"Brewer’s Conjecture and the Feasibility of Consistent, Available, Partition-Tolerant Web Services" (The formalization of CAP, which caused a ton of very poor trade-off thinking in the decade that followed by defining Availability in a very goofy way. Still a classic.)
"Practical Byzantine fault tolerance" (Moved the conversation on Byzantine faults forward significantly).
This is just a short selection. There's so much good stuff going back in the 70s and 80s distributed database literature, for example (and in the modern systems and DB literature too).
Lamport lists his collected works on his site[0] along with some commentary for each one, and directly mentions, The Maintenance of Duplicate Databases, as the origin for logical clock paper. Many of the notes have interesting commentary.
> Many computer scientists claim to have read it. But I have rarely encountered anyone who was aware that the paper said anything about state machines. People seem to think that it is about either the causality relation on events in a distributed system, or the distributed mutual exclusion problem. People have insisted that there is nothing about state machines in the paper. I've even had to go back and reread it to convince myself that I really did remember what I had written.[1]
I gradually realized that Lamport is more of the godfather of distributed system than Hinton is to deep learning. Lamport is less prominent than Shannon is to information theory. Shannon is the closest to any title of "gold-like" figure to a scientific discipline of universal relevance in modern society.
Lamport specifically revealed a philosophical connections between computer systems and physics, in the parallel of distributed consensus to relativity theory. To me, the enlightenment is that, the relations between events happening in a distributed system, is more fundamental than their absolute ordering, thus the central role of an "observer". I haven't really analyzed if this realization is from Lamport's papers, or my general reading and thinking, but I am moderately confident that general readings are heavily influenced by Lamport's papers, or can be traced back to be compatible with Lamport's thinking. I have not seriously study if this connection is valid in depth, which might be another pure amateur speculation of mine.
One thing I think Lamport falls short is that his writing is not easy to read and understand. I unconciously feel that Lamport (and Dario from Anthropic) probably share a hidden sense of intellectual supriority grew from their own experience throughout their career. So their writing (and Dario's gospel) all share a unchangable sense of narration from their own delicate and graceful ideas, much less of faciliating the understanding to their audience. In this cateogry, Shannon is abosolutely superior in any measure, in his writing, ideas are so naturally presented, although the implications of the ideas remain elusive due to the inherent depth.
Also, among the 3 prominent figures of modern AI: Hinton/Bengio are more like Shannon, Lecun is closer to Lamport.
Enough random rambling. Lamport, as indicated by the outweight presence in this list, is no doubt the single most important scientist in distributed systems.
> the relations between events happening in a distributed system, is more fundamental than their absolute ordering
The important thing in most distributed systems is having an order. Having a single observer serialize events as it receives them is so much more tractable than trying to use absolute order. Using absolute order requires very precise time synchronization which is hard; using absolute order requires knowing when you have received all the reports of events that already happened which is hard. Determining a designated observer isn't typically easy, but having it determine the order it observes events is easy. If two events happen at a similar time it's typically not a big deal which one is considered first as long as all nodes will agree on the result --- let the designated observer just pick the first one it sees works pretty well. If your report takes an unexpectedly long time to make it to the designated observer, then it won't be first and you'll deal.
Much better than trying to figure out unknowable questions of relativity. :P
yes, it can be easier to have a central serializer for events, but that certainly makes things problematic for fault tolerance and basically excludes large-scale solutions. using an agreed-upon post-hoc ordering based on timestamps is certainly another way, but really only if you're working in a paradigm that lets you impose it (like mvcc) or doesn't care (like crdt).
personally I find when you view events in a distributed system as a partial order, its more liberating than confusing. its not unusual to assume that there is some kind of canonical event ordering that we have to preserve, when its often just not semantically important. so its a useless constraint that can
impose complexity and limit the solution space. the partial ordering exposes the real causal constraints.
Computer scientists are in denial about it, but CS is a branch of theoretical physics, not mathematics. You can point to this or that model of computation, such as lambda calculus or mu-recursive functions and try to claim its abstracted well beyond the particular laws of physics for some specific universe, but they all have some kind of rate limit built into them... and where does the motivation for this idea, that it takes something (time, space, work) to compute something ultimately come from? That's right - from underlying physics itself[1] - from the Bekenstein bound or Bremermann's limit or the like.
