Re: Thousands of parallel connections

Chris Travers <chris@metatrontech.com> writes:
> Tom Lane wrote:
>> That does not add up: the graph can't have a negative y-intercept.
>> There should be a substantial cost to run the postmaster at all,
>> and then an essentially fixed cost per connection --- assuming
>> that all the connections are running similar queries, of course.
>> You're telling us the first 40 connections require zero RAM.

> That is strange.  Is it really linear or does the cost go up somewhat
> after the first few?

Well, if you have significant contention problems then the speed could
be worse than linear --- but he was talking about memory usage.  AFAICS,
a backend doing a particular query should need X amount of RAM pretty
much independently of how many others there are.  The only data structure
I can think of that would be impacted at all is QuerySnapshot, and at
4 bytes per sibling backend it's *way* down in the noise...

            regards, tom lane