AI Summary
This video is the first in a four-part series on scaling database writes. It addresses a common interview question about handling write throughput at scale, arguing that vertical scaling on a single powerful database instance is often sufficient and that choosing the right database technology is more important than prematurely adding complexity.
Chapters
The interviewer's question about handling 10,000 users may indicate they are stuck in 2010 thinking or want to test your knowledge of vertical scaling. It's a prompt to discuss scaling strategies.
You can rent machines with hundreds of cores for a few thousand dollars a month, so you don't need a massive database cluster to handle large scale. This is a cost-effective and simpler approach.
OpenAI scaled ChatGPT with a single primary database for writes, albeit with hundreds of read replicas for queries. This demonstrates that a single write node can handle significant load.
Calculate your actual write throughput. For example, 10,000 users writing every 10 seconds equals 1,000 transactions per second (TPS), which is trivial for a single database instance.
If you outgrow a single Postgres instance, question whether it's the right database. Postgres uses a B-tree index that updates on every insert, while databases like Cassandra use an LSM tree that appends to a log, making writes faster.
This is part 1 of 4 on scaling writes. The full breakdown is available on hellointerview.com.
The key takeaway is to avoid premature complexity: vertical scaling and choosing the right database can handle most write loads, and you should calculate your actual throughput before designing a distributed system.
Mentioned in this Video
Study Flashcards (4)
What is the approximate cost of renting a machine with hundreds of cores per month?
easy
Click to reveal answer
What is the approximate cost of renting a machine with hundreds of cores per month?
A few thousand dollars.
00:16
How did OpenAI scale ChatGPT's write operations?
medium
Click to reveal answer
How did OpenAI scale ChatGPT's write operations?
They used a single primary database for writes, with hundreds of read replicas for queries.
00:28
If 10,000 users write every 10 seconds, what is the write throughput in TPS?
easy
Click to reveal answer
If 10,000 users write every 10 seconds, what is the write throughput in TPS?
1,000 TPS.
00:42
What is the main difference between Postgres and Cassandra regarding write performance?
medium
Click to reveal answer
What is the main difference between Postgres and Cassandra regarding write performance?
Postgres updates a B-tree on every insert, while Cassandra uses an LSM tree that appends to a log, making writes faster.
00:55
💡 Key Takeaways
Vertical scaling is viable
Challenges the assumption that massive scale requires distributed databases, offering a simpler and cost-effective alternative.
00:16OpenAI's single primary
Provides a real-world example that a single write node can handle immense scale, validating the vertical scaling approach.
00:28Do the math first
Emphasizes the importance of calculating actual throughput before designing for scale, avoiding unnecessary complexity.
00:42Database choice matters
Highlights that the choice of database technology can have a significant impact on write performance, not just scaling.
00:55Full Transcript
[00:02] "We might have 10,000 users. How will we handle rights?" There's only two possibilities. Either A, they're hopelessly stuck in 2010 where that number was large, or B, they want to make sure you know about vertical
[00:16] scaling. You don't need a massive database cluster to handle immense scale. You can rent boxes with hundreds of cores for a few grand a month. Maybe you don't believe this operates at scale. OpenAI scaled out their service
[00:28] of ChatGPT with only a single primary serving rights. Now, that's admittedly with hundreds of read replicas for queries, but you get the point. So, do the math first. What's your actual write throughput, and does it fit on one big
[00:42] beefy box? If we assume 10,000 users are making rights every 10 seconds, that's 1,000 TPS. Nothing. That reflection helps your interviewer know you don't sprinkle complexity on a system that
[00:55] doesn't need it. If you really have outgrown a single Postgres instance, ask if it's even the right database. Postgres updates a B-tree on every insert, whereas other databases like Cassandra use an LSM tree, which
[01:09] basically just append to a log. This is way faster for rights. scaling rights in this series, so follow for more and view the whole scaling rights pattern breakdown on our website hellointerview.com.
