ºÚÁÏÃÅ

Episode transcript

Ìý

Scott Shaw: Welcome, everybody, to the ºÚÁÏÃÅ Technology Podcast. I'm going to be the host today. My name is Scott Shaw, and I'm joined today with co-host Rebecca Parsons. Would you like to introduce yourself, Rebecca?

Ìý

Rebecca Parsons: Hello, everyone. This is Rebecca Parsons, another one of your regular co-hosts. I'm the CTO Emerita, and we are joined today by another Thoughtworker based in India who has just written a book — or finished a book! [chuckles] Unmesh, would you like to introduce yourself?

Ìý

Unmesh Joshi: Hey. Hi, everyone. I'm a principal consultant at ºÚÁÏÃÅ and been with ºÚÁÏÃÅ for 16 years now. As Rebecca said I’d been involved in writing about distributed systems for the last few years which is now compiled into a book which is out.

Ìý

Scott: That's quite an accomplishment. Unmesh and I worked together 10 years ago or more back in Pune, India, so it's a pleasure to be talking to you again. The book is called Patterns of Distributed Systems which is a pretty big topic, and I think it's also something that's probably not that well documented elsewhere, and so I think this is a unique book and probably fills a needed gap in the market right now. I wonder, how was it that you first came to want to write a book? How did this get started initially?

Ìý

Unmesh: It's a long journey. I think it started back in 2015, 2016, when some questions started coming to my mind. Typically if you see the digital transformation projects or any project for that matter, you work on for last several years, they involve cloud services, cloud like Amazon, GCP, and Azure and also products like Kafka and Cassandra and other distributed databases, or Kubernetes now that everyone uses that.

Ìý

I was like everyone using it or started using it back then 2017-ish. I was trying to learn all of these things, struggling my way through it. At the same time, there was some interesting work that a team in our Pune office was doing, and I just happened to work with that team to help them out. This was for writing software for a large optical telescope called Thirty Meter Telescope, and a lot of problems that they had to solve, essentially keep data distributed and distributed copies of data in sync have different telescope subsystems coordinate and work towards some task.

Ìý

All those I thought were similar to what I was learning about systems like Kafka or cloud services, but for this Thirty Meter Telescope thing, the challenge was that they had to build things from scratch. There was a need to understand all these things in a little more depth so that you are confident that the decisions that you are taking are right. At that time, I started going a little more deep and trying to understand all these things.

Ìý

Things like Paxos, for example, or things like Lamport clock. When I went to books, the books were very theoretical, I thought. They explain a concept; they talk about asynchronous systems and synchronous systems, even things like Lamport clock or Paxos, but there is always some mathematical abstraction or description in terms of pseudo code, and you don't really see how Lamport clock is used in practice.

Ìý

There was that huge gap now that I struggled with. What I decided was to really go to open-source code bases, and that's the beauty of open-source. I think there is so much code available for you to look at. I started going through Kafka's code base and Cassandra's code base and various other code bases. Then I figured out that-- I could connect dots. What I was reading about and what I could see in the code I could connect those dots, but there were also some gaps.

Ìý

What was explained in the book versus what was actually implemented, there were some gaps. What I decided to do was-- Kafka as a whole is a big system. It does so many things, so if I focus on maybe step by step, if I have to build something like it what all decisions I need to take and what are the different components I need to code. I started doing that. I coded my own miniature Kafka. I coded my own miniature Cassandra and I made sure that the code is not toy code.

Ìý

It resembles what I actually see in production code of Kafka or Cassandra or similar tools. That worked pretty well because then I had real code to look at. When I was talking to people they also had real code to reason about, and that worked pretty well on that Thirty Meter Telescope project. I did a couple of workshops using this technique. Just don't talk about distributed system theory, but let's say let's build Kafka-like thing in a few days.

Ìý

This approach of tackling one problem at a time and then the next problem, and then weaving multiple things together, and then you have this big system which looks complex. If you look at it as a whole, it looks complex, but once you focus on one thing at a time it's a bit easier. Then I happened to talk to Martin about what I was doing, and Martin has done this kind of thing in the past with various systems, like his Patterns of Enterprise Architecture book, I think is a similar thing done to the world of object traditional mapping frameworks.

Ìý

He liked the idea and he actually suggested that I write patterns because the small components and small problems I was trying to solve one at a time, now he thought that can be very well explained as a pattern. That's where it started. That's when around start of 2020, the start of the pandemic, I started documenting these patterns, and in last three years, I think we got to a stage where we had enough patterns to maybe explain something like Kubernetes or Kafka and then we decided to compile it into a book.

Ìý

Scott: It was helpful, I guess, that we were locked down in a lot of places, and you were stuck at home, I imagine and had a lot of time on your hands to write?

