Host: Computer science may well not be a growth industry going forward. This is Mike Stonebreaker. He's a Turing Award winner, famous for his fundamental contributions to database systems, like creating Postgres and more. What was the hardest part of that implementation? Query optimizer. It's just algorithmically difficult. How do you identify the people who aren't smart? Well, I mean, it's it's very easy. He shared interesting technical takes from his experience. On our benchmarks, large language models get zero percent. Why did you disagree so much with MapReduce? That wasn't the only thing Google was stupid about. I'm curious your thoughts on unsolved problems in databases and what you think the future might look like. Here's the full episode. The first thing I want to go over is the story of how Postgres got started. But for that, I kind of want to start the beginning. How did you get into building database systems?

Mike Stonebraker: When I graduated, I had the good fortune of being hired at Berkeley, and it was clear I had to, you know, continuing what I did for my PhD was not going to go anywhere. Then, as well as today, you're way ahead if you get adopted by a mentor who knows the ropes. So Jean Wong, who is still alive and still still kicking, took me under his wing and said, "Well, let's do something together." And this was 1971, which was the year after Ted Codd wrote his pioneering paper in in CACM. Jean Wong said, "Let's let's take a look at database stuff." And at the time, the the competitors were a thing called the Coddiesel proposal, which you're probably too young to have ever heard of. And so it was a low level spaghetti network proposal where you where you executed queries by following pointers. And then the the alternative was the IBM proposal, which was a higher a thing called IMS, which is still available. And it's hierarchical data. It's a tree. You're organized your data as trees. And even at the time, IBM realized that trees were not general enough to to solve many people's problems. So they hacked on a way to make it a limited network structure. So it was clear that was a horrible hack. The Coddiesel proposal had all kinds of bad properties besides being low level and and really hard to debug. It also had the property that if anything changed in your what's now called your schema, you basically had to throw away everything and do it all again because it was absolutely rooted at the physical level. Whereas Ted Codd stuff made perfect sense. And so Jean said, "Well, let's build one of these puppies. That's clearly the next thing to try." So we started building Ingres in 1972 while I was an assistant professor at Berkeley. As you know, if you're an assistant professor, you have to you have about you get five years to prove that you're a big shit, and they fire you or they give you tenure. So Ingres was my ticket to getting tenure, which happened in 1976. That was where it started. And then again, you know, happenstance. At the time, you know, a lot of people would build prototypes, which were sort of studenty like code, which means you could get it to run, but if you gave it to anybody else, they couldn't. So we put in the first 90% to get something we could run, and then for whatever reason, we put in the next 90% to get it to where it really worked. So the University of California version of of Ingres was really worked. And so over the next couple years, about a hundred universities started running it because Unix became the big thing, and so this was a database a free database system that ran on Unix, and so it was quite popular in in the academic world.

Mike Stonebraker: And so we got started getting you know lots of visitors at Berkeley who would say, "Gee, this is really nifty looking stuff. What's the biggest biggest Ingres application you have?" And we'd be forced to say, "Not very big." And so this was brought home in spades when Arizona State University considered running Ingres on their student records data, all 40,000 students worth, and they could get over that they had to get an unsupported operating system from Bell Labs. They could also get over they had to run an unsupported unsupported database system from these guys at Berkeley, but the project went down in flames when they realized there was no COBOL available for Unix, and they were a COBOL shop. So unsupported operating system, unsupported database system, no COBOL doomed us to you know irrelevance. And it was clear the only way out of that was to start a company. And so in 1980, we got venture capital as it existed then and started Ingres Corporation to move Ingres to to Dex VMS, you know a real a real operating system, and we had a real company that would support Ingres, and that was the start of the commercial journey.

Host: I saw that Ingres was competing with Larry Ellison's offering at Oracle. Yes. I saw that Ingres was was certainly better than what they were offering, but they were still competing somehow. How how did they compete?