Even apparently non-physically-realizable models of computation like non-deterministic Turing machines are ultimately informed by and motivated by concepts in physics... otherwise they would just be examples of chmess[2] and of no interest to anyone. Computer science is of course somewhat abstracted from the details, but no more so than, say, thermodynamics, where concepts like entropy or Gibbs free energy can be studied in the abstract without reference to whether we are talking about a gas of non-interacting molecules or the spins of a bunch of electrons trapped in a lattice.
So, it's of no surprise whatsoever that the fundamental problems of distributed computing are ultimately the same as those found in the relativity of simultaneity[3]. You've all been studying the same things all along, just with different tools and at different levels of abstraction.
Maybe thats why one of the not-so-bad ideas how to go about distributed systems came from a guy who was trained as a physicist and used to complain to his fellow programmers that “a lot of systems actually break the laws of physics”[1]?
> In distributed systems there is no real shared state (imagine one machine in the USA another in Sweden) where is the shared state? In the middle of the Atlantic? - shared state breaks laws of physics. State changes are propagated at the speed of light - we always know how things were at a remote site not how they are now. What we know is what they last told us. If you make a software abstraction that ignores this fact you’ll be in trouble.[2]
I suspect this is old news for you, but just in case you haven't heard of it, check out Feynman's Lectures on Computation. A surprising amount of the book is still relevant, and it's fun how much he always brings everything back to the physics.
No disrespect meant to Mr. Armstrong, but it's possible it's never listed because it's basically a textbook. It's 295 pages. The rough average of all papers in OP is like 15 pages.
I would say 300 is more than slightly longer. Varies by institution and style, but 100k words, 4-5 collated papers, 100-150 pages maximum are quite common.
I would be concerned as an examiner if this came across my desk, more so if I read the colophon where the author comments that they intended to write their own typesetting system, before reading Knuth and wisely concluding that they were unlikely to do anything better than TeX. Top tier yak shaving there.
In distributed systems the fist lesson is resilience is more important.
Spend enough time in a computer lab, and you will see things halt and catch fire on occasion. Especially if it has a bunch of GPUs pinning the utilization 24/7, or a cheap power supply in the cluster. =3
Not OP, but a good resource for distributed consensus specifically is Tim Roughgarden's YouTube playlist "Foundations of Blockchains" [0] (hear me out, despite the title - see below). It's 85 videos over 12 lectures, rigorous, very well explained (but assumes some CS fundamentals). Not original work, but gives the conceptual framework and background so that one can read these classic papers in context.
Per-lecture reading lists are on the course page [1]. He also points at Elaine Shi's Foundations of Distributed Consensus and Blockchains [2] and Andrew Lewis-Pye's Consensus in 50 pages [2] as background, though these are more textbooks, not papers.
Remarkably, the first 7 lectures deal with permissioned systems, recapitulating the classic results of consensus in distributed systems (Dolev Strong, FLP impossibility, CAP) - no blockchain in sight. This takes us to the state of the art at the end of the 1990's (with algorithms that can achieve consensus in the presence of byzantine failures, namely Byzantine Paxos and PBFT, though he discusses a modern variant, permissioned Tendermint from 2014).
Lecture 8 stays permissioned and proves consistency and chain quality for the longest-chain rule. Only at Lecture 9, with proof of work, does anything specifically blockchain appear; then L10 block rewards and selfish mining, L11 transaction fee mechanism design; L12 proof-of-stake sybil resistance with the whole litany of attacks possible there.
Very good series in my view, and shows how little technical merit this whole blockchain circus has - nearly all the great properties people tout (reliability, consistency, audibility, availability) can be achieved with good old permissioned tech more efficiently, with pretty instant and deterministic finality.
Anyway, I found the series worth watching for the classical consensus material alone.
Yes, but it is boring telecom projects... so most people aren't interested.
Scala is kind of a more modern alternative.
If doing a twitter like platform, elixir with Phoenix channels could handle around >20k users per host. Very few other options work for that use case. =3
There's still a lot of WhatsApp Erlang tooling being published, so I'm pretty sure it's mostly still Erlang. I think I still know a couple people who would share with me if it was rewritten, although I left in 2019.
