Sticky / round robin to solve delay in eventual consistency

[gedw99]

There is a natural delay in replication

Once your round robin load balanced via NATS to 3 instances running in the same DC you might get different query results etc.

fky.io solved it at the LB level by detecting a mutation and making the user who made the mutation be sticky for a few seconds so they get correct results .

I have heavily summarised the technique here. I full explanation is here: https://fly.io/blog/globally-distributed-postgres/

So because fly is NOT multi Master they also has the detect and redirect based on if it’s a SQL read versus a SQL write. We don’t have that in the Multi Master architecture.

It’s a simple sort of brute force approach and could be useful