Mike Stonebraker: Larry Ellison is a fabulous salesman, and he at the time he he would he made present tense and future tense indistinguishable, and so he basically lied to customers. He would ship stuff that didn't work and have his initial customers help him debug it. So I think he he engaged in what I consider very shady business practices, but lying to customers I think is is you know unconscionable. So for instance, there was a thing called referential integrity, which is if you if you fire an employee and he's the last person in a given department, do you want to delete the department or do you want to have it be a department a ghost department? It's it's all that kind of stuff. And so Ingres Corporation implemented referential integrity. Oracle Corporation wrote two manual pages that said, "Here's the definition of referential integrity," which everybody agreed to, and then he then down at the bottom it said, "Not yet implemented."

Host: Interesting. Yeah, I had interviewed someone who worked at Sun Microsystems, and they had a similar opinion that they Larry Ellison was a little bit shady. So it seems to be a commonality. I also saw somewhere else in something that you had said was that when Oracle acquired MySQL, that everyone kind of got afraid of that and moved to Postgres. That was the genesis of Postgres replacing MySQL as the preferred open source relational database system. So you you created Ingres, and there was a lot of technical innovations in it, so that it was better than the the incumbents, but ultimately it it went away, and you developed Postgres. What what was the thing that Ingres didn't do that Postgres would do?

Mike Stonebraker: The big thing that guided us at the very beginning was the the original reasoning for the academic version of Ingres was we were going to support a geographic information system that the neighboring professor Praveen Varia wanted. And so to support a GIS system, you need points, lines, polygons, line groups, that sort of stuff. And it was clear that Ingres couldn't do it because the data types we put into Ingres were the standard ones: integers, floats, text strings, and you couldn't support you couldn't efficiently support GIS types on top of that. So as a GIS, the academic version of Ingres was a complete failure, and that was in the back of of our mind. The other thing that happened this is a little out of chronicle chronological sequence, but it helps make the point is that the commercial version of Ingres, I think, around 1985, you know, there was ANSI had just proposed a date and time standard for relational databases, and so commercial Ingres implemented date and time, you know, using the standard Gregorian calendar. And so I was associated with the commercial version of Ingres as well as I was still at the University of California as a professor. So I got a call from from an Ingres customer who said, "You know, you implemented date and time wrong." And I said, "Huh? We implemented the Gregorian calendar, and you can subtract, and you know, if it has you know days have 30 or 31 months except for February, except for leap years. So subtraction on dates works exactly the way you would expect it to." But he said, "That's not what I want." In his particular world, he he said he was he was dealing with with bond financial instruments, and for some reason, I mean, you got the same amount of interest on on his financial bonds during each month, no matter how long the month was. So he had the date you bought the bond, the date you sold the bond. He wanted to do a subtraction, multiply it by the coupon rate, and say that's what that's the interest we paid you. But of course, his version of subtraction was March 15th minus February 15th is 30 days, because that's the definition of his calendar. And so he had to retrieve two dates out to user code, do the subtraction in user code, put the answer back, and it cost him a factor of two or three in efficiency. And he said, "Why can't I just overload your definition of subtraction with what I want?" And of course, with Ingres, it was hard coded, and the problem was this is a case where you wanted bond time, just like you wanted points, lines, and polygons. And so Postgres was engineered to have an extendable type system, so you could have whatever data types you wanted, and they were very efficient, and that was the main gist of Postgres was that it had that flexibility. And as you know, in business data processing, a lot most people were happy with the standard data types, but relational databases started to spread to all kinds of other places. What are called abstract data types or you know stored procedures, you know bunch of names they're called, you know had had great applicability, and so Postgres that was that was the big thing in Postgres. We also Postgres also supported what the AI guys at the time wanted in the way of inheritance. We also supported time travel, but the implementation absolutely sucked, and it got taken out after a while. So there were a huge number of of really nifty things in Postgres.

Host: You mentioned you you want to hire extraordinary software engineers, and I think you've you've said before that you have no trouble finding those people. How do you identify those people in your hiring that they're the extraordinary ones?

Mike Stonebraker: It's usually pretty obvious. I mean, I have a good feel for how difficult stuff is. If they get three x the the amount done, you know, in school that I think is reasonable, then then they're incredible. On the flip side, you had this interesting quote. I wrote I wrote it down. He said, "I can't stand people who are who aren't really smart. It's challenging to talk to them. How do you identify the people who aren't smart?"