Erlang is simply the right language to build a chat server with, and it doesn't make sense to rewrite it in another language, even though Facebook did that with the server for Messenger. :P
Certainly parts of the WhatsApp stack have changed considerably. No more FreeBSD / no more only a handful of OS processes per node, much less Mnesia (not sure if there's any Mnesia left... I don't get that level of information). I've seen signs that they may have returned to large nodes, so they might be back to millions of connections per node, but that was very much not the case when I left; connection counts had dropped on our FreeBSD nodes because we were doing a lot more on the chat channel and then again because the Facebook nodes we migrated to were so much smaller than our FreeBSD nodes.
It's the thing that baffles me the most with those people, the Knuths and Lamports and Carmarcks, and whatnot.
It's not like they have light schedules - they do research, they have classes, they attendconference, and they have jobs, etc...
And some benefit from tenure, which give them time to deeply focus on some topics.
But I can't imagine how I would write tex in ten years in Pascal on archaic machines if I had all the time to myself.
Doing that in the middle of understanding more computer science than my eyes would ever read ?
What do they do with their time ? What do they don't do ? Are they able to skip the eating and pooping and sleeping and cleaning their house and picking up parcels at the post office and grocery shopping parts ?
Or a proto Unix being written in three weeks. Concretely, what are the days of those three weeks made of ? Especially since it's famously when the spouse was NOT there to handle the housekeeping...
I see not amount of "life hacks" and "atomic habits" and whatnot helping with that. But it's probably just as natural to them as biking - and you can't explain biking, I guess ?
A favorite Knuth story to illustrate: He was re-writing the 8080 code for the processor that ran our new Alphatype CRS Phototypesetter, since the manufacturer's code wasn't suitable. Real-time, 7 levels of simultaneous interrupt stuff, since it controlled two stepper motors that moved the lens around the photographic paper, plus four custom font outline-tracing S100 cards that flashed vertical segments of characters at just the right moment on the CRT that the lens then focused onto the paper. No debugability other than a 4x4 keyboard and a 15(?) character display. I was responsible for the code on the other end of the RS232 port that connected it to the DEC mainframe, sending down character outlines on the fly to the very limited memory on the 8080, along with the properly-sorted character placement info ultimately from TeX.
Anyway, we'd been doing morning, noon, and night debugging sessions together for a number of days. On maybe the 3rd morning, I arrived at his office and was speaking with his secretary of many years, Phyllis Winkler, when Knuth came in. He handed her a stack of maybe a dozen or two pages, saying "Here's a paper I wrote; please type it in." When? When did he write it? It's still a mystery to me.
Obligatory shout out to Morris and Kaashoek's class at MIT - they update the reading list every year, and it's always great: https://pdos.csail.mit.edu/6.824/
This is not a bad list for sure. Here are some deeper cuts for those looking for something a bit less mainstream:
"The Maintenance of Duplicate Databases" https://datatracker.ietf.org/doc/html/rfc677 (AFAIK the genesis of the use of logical clocks in distributed systems).
"Chain Replication for Supporting High Throughput and Availability" https://www.usenix.org/legacy/event/osdi04/tech/full_papers/... (Chain replication is how a huge percentage of real-world cloud-scale data replication is done).
"Brewer’s Conjecture and the Feasibility of Consistent, Available, Partition-Tolerant Web Services" (The formalization of CAP, which caused a ton of very poor trade-off thinking in the decade that followed by defining Availability in a very goofy way. Still a classic.)
"Paxos Made Live" https://research.google/pubs/paxos-made-live-an-engineering-... (Brought a much-needed engineering perspective to a conversation that was largely theoretical up until this time.)
"Practical Byzantine fault tolerance" (Moved the conversation on Byzantine faults forward significantly).
This is just a short selection. There's so much good stuff going back in the 70s and 80s distributed database literature, for example (and in the modern systems and DB literature too).
Lamport lists his collected works on his site[0] along with some commentary for each one, and directly mentions, The Maintenance of Duplicate Databases, as the origin for logical clock paper. Many of the notes have interesting commentary.
> Many computer scientists claim to have read it. But I have rarely encountered anyone who was aware that the paper said anything about state machines. People seem to think that it is about either the causality relation on events in a distributed system, or the distributed mutual exclusion problem. People have insisted that there is nothing about state machines in the paper. I've even had to go back and reread it to convince myself that I really did remember what I had written.[1]
0: https://lamport.azurewebsites.net 1: https://lamport.azurewebsites.net/pubs/pubs.html#time-clocks
I am a Lamport admirer.
