Connection pool with single connections

Question

I would like to get the opinions on situation I'm facing regarding the connection pools.

I'm SW developer, working on some multitenant application. We've one DB and each tenant (client) has it's own schema. For each connected tenant, the solo process is started and it gets a solo DB connection. In the near future, I would need to use it in 300+ simultaneous tenant environment.

From what I read, using a lot of connections (100+) to postgres is not advised. The one solution is to use the connection pool. The other one, to use more DB servers.

I was thinking about the connection pooler (pgBouncer, pgPool). But in current state of our application, it is a bit of problematic. Here is the list of the "problematic" items and proposed solution:

1. Single connection to the server for whole lifetime of the process - This is because, we're using the temp tables and prepared statements heavily. The temp tables duration are variable but in most circumstances, they span multiple transactions.

   • Because connection pooler will return the "free" connection I cannot be sure if the given temp table was created in the returned connection. But I think I could workaround it creating "temp tables" in predefined schema (but then I would need to have some background task to clean the orphaned temp tables from the processes that aren't cleanly closed or crashed - the postgres temp tables are dropped on connection close automatically). For prepared statements I haven't found the workaround.

2. Use of "set search_path="custom_schema",public;" - This is done on application start for each tenant, so the correct tables are used.

   • This could be fixed by issuing the set search_path=... command on each transaction. It should be cheap/fast.

3. Use of triggers that depends on the temp tables - This is used for automatic logging of some stuff. It use the temp tables that are automatically created on the application start.

   • Don't have the solution yet as cannot use the "custom table approach" mentioned above because of the table name should be constant (multiple tenants would then create the same table more than once, which is bad).

So, I don't know if I should start to think about redesigning or not. If I could just add more DB servers and everything would run, than no need to redesign it.

What do you think?

Thank You

Answer 1

这些事情在pgbouncer的pool_mode = session中不是问题。