I felt like I didn't understand the blog post very well, either.
It spends a lot time building a benchmark framework and measuring things. At the end, hstore is the choice supported by the data. But instead the author writes:
> If we go strictly with the data, the best way looks to be the hstore-powered solution, but I think the HLL is probably the right choice.
?!? Why bother defining a framework for measurement if you're just going to substitute your own judgment? Perhaps the framework didn't capture something important -- like the absolute number of people looking at any given post, concurrent updates, etc.
I'm also confused by the benchmark. The autocannon script has:
// 10% of the time, get *all* the users
request.method = 'GET'
request.path = `/posts/${postId}/seen-by/users`
Does this imply there's a path it can hit to get the list of users who saw a post? HLL is a probabilistic data structure, so it doesn't support that with 100% accuracy. You could reverse engineer it from the list of users, but the performance will vary with the number of users, which didn't seem to be captured by the benchmark. I tried to look at the source code, but the GitLab URL was 404ing for me.
It spends a lot time building a benchmark framework and measuring things. At the end, hstore is the choice supported by the data. But instead the author writes:
> If we go strictly with the data, the best way looks to be the hstore-powered solution, but I think the HLL is probably the right choice.
?!? Why bother defining a framework for measurement if you're just going to substitute your own judgment? Perhaps the framework didn't capture something important -- like the absolute number of people looking at any given post, concurrent updates, etc.
I'm also confused by the benchmark. The autocannon script has:
Does this imply there's a path it can hit to get the list of users who saw a post? HLL is a probabilistic data structure, so it doesn't support that with 100% accuracy. You could reverse engineer it from the list of users, but the performance will vary with the number of users, which didn't seem to be captured by the benchmark. I tried to look at the source code, but the GitLab URL was 404ing for me.