I gradually realized that Lamport is more of the godfather of distributed system than Hinton is to deep learning. Lamport is less prominent than Shannon is to information theory. Shannon is the closest to any title of "gold-like" figure to a scientific discipline of universal relevance in modern society.
Lamport specifically revealed a philosophical connections between computer systems and physics, in the parallel of distributed consensus to relativity theory. To me, the enlightenment is that, the relations between events happening in a distributed system, is more fundamental than their absolute ordering, thus the central role of an "observer". I haven't really analyzed if this realization is from Lamport's papers, or my general reading and thinking, but I am moderately confident that general readings are heavily influenced by Lamport's papers, or can be traced back to be compatible with Lamport's thinking. I have not seriously study if this connection is valid in depth, which might be another pure amateur speculation of mine.
One thing I think Lamport falls short is that his writing is not easy to read and understand. I unconciously feel that Lamport (and Dario from Anthropic) probably share a hidden sense of intellectual supriority grew from their own experience throughout their career. So their writing (and Dario's gospel) all share a unchangable sense of narration from their own delicate and graceful ideas, much less of faciliating the understanding to their audience. In this cateogry, Shannon is abosolutely superior in any measure, in his writing, ideas are so naturally presented, although the implications of the ideas remain elusive due to the inherent depth.
Also, among the 3 prominent figures of modern AI: Hinton/Bengio are more like Shannon, Lecun is closer to Lamport.
Enough random rambling. Lamport, as indicated by the outweight presence in this list, is no doubt the single most important scientist in distributed systems.
> the relations between events happening in a distributed system, is more fundamental than their absolute ordering
The important thing in most distributed systems is having an order. Having a single observer serialize events as it receives them is so much more tractable than trying to use absolute order. Using absolute order requires very precise time synchronization which is hard; using absolute order requires knowing when you have received all the reports of events that already happened which is hard. Determining a designated observer isn't typically easy, but having it determine the order it observes events is easy. If two events happen at a similar time it's typically not a big deal which one is considered first as long as all nodes will agree on the result --- let the designated observer just pick the first one it sees works pretty well. If your report takes an unexpectedly long time to make it to the designated observer, then it won't be first and you'll deal.
Much better than trying to figure out unknowable questions of relativity. :P
yes, it can be easier to have a central serializer for events, but that certainly makes things problematic for fault tolerance and basically excludes large-scale solutions. using an agreed-upon post-hoc ordering based on timestamps is certainly another way, but really only if you're working in a paradigm that lets you impose it (like mvcc) or doesn't care (like crdt).
personally I find when you view events in a distributed system as a partial order, its more liberating than confusing. its not unusual to assume that there is some kind of canonical event ordering that we have to preserve, when its often just not semantically important. so its a useless constraint that can impose complexity and limit the solution space. the partial ordering exposes the real causal constraints.
> absolute order requires very precise time synchronization which is hard
Presumably relativity is the reason precise time synchronization (and thus absolute ordering) is hard.
Hot take of the day:
Computer scientists are in denial about it, but CS is a branch of theoretical physics, not mathematics. You can point to this or that model of computation, such as lambda calculus or mu-recursive functions and try to claim its abstracted well beyond the particular laws of physics for some specific universe, but they all have some kind of rate limit built into them... and where does the motivation for this idea, that it takes something (time, space, work) to compute something ultimately come from? That's right - from underlying physics itself[1] - from the Bekenstein bound or Bremermann's limit or the like.
Even apparently non-physically-realizable models of computation like non-deterministic Turing machines are ultimately informed by and motivated by concepts in physics... otherwise they would just be examples of chmess[2] and of no interest to anyone. Computer science is of course somewhat abstracted from the details, but no more so than, say, thermodynamics, where concepts like entropy or Gibbs free energy can be studied in the abstract without reference to whether we are talking about a gas of non-interacting molecules or the spins of a bunch of electrons trapped in a lattice.
So, it's of no surprise whatsoever that the fundamental problems of distributed computing are ultimately the same as those found in the relativity of simultaneity[3]. You've all been studying the same things all along, just with different tools and at different levels of abstraction.
