Re: Urgent: 10K or more connections

Sean Chittenden <sean@chittenden.org> writes:

> Some light weight multi-threaded proxy that
> relays active connections to the backend and holds idle connections
> more efficiently than PostgreSQL...

What excuse is there for postgres connections being heavyweight to begin with?
The only real resource they ought to represent is a single TCP connection.
Servers that manage 10,000 TCP connections are a dime a dozen these days.

Any database context that has to be stored for the connection, the state of
binary/text or autocommit mode or whatever, will have to be maintained by any
pooling interface anyways. And I think both of those examples are now much
cleaner more or less stateless per-request flags anyways.

Basically what I'm asking is, hypothetically, if postgres were implemented
using threads instead of processes, are there any per-connection resources
that really couldn't be completely disposed of when the connection was
completely idle between (ie at the start of) transactions?

Ideally if every per-connection resource could be completely disposed of
whenever the connection was completely idle then you wouldn't need a whole
extra layer for the communication to traverse and a whole extra layer of
complexity for the protocol semantics to be maintained. A multithreaded server
could easily handle 10k-40k mostly idle connections without any unusual
resource needs.


--
greg