Ìý

Unmesh: Lockdowns and during the pandemic, I think that gave me a lot of time to work on these things.

Ìý

Scott: Did you find it was a difficult transition from writing code to communicate to other people, to writing in English, to describing these things in English in a concise way? Did it take you a while to get into the pattern of writing patterns?

Ìý

Unmesh: No. Actually having code in front of me or in front of people while I explained things in English, that actually worked pretty well. I think that's one of the beauties of the patterns approach. I think that patterns, you have some well-defined names. Like Lamport clock is a well-known name, or a write-ahead log is a well-known name. Or even Paxos is a well-known name.

Ìý

When you read just English with, let's say pseudo-code or some abstraction you need to imagine things, and then you make some vague impressions in your mind of what that thing really is. Now, as opposed to that, when you have real code in front of you which highlights what Paxos looks like or what Lamport clock looks like, it's actually very easy. Even for me, writing was easy after the code and some examples were done.

Ìý

The hardest part was to get that code that shows the essence of a particular pattern, and then writing about that around that was the easier thing.

Ìý

Scott: I think that's interesting. I think it reflects the approach to technology that we have at ºÚÁÏÃÅ, which is you can talk all you want about your opinions, but until you write the code, you don't actually know and the opinion is walk, code, talk. I think that's the way we try to ground things usually in practice. I think that's an interesting way to approach it.

Ìý

Unmesh: Martin often calls and, I like his sentence, he says code is like mathematics of our profession because when you write code and execute it, you cannot cut corners. When you're drawing on the board or writing about something. You can cut corners and that's true about a lot of books on distributed systems as well. Even if the focus is to prove that something works, there are certain assumptions and those assumptions might fail in practice or you can very well ignore some of those assumptions that are out there.

Ìý

If you're not aware of those you can run into real issues in production.

Ìý

Scott: I wonder how many other people have been through this experience. I know I did years ago when I first naively tried to start building distributed systems and then ended up reinventing a lot of the things. I pulled this book out because this was a really influential book for me. I don't know if you can see it, but it's Jim Gray's book on Transaction Processing. I was so relieved when I found, oh, somebody's actually documented how these things work and I think your book is probably going to be similar for people.

Ìý

It's ironic that we live in a world where distributed computing is ubiquitous. With the cloud, everything is distributed, and yet there are so many abstractions and tools and open-source frameworks that remove a lot of the details of that away from us. That, even though we're using it more, I think developers who are just starting in this probably learn less about it because they don't have to actually build it themselves. Do you try to get beyond those abstractions?

Ìý

Unmesh: Oh, yes. That was one of the one of the main goals that to go beyond abstractions and to show real code. I really liked when I saw Jim Gray's book that you showed, because that was something that I wanted to do with distributed systems. It's sad that Jim Gray's book came in the 1990s. It's excellent because it does the same thing that I wanted to do with distributed systems.

Ìý

It doesn't leave things to imagination. It gives you very concrete things that you observe in systems at the code level. You can read and be sure that this is something that you can expect. For even things like as I said, the book has four broader sections, one bigger chunk is about replication. One is about how you manage time in distributed systems and then how you manage a cluster server to communicate with Azure. Those are the third sections.

Ìý

Something as obvious as replication has a lot of complexities. I think it's not obvious to developers when they read about replication because the term, at least people think that it's a well understood term where in practice it is really not because because replicating data on multiple servers and then making sure that they agree on the data that's replicated it requires consensus.

Ìý

Really systems which give you that a guarantee really need to implement something like Paxos or Raft. Now you see that most systems, or most data systems including Kafka and all the newer databases which implement this replication, they implement Raft or Paxos as that replication mechanism. Having concrete understanding of what goes on in something that's at least perceived to be obvious is important. I think in the book I try to go into the details of at the implementation level, how this replication mechanisms work.

Ìý

Rebecca: One of the things I was wondering about a feature, if you will, of patterns is that you can talk about when they are applicable. Given the fact that some of the things that you're talking about are somewhat fundamental notions, there are, for example, lots of ways you might decide to use distributed time. As you were looking at these different code bases, did you mostly run across essentially the same implementations, or were there times when there might have been different implementations for one of these more fundamental constructs that you were able to examine?

Ìý

Unmesh: Yes. I think again, the strengths of the patterns approach is, you look at multiple things, and the rule of three is generally followed that you look at at least three things before you call something as a pattern and then there is this concept of similarity. In various systems the implementations are similar, but not exactly the same. Patterns allow you to document what similar thing is across these systems. There can be variations that you then document as well.

Ìý

About the replication and distributed time as I said, if you consider replication that I said is complicated than what people think it is. If you consider Raft implementation that's used in products like CD or more modern new SQL databases and you look at Kafka's replication protocol there are similarities, but they are not exactly the same. For example, the way Kafka maintains high watermark and Raft has something called a commit index to differentiate between what is a stable portion of the log and what is the portion that you're not sure about.