Mike Stonebraker: Well, I mean, it's it's very easy. You talk to them, and and it it rapidly you can rapidly surface whether they're smart or not. You know, what was your master's thesis? What did you do? Well, how did it how did it exactly work? Well, how did you deal with error conditions? How many processes did you have? Why didn't you use threads? I mean, you you ask them technical deep technical questions.

Host: You you gave a talk, and I think there's also a paper behind it of this idea that one size fits all database systems not optimal. One size actually fits none, and that what you really want is database solutions that target specific needs. What database offerings you see today that are one size fits all?

Mike Stonebraker: In 2004, when I wrote the paper, we had an academic project which was building what became Streambase, and so a stream processing engine looks nothing like a relational database, and we had the gist of an idea for column stores for the for data warehouses, which was popularized by Vertica, looks nothing like a row store. So here were three wildly different implementations that had no resemblance to each other, and in each case, they were an order of magnitude faster than the other guys. So it's pretty clear that one side you know that in with those three instances, you give up an order of magnitude when you're running a database system that isn't that isn't architected for your kind of stuff. I think that's still true. I mean, I think ClickHouse is a column store, Pinecone is faster than user defined types on on text based vector processing, and so I think it's it's still very much the case, and I think there's no difficulty putting a common parser on top of multiple implementations. Postgres has so far chosen not to do that. They don't implement a column store, and so I think they are not they are not competitive, you know, on sizable data warehouses. They also don't have multi node support. Again, for people with big data warehouses, that's table stakes. So I think it's just as true today as it ever was. I think that what is true is that if you want to get going, you have a database problem. You know, the answer is choose Postgres, and there's a huge programming community, all kinds of all kinds of you know data type implementations. It's free, and you can find Postgres talent easily and get going. And and so I think it's it's it's a great choice for lowest common denominator. And until you're trying to do a million transactions a second, it works just fine. Until you're trying to support a petabyte data warehouse, it it I say the low end, it's absolutely the right one size fits all. At the low end, it's Postgres. At the high end, that's just not true.

Host: GPUs do they make available some new opportunities to optimize databases?

Mike Stonebraker: Probably, but I think the the big challenge is that GPUs are you know SIMD SIMD you know single instruction multi data, and that's that's the anathema of indexing. And so whenever indexing is the right answer, they're probably not a good idea. And I think also you've got to architect them so that the so that the bandwidth so that the bandwidth from storage is is not not the bottleneck, and so. If they're an add-on to the CPU, as often as not, the bus connecting it to the GPU to the CPU is a bottleneck.

Host: Can you explain why indexing would be not as effective when there's SIMD?

Mike Stonebraker: So let's let's say I'm I'm looking for Ryan's I'm looking for Ryan's salary and I have a B-tree. So you go to the root of the B-tree. You find you find the divider that has both sides of Ryan. You follow the pointer. That's a memory access for sure. Then you do it all again. And you do this like three or four times. So that doesn't parallelize well. So the answer is indexing doesn't parallelize well.

Host: You mentioned B-trees. When you first implemented that first version of Ingress, did you write all of that by hand? Because I imagine there's probably not some existing B-tree library or something.

Mike Stonebraker: Yeah, we wrote the original version of Ingress was all written from scratch. What was the hardest part of that implementation? The query optimizer. And why was that hard? It's tough. It's it's just algorithmically difficult. It's still if you ask most any senior database programmer what's the hardest hardest part, they'll still say the optimizer.

Host: MapReduce came out at some point in the early 2000s, and it kind of took the data world by storm. People were really impressed by it. They thought Google really knows what they're doing. This is the best thing since sliced bread. But it seems like when I look at the literature and what you thought at the time, you kind of disagreed heavily. Why did you disagree so much with MapReduce?