[1]: https://en.wikipedia.org/wiki/Limits_of_computation
[2]: https://link.springer.com/article/10.1007/s11245-006-0005-2
[3]: https://en.wikipedia.org/wiki/Relativity_of_simultaneity
Related hot take
Maybe thats why one of the not-so-bad ideas how to go about distributed systems came from a guy who was trained as a physicist and used to complain to his fellow programmers that “a lot of systems actually break the laws of physics”[1]?
> In distributed systems there is no real shared state (imagine one machine in the USA another in Sweden) where is the shared state? In the middle of the Atlantic? - shared state breaks laws of physics. State changes are propagated at the speed of light - we always know how things were at a remote site not how they are now. What we know is what they last told us. If you make a software abstraction that ignores this fact you’ll be in trouble.[2]
[1]: “The Mess We’re In”, 2014 https://www.youtube.com/watch?v=lKXe3HUG2l4 [2]: https://news.ycombinator.com/item?id=19708900
I suspect this is old news for you, but just in case you haven't heard of it, check out Feynman's Lectures on Computation. A surprising amount of the book is still relevant, and it's fun how much he always brings everything back to the physics.
As always, lists like these don’t include Joe Armstrong's PhD thesis “Making reliable distributed systems in the presence of software errors” - http://erlang.org/download/armstrong_thesis_2003.pdf
No disrespect meant to Mr. Armstrong, but it's possible it's never listed because it's basically a textbook. It's 295 pages. The rough average of all papers in OP is like 15 pages.
One can skip the Erlang-specific things/the description of the programming language.
Just the first 2 chapters (~30p) + the Conclusion (~10p) contain a lot of useful food for thought.
His thesis is only slightly longer than typical PhD theses, I think.
Yes, but the other things in the OP list aren't theses, they're journal papers.
I would say 300 is more than slightly longer. Varies by institution and style, but 100k words, 4-5 collated papers, 100-150 pages maximum are quite common.
I would be concerned as an examiner if this came across my desk, more so if I read the colophon where the author comments that they intended to write their own typesetting system, before reading Knuth and wisely concluding that they were unlikely to do anything better than TeX. Top tier yak shaving there.
reliable? we all wish...
In distributed systems the fist lesson is resilience is more important.
Spend enough time in a computer lab, and you will see things halt and catch fire on occasion. Especially if it has a bunch of GPUs pinning the utilization 24/7, or a cheap power supply in the cluster. =3
Are there any other works that should also be included that you know of?
Not OP, but a good resource for distributed consensus specifically is Tim Roughgarden's YouTube playlist "Foundations of Blockchains" [0] (hear me out, despite the title - see below). It's 85 videos over 12 lectures, rigorous, very well explained (but assumes some CS fundamentals). Not original work, but gives the conceptual framework and background so that one can read these classic papers in context.
Per-lecture reading lists are on the course page [1]. He also points at Elaine Shi's Foundations of Distributed Consensus and Blockchains [2] and Andrew Lewis-Pye's Consensus in 50 pages [2] as background, though these are more textbooks, not papers.
Remarkably, the first 7 lectures deal with permissioned systems, recapitulating the classic results of consensus in distributed systems (Dolev Strong, FLP impossibility, CAP) - no blockchain in sight. This takes us to the state of the art at the end of the 1990's (with algorithms that can achieve consensus in the presence of byzantine failures, namely Byzantine Paxos and PBFT, though he discusses a modern variant, permissioned Tendermint from 2014).
Lecture 8 stays permissioned and proves consistency and chain quality for the longest-chain rule. Only at Lecture 9, with proof of work, does anything specifically blockchain appear; then L10 block rewards and selfish mining, L11 transaction fee mechanism design; L12 proof-of-stake sybil resistance with the whole litany of attacks possible there.
Very good series in my view, and shows how little technical merit this whole blockchain circus has - nearly all the great properties people tout (reliability, consistency, audibility, availability) can be achieved with good old permissioned tech more efficiently, with pretty instant and deterministic finality.
Anyway, I found the series worth watching for the classical consensus material alone.
[0] https://www.youtube.com/playlist?list=PLEGCF-WLh2RLOHv_xUGLq...
[1] https://timroughgarden.github.io/fob21/
[2] https://elaineshi.com/docs/blockchain-book.pdf
[3] https://lewis-pye.com/2022/08/15/consensus-in-50-pages/
Does anyone use Elixir/Erlang anymore?