Ìý

It is the same the way they use something that I call as a generation clock or Kafka call it epoch and Raft calls it its own term, but that's really to detect if there are stale leaders out there which are trying to cut across your new leader. Those things are similar but not exactly the same. What I have tried to document is the similar portions with differences documented as well.

Ìý

For distributed time, it's similar, I think. Lamport clock is very, very fundamental and all key value stores use it essentially to detect ordering of rights, but what you see in practice is a variation of that Lamport clock called hybrid clock that's used. Because Lamport clock as as it's documented, is just a simple integer counter that you use, and then you pass it across all your messages and they're stored with the values that are stored as well to make sure that you detect which rights happen before other.

Ìý

In practice in the data world, you typically will like to have real date time and real timestamp attached to these values to detect which values are written earlier and later. This Lamport clock technique is is used along with system timestamp. You use that counter, but a system timestamp as well. I'm saying system timestamp, but this is another thing that's not well understood. I think that different servers when they use system time, the time that your machine gives you, or the server gives you, there is no guarantee that those will be in sync.

Ìý

That's a big problem in practice because then you can have issues where your newer rights, if they are assigned an older timestamp or a lower timestamp, they can get overwritten by something that was written before but was assigned a higher timestamp. That's a big challenge. Most practical systems, they combine this Lamport clock technique along with this system timestamp, and that's the hybrid clock technique that's used in a very similar way across all data systems.

Ìý

Like, MongoDB uses it the same way, CockroachDB uses it the same way, YugabyteDB uses it the same way. Other advantage of, again, the patterns approach is that if you see all these systems, they are written in different programming languages, from Scala to C++ to Rlang, to Golang, to Rust. The broader implementation structure of all these things, it remains the same, even if the language of implementation differs.

Ìý

Scott: It's interesting. I see some parallel between the work that you've done and the Jepsen Project in verifying perhaps the claims of some of these systems and actually digging into the details to see how they really work. Even with open source, there are so many claims made that are a little bit superficial that don't really go into these details. I'm curious if you found any contradictions when you were digging into the code or errors or things that didn't quite work the way they were advertised.

Ìý

Unmesh: Not contradictions, but things that are omitted from the descriptions. Like Paxos, for example, it's very, very popular replication algorithm or consensus algorithm. There are some obvious things. Paxos stops at the point where it says that okay, multiple nodes now agree on one value or one command to execute, and then one of the nodes knows about it. Then, how do you make sure that all the nodes know about it?

Ìý

Because in a distributed database, for example, it's not enough for one node to know about it. You need to have all the replicas to remain consistent or to have the same data. Paxos doesn't talk about that. What's called full replication that you make sure that everyone finally reaches the same state, Paxos doesn't talk about that. This looked like an obvious question that anyone would ask.

Ìý

I was searching in implementations, there are some implementations of Paxos which are popular in academic circles, but they don't solve this problem, and they don't even talk about this topic. It's really fascinating and surprising that it required a new PhD thesis, that's Raft, which clearly talked about that. This is also a problem that you need to solve, along with basic consensus. It gives you the complete description of that.

Ìý

Or there are people or the implementations which claim that they use Paxos like Cassandra. They use their own extension to solve this particular issue. That extension works, I think, but that's an extension to what is proven by Paxos. There were things like that. That's the advantage that I found while working on this, that you get these questions. Even when these questions that people hesitate to talk about, they are real questions that can bite you in your production implementations.

Ìý

Like with Raft, for example, Raft's leader direction and any Paxos implementation, for example, can go into LiveLog. All the nodes, they just keep on tripping the leaders and just go into a phase where there is data detection happening with no real work that any node is able to do. This is a real problem. Even if you have good Raft implementations, they can go into this problem. A couple of years back, I think there was an outage in Cloudflare system. It was, I think a six-hour long outage.

Ìý

This primarily really happened because of this LiveLog in the Raft implementation of its CD. This can happen in your Kafka cluster or Kubernetes cluster, which use this Raft algorithm. Some of these loopholes in these implementations, you at least need to be aware of that, okay, something like this can go wrong. Even if the documentation doesn't say aloud, that this can go wrong. This can go wrong in practice.

Ìý

Or even things like I talked about hybrid clock. With hybrid clock or the basic characteristic of Lamport clock, that you only get partial order. When I say partial order, if there are two independent writes happening on two different nodes, you cannot really figure out which happened first and which happened later. It's only when two nodes are talking to each other or writing to the same key, let's say, you can detect the order.