Mike Stonebraker: Well, I think there were a lot of not very enlightened people who said Google Google is really smart. They must know what they're doing. And so we'll do whatever they say. And so they would they would they would engage in Hadoop or engage with Hadoop. But Hadoop is ridiculously inefficient. And so at the time, you know others, you know Dave DeWitt and others who who were involved in our 2011 paper, we understood distributed databases and understood that you could beat the heck out of Hadoop with a distributed database system, which is basically what that 2011 paper says. And of course it was it's true. And but that wasn't the only that wasn't the only thing Google was stupid about. So Google also had the opinion that eventual consistency was the right way to do concurrency control. And so that was postulated from on high by Google all during that same period of time. And it wasn't and all the database people said you know you're out of your frigging mind because it doesn't it solves one particular kind of problem but only and that very rarely occurs in practice.

Host: Why did they pursue eventual consistency?

Mike Stonebraker: Okay, well the idea is that you have an East Coast database and a West Coast database and they're replicas. So you want them to be the same. If you say I'm going to do a transaction, I'm going to decrement by one the number of widgets in the West Coast warehouse. Then I'm going to before I commit that transaction, I'm going to update the East Coast warehouse, pay pay a message over and back to update it. And then to make sure everything goes well, it takes it takes another round trip a message to make sure that both of them actually do the commit correctly. So it's expensive to do a distributed commit. It still is. And so the idea was well you you do the you do the West Coast update, you decrease the widgets by one, you just send a message asynchronously and not in a transaction. So that eventually the East Coast warehouse gets decremented by one. So meanwhile, if you're on the East Coast, you you decrement you know foodstuffs by one, you send an asynchronous message. Eventually the West Coast gets it. And eventually everything settles out. So if you're allowed to to go below zero, then what will happen is if the East Coast guy and the West Coast guy simultaneously sell the last widget, then then eventually the state of the warehouse will be minus one and somebody won't get their widget their widget. And so if if you're allowed like Amazon to say usually ships in 24 hours, then maybe you're kind of allowed to oversell, but most enterprises can't do that. And so eventual consistency just doesn't work. So we talked a million hours ago about referential integrity. So referential integrity in a sales system is integrity constraint is stock is greater than minus one. And that fails with eventual consistency. And so Jeff Dean of Google finally figured that out. And when they did Spanner, Spanner had a conventional transactional system. And so Google completely abandoned eventual consistency and completely abandoned MapReduce. So the trade-offs basically correctness or performance. So it's performance versus data integrity. And if you don't care about your data, then you're willing to deal with with bad things happening.

Host: You've influenced industry significantly from academia, and my one thought that I had is what why not work directly in industry or why why do you prefer the position of being in academia and having influence in the way that you have versus just taking a job at AWS or something like that, being a very distinguished engineer there?

Mike Stonebraker: Because that gives you a boss, and that gives you company rules, limits your ability to publish, limits your ability to go talk at conferences, limits your ability your ability to go go poke at what what various competitors are doing that they won't tell their competitors. But mostly, I really like being in startups, and I and I after the commercial version of Postgres got acquired by Informix, you know, I was working part time for Informix, which was a 2,000 person company, and I didn't feel like I could make a difference because it was bureaucratic and and whatever the president wanted, he got. So I think I'm just not cut out for I'm not cut out for politics. I don't do that very well, and I have a hard time interacting with people I think are dumb. And that again, so I guess I I have I have some problems with with big companies.

Host: I I wanted to talk a little bit about dbos. I just thought it was a really interesting technical model. Can you explain what dbos is?