Yes, but it is boring telecom projects... so most people aren't interested.
Scala is kind of a more modern alternative.
If doing a twitter like platform, elixir with Phoenix channels could handle around >20k users per host. Very few other options work for that use case. =3
WhatsApp backend in Erlang + used in telecoms
The "WhatsApp used Erlang" story was a decade ago, doubt it is true now.
There's still a lot of WhatsApp Erlang tooling being published, so I'm pretty sure it's mostly still Erlang. I think I still know a couple people who would share with me if it was rewritten, although I left in 2019.
Erlang is simply the right language to build a chat server with, and it doesn't make sense to rewrite it in another language, even though Facebook did that with the server for Messenger. :P
Certainly parts of the WhatsApp stack have changed considerably. No more FreeBSD / no more only a handful of OS processes per node, much less Mnesia (not sure if there's any Mnesia left... I don't get that level of information). I've seen signs that they may have returned to large nodes, so they might be back to millions of connections per node, but that was very much not the case when I left; connection counts had dropped on our FreeBSD nodes because we were doing a lot more on the chat channel and then again because the Facebook nodes we migrated to were so much smaller than our FreeBSD nodes.
I doubt they changed out Erlang/Elixir, as it is mature backbone infrastructure. It has run many telecoms for over 30 years without issue.
The problem is finding talent that knows both distributed systems, and Erlang/Elixir. But Erlang/OTP can replace a lot of expensive kludges. =3
Some other classics (applied dist systems):
- DDB: https://www.allthingsdistributed.com/files/amazon-dynamo-sos...
- MapReduce: https://www.cs.princeton.edu/courses/archive/fall13/cos518/p...
- Spark/RDDs: https://www.usenix.org/system/files/conference/nsdi12/nsdi12...
- BigTable: https://static.googleusercontent.com/media/research.google.c...
Lamport is the author of more than half of these, it's crazy how influential he is in this space
In addition to making LaTex which has almost nothing to do with it
It's the thing that baffles me the most with those people, the Knuths and Lamports and Carmarcks, and whatnot.
It's not like they have light schedules - they do research, they have classes, they attendconference, and they have jobs, etc...
And some benefit from tenure, which give them time to deeply focus on some topics.
But I can't imagine how I would write tex in ten years in Pascal on archaic machines if I had all the time to myself.
Doing that in the middle of understanding more computer science than my eyes would ever read ?
What do they do with their time ? What do they don't do ? Are they able to skip the eating and pooping and sleeping and cleaning their house and picking up parcels at the post office and grocery shopping parts ?
Or a proto Unix being written in three weeks. Concretely, what are the days of those three weeks made of ? Especially since it's famously when the spouse was NOT there to handle the housekeeping...
I see not amount of "life hacks" and "atomic habits" and whatnot helping with that. But it's probably just as natural to them as biking - and you can't explain biking, I guess ?
A favorite Knuth story to illustrate: He was re-writing the 8080 code for the processor that ran our new Alphatype CRS Phototypesetter, since the manufacturer's code wasn't suitable. Real-time, 7 levels of simultaneous interrupt stuff, since it controlled two stepper motors that moved the lens around the photographic paper, plus four custom font outline-tracing S100 cards that flashed vertical segments of characters at just the right moment on the CRT that the lens then focused onto the paper. No debugability other than a 4x4 keyboard and a 15(?) character display. I was responsible for the code on the other end of the RS232 port that connected it to the DEC mainframe, sending down character outlines on the fly to the very limited memory on the 8080, along with the properly-sorted character placement info ultimately from TeX.
Anyway, we'd been doing morning, noon, and night debugging sessions together for a number of days. On maybe the 3rd morning, I arrived at his office and was speaking with his secretary of many years, Phyllis Winkler, when Knuth came in. He handed her a stack of maybe a dozen or two pages, saying "Here's a paper I wrote; please type it in." When? When did he write it? It's still a mystery to me.
I haven't updated this in quite a while, so probably more than a few broken links: my list of distributed systems reading lists.
https://gist.github.com/macintux/6227368
Obligatory shout out to Morris and Kaashoek's class at MIT - they update the reading list every year, and it's always great: https://pdos.csail.mit.edu/6.824/