Ìý

If there are independent keys written on independent servers, you can't detect the order. For that, Google has their own infrastructure with something called as true time. Open-source implementations, which can be deployed on a data center or different cloud settings than Google, they don't have luxury of this true time. Then what they need to do is they need to assume what can be the maximum clock drift or clock skew across servers. I have documented this as a pattern called as clock-bound wait.

Ìý

That you just wait to make sure that the timestamp at which you are writing, all the other servers have actually reached that timestamp before you make that write visible to others. Some of these are subtle things. If these assumptions go wrong, that maximum clock skew if you assume to be, let's say, 200 milliseconds, and in reality, it's 500 milliseconds or more than that, which can very well happen, then you can see those writes done in the opposite order to what they really are done. These things, I think you will get to know or at least question some of the documentation.

Ìý

Rebecca: I think that speaks to the whole purpose, though, of writing a book that lays out not just the pretty abstraction, but the reality of the implementation details because you're not going to run across a problem, most likely with a pretty abstraction. You're going to run into a problem in production when reality hits the code. It is truly amazing how often people will use libraries and system tools like this without really understanding what's under the hood.

Ìý

You can understand that because maybe it works 99% of the time. It's a lot of work to do what you've done over the last three years to really understand these things. I do think that the role that this book has is it provides the mechanism where you can get-- it's the middle ground, if you will, between the nice pretty abstraction and somebody having to troll through the entire commit history of Kafka to figure out how it works. This is a nice middle ground.

Ìý

Unmesh: Absolutely, I think. One of other observations I had when I started my career, just to relate to the times we are now, to the times back in the late '90s, I think. As an industry, I think when we don't know how to make sure that people understand certain concepts, people fall back on certifications. You see the times where everyone is asking for Amazon certifications and GCP certifications and stuff like that, because we don't know how to make sure that people understand how these systems work.

Ìý

Understanding these patterns or the basic principles of how, for example, Amazon builds the services that it provides, it gives you a good way to understand broader spectrum of systems. You don't need to do individual certifications because those are surface level things that you will get below with that. I think something like this will help solve or breach that gap that we are pacing today.

Ìý

Scott: There's so much that a software engineer has to know today. What of this material, or how deeply into this material, do you think people should have to delve just to reach the lowest bar in qualification to work in cloud or with Kafka or something?

Ìý

Unmesh: In the book, we actually have divided it into two parts. That's very typical of patterns' books, I think. First section just gives the overview of the patterns, and it talks about all the problems and just a cursory view of how a particular pattern solves it, and then how they link together. To start with, I think it's easier to just get this broad view of what things can go wrong and how different implementations tackle it.

Ìý

Then when you are working on something and maybe need to deep dive into one specific thing, then you can go and read a pattern that talks about that thing. That's the best way to use this, I think. Yes, because that's another aspect of pattern's book as well. I mean, you can't read it end to end. It's a hard thing to read, end to end. You have to focus on one thing and read it.

Ìý

Scott: Given our limited attention spans now, I actually like that. I like that I can pick it up and absorb one thing and then put it down till the next time.

Ìý

Rebecca: We did a podcast about three years ago when you first started writing, what would you say is the major change in the things that you've talked about, from that first podcast to today? Was there one big section that you wanted to finish and that's been done in that time?

Ìý

Unmesh: Oh, yes. Two major things. I think documented patterns of distributed time, that's a major chunk. Also detailed out things like Paxos and replicated log, which are foundational to any consensus mechanism used for replication. Those two are the major changes since last time we talked.

Ìý

Rebecca: Have you soured on ever writing a book ever again? [laughs]

Ìý

Unmesh: Yes, possibly. After some break that I need.

Ìý

Scott: Will you be doing any workshops?

Ìý

Unmesh: Yes. I'm doing workshops, four already, which I was doing once a quarter, I think, while I was working on the book. I will do more of those in future, I think.

Ìý

Scott: I'm sure there's going to be an audience for it. There'd be a big demand for it. Some of the conferences that I've been to, I think people would be really interested. Anything else you wanted to say, Unmesh?

Ìý

Unmesh: No. Just read a book and send me questions if you like. One of my hopes is that people will get questions and openly start talking about obvious things that might be missing from popular implementations. Yes. I like to use this term wise fool. Wise fool is someone, you know the story of Emperor's new clothes, and it's the key that says that the emperor is wearing no clothes, so when everyone was hesitating-- so we need more of those talking about obvious things and discussing them aloud.

Ìý

Scott: Yes. I agree. Well, once again, congratulations. I think this is a huge accomplishment, and you should be very proud of yourself for having got through that. I'm looking forward to reading more of the book. I've read some of the patterns, but I'm looking forward to reading more of them, and best of luck.

Ìý

Unmesh: Thank you. Thank you, Scott. Thank you, Rebecca.

Ìý

Rebecca: Thanks for joining us.