaki]

Creating more than one temporary table in a Postgres regress test seems to cause the next regress test in schedule to fail with a read RPC timeout. I suspect that this issue captures the reason. This seems to be an issue on CentOS but a non-issue on Mac. Here is a minimal repro:

• repro_schedule

  test: first
  test: second

• sql/first.sql

  CREATE TEMP TABLE a (i int);
  CREATE TEMP TABLE b (i int);

• sql/second.sql

  SELECT 1;

  results/first.out would be identical to expected/first.out. However, results/second.out would look like

  ysqlsh: FATAL:  Timed out: Read RPC to **** timed out after 9.718s

  which is not expected.

If the schedule had any more tests before and after, they wouldn't be affected. Only second is affected because it is right after first.

[JDNdeveloper]

Just wanted to add that it seems like we might have this issue even if the test is last in the schedule. The cleanup code executes some SQL statements which seem to hit the Read RPC timed out error as well.