Thank you. I was thinking or Redis ordered sets (ordered by time of signing up) but was waiting whether someone will bring it up instead of (D)ETS and Mnesia - they do seem to offer similar features, but it feels like a few bits and pieces need to be added manually, while Redis is more up to date with today’s needs and has it all out of the box… So I would just ZADD everyone into ordered list, then save the first 30 in bulk to Postgres, but keep the list as a “waiting list” around, maybe just using Redis persistence options…

Then I was thinking also about not scaling prematurely because I believe Postgres is in fact much more powerful than most people realize However, I am worried mostly about the atomicity - INSERT with RETURNING the “rank” in the table, which might be not so performant anymore…?

This is a university system and there’s gonna be dozens, maybe hundreds, of courses - I was thinking each having its own GenServer taking care of the queue. And they open the floodgates at the same moment for all the students, which is 40k people – so it’s huge sudden spike in traffic, but it’s also true that the atomic writes, which I’m afraid of the most, would likely be spread at least over several seconds, so probably not reach the rate of thousands per second.

I’m just trying to prevent the repeating scenario that the system always goes down within seconds of opening and then they’re putting out fires for the whole day while people at home are crazily hitting refresh button all day, desperately needing to enroll before it’s all full

How many instances of web server are you going to have? Is it going to sustain the 10000 requests in a second?

If one server instance is fine, then a single GenServer per course sounds like a good solution. Looks like your constraint is just a counter (30 seats). Assuming you have a reasonable DB (INSERT is just a few milliseconds), you can just write the record immediately and have all the consistency guarantees. So the GenServer would be something like:

1. Start the GenServer with course id,
2. On init (or handle_continue) get the booked seats,
3. On each “book a seat” message, check the counter, write to DB if seat available, return result.

Be sure to bump the timeout on the GenServer.call when sending the “book a seat” message. I think this should be good enough, but it’s hard to tell without real performance tests.

If it turns out that you need multiple instances of the web server, then things get a bit more hairy. You could go with Erlang distribution or use out-of-process cache like Redis.

Yep, I mostly agree with what @stefanchrobot said.

You could theoretically go with a singleton GenServer (a single global process), but that will make horizontally scaling your app to several instances and ensuring you don’t loose data in case of restarts a bit more difficult.

Honestly, I would rather:

• Implement it efficiently with standard database queries. You can go a long way with a database like Postgres before you hit such a bottleneck. How likely it is that you would actually get 10.000 requests per second? Remember you can always optimize your solution when traffic actually grows to the point you start to see scalability problems.

• If you need more performance, I would implement this with Redis rather than in a GenServer. You could use a simple integer counter, or even a set, so you can efficiently check if the course is full and only write to database if not (in cases when the sets are much bigger and one is fine with an approximated count Redis even provides HyperLogLog). With Redis you would get performance, but also durability if you use the WAL, so you don’t risk loosing bookings if your server crashes or restarts.

Of course, it is possible with Elixir to run a single global GenServer in your cluster taking care of each counter, but you would need to take special care in your logic to make it work properly in all cases (think about restarts, deployments, etc.).