Félix GV

🧵 How do Quorums Work? I’m continuing to read through the 2nd edition of the excellent DDIA book, as I prepare for teaching a dist db class next semester. I highly recommend the book for all software engineers.

The book says that the quorum condition is met (w + r > n) but that does not seem true, at least not at the point in time where w = 1 and r = 2. The quorum is not met *yet*…

🧵 How do Quorums Work? I’m continuing to read through the 2nd edition of the excellent DDIA book, as I prepare for teaching a dist db class next semester. I highly recommend the book for all software engineers.

If the fastest 2 replicas return inconsistent results, then the client must wait on a response from the 3rd replica to serve as a tie breaker. If 2 out of 3 results are the old value, then that is what would be returned to the client.

🧵 How do Quorums Work? I’m continuing to read through the 2nd edition of the excellent DDIA book, as I prepare for teaching a dist db class next semester. I highly recommend the book for all software engineers.

My understanding of quorums is that reads are sent in parallel to all replicas, and that in the common case there is a "fast path" scenario where the result can be returned after waiting only for the fastest 2 of 3 replicas.

🧵 How do Quorums Work? I’m continuing to read through the 2nd edition of the excellent DDIA book, as I prepare for teaching a dist db class next semester. I highly recommend the book for all software engineers.

The book then concludes that, because of this, the first read request returned the new value while the second read request returned the old value—a breach of Linearizability.

🧵 How do Quorums Work? I’m continuing to read through the 2nd edition of the excellent DDIA book, as I prepare for teaching a dist db class next semester. I highly recommend the book for all software engineers.

While going through the Linearizability portion of chapter 10, I saw a description of quorums that left me perplexed. The book describes a race condition where a write to 3 replicas has reached only 1 of them, and meanwhile two seperate read requests touch 2 replicas each.

🧵 Does @apache.org #Pinot offer strongly-consistent writes nowadays?

The key question is: what is applying the writes? 1. A Kafka consumer? (definitely async, even if nearline) 2. Some incremental batch ingestion job? (Periodic / async, unless the analyst can wait on that job in a blocking fashion?) 3. Some REST/gRPC endpoint? That part I’m not sure if it exists 🤔

🧵 Does @apache.org #Pinot offer strongly-consistent writes nowadays?

The DDIA v2 book makes it sound like this by saying: "From an analyst's point of view, data that has been modified with inserts, updates, or deletes is immediately reflected in subsequent queries. […] Apache Pinot, […] and many other databases do this."

Don’t Sell Yourself Short With New Year Resolutions Last night, I was thinking of a new year’s resolution where I would do one more push up each day, until I reached 365 in a day at the end of the year. Then I thought: why wait? I can already do 365 push ups in a day.

It’s not that hard. I just need to grease the groove and take the time to do it. So today I did 24 sets of 10 reps before lunch, and another 12 sets of 10 before dinner, with 10m breaks after each. And I topped it off with 3 sets of 2 reps each with 27 kg (60 lbs) on my back.

Which Slop is Right? I asked two AIs exactly the same prompt: Does postgres guarantee atomicity of a transaction at all isolation levels? 🤔

And I got these two answers: 1. Yes, PostgreSQL guarantees the atomicity of a transaction at all isolation levels. 2. PostgreSQL does not guarantee atomicity of a transaction at all isolation levels. Which one is wrong 🤷‍♂️ ?