Mike Stonebraker: We started the academic project in 2019, 2020, something like that. And the gist of it was at that point Mate Mate Zaharia, who was on the faculty at Stanford, was also one of the founders of Databricks, was the original creator of Spark. And so he said at the time Databricks you know basically was running people's Spark jobs on the cloud. And so he said at any given time we might be orchestrating a million Spark jobs. And so we have to write a scheduler that's going to decide who to run next at scale a million. And he said there was no we tried all the all the schedulers written by the OS folks and they they couldn't they didn't scale. So we put all the scheduling data in a Postgres database, and basically a Postgres application was doing scheduling. And then it sort of clicked that by and large most everything you do in an operating system is managing data at scale, and you should do that using database technology. So why don't we just replace at least the upper half of Linux with a database system? So that was the gist of the academic project, and we worked on it at Berkeley and Stanford in the early early 20s, and it was it was very successful. It clearly it clearly worked. And in the process, the Stanford folks wrote an extension to JavaScript so that you could program. You need some programming world that can can talk to your implementation. So if you're doing what amounts to a programming language, and you're running on top of what amounts to an operating system that is a database, then the obvious thing to do is put all your state in the database, and that's exactly what they did. And so we had an innovative programming language model, an innovative operating system model, and and of course then the idea was well can we start a company? And so we talked to the VCs who to a person said you're you're dreaming if you think you're going to displace Linux. However, this programming language stuff is really nifty. We had what amounted to extensions to JavaScript that would allow any any program to have all of the nice features of a database system. You know stuff was durable. You could have transactions. If it failed, you'd fail over. You know it's all that nifty stuff. So we got funded to start a company in 2023, and that was dbos incorporated.

Mike Stonebraker: And I think the key thing so far is probably two thirds of the customers are doing agentic AI, which means that they have a large language model surrounded by a bunch of stuff that that adds more signal. And so far, most of agentic AI is read only, meaning you want to produce a prediction for whether Ryan is is going to be a good customer or not. And so that just runs some stuff and then produces a new thing that's given to somebody. So basically read only, which means that you're not you're not actually updating Ryan's credit rating, or and so I think fairly quickly this the whole world is going to move to using you know agents to do read write applications, and that's going to make that's going to make them very very databasey, and dbos does that stuff really really well. And so you know for instance if you want to write an agent or two agents that move $100 from my account to your account, and so you debit my account, you increment your account, and these two agents have to agree to commit, or you have to back everything out, which is to say the workflow needs to be what I call atomic, which is it all happens or it looks like it never happened.

Host: I'm curious your thoughts on unsolved problems in databases and what you think the future might look like.

Mike Stonebraker: Okay, so I think two different things that I'd like to talk about. The first one is like everyone else, three years ago we started to look at what were large language models good for. So we've been trying to get what's now called text to SQL. To we've been we've been trying to make it work on real world databases, especially real world data warehouses. So we've been trying the technology on four different production databases warehouses where we've gotten the workload the actual workload that's run and you know from the actual users using the system, and we've gotten them to reverse engineer the text that corresponds to that SQL. So we have text and SQL for we have four benchmarks. [...] Well, on our benchmarks, large language models get zero percent, and if you enhance them with rag and and all the tricks goes to 10 percent, and if you give as a prompt the from clause, in other words, all the actual tables that need to be accessed and all the actual join clauses that need to be joined, then accuracy goes to about 35 percent. So the definition of this stuff doesn't is not ready for prime time and not going to be for a while if ever. So what what's the difference? Number one data warehouse you know LLMs are trained on the pile. Data warehouse data is not in the pile... Number two, query complexity... Number three, the schema in Spider and Bird is clean... In data warehouses... column names are often underscore z underscore blah... So we've been so what do you do? So well, first of all, we published our benchmark. It's a thing called Beaver, which is an anonymized and abstracted version of these four data warehouses. And so if you think you're really good at doing text to SQL, try a real benchmark, not a fake one.

Host: If you could go back to yourself when you just graduated, give yourself some advice, knowing what you know today, what would you say?

Mike Stonebraker: Back when I first took the job at Berkeley, without thinking about it much, we said let's write a database system, and we we knew nothing about databases, nothing about implementations. We were not skilled programmers like Bill Joy. So starting off doing something that was that crazy was really pretty crazy. [...] the better question is if you were starting out today, what would you major in? Because I think you know computer science may well not be a growth industry going forward, and I'm not sure I would recommend 18 year olds to major in computer science. I mean I think health healthcare and and the building trades are are safe bets, and everything else looks much riskier. [...] Both my wife and I said follow follow your passion. Somehow the money will work out, and I don't believe that for a minute. But I think that's what you have to tell your kids. [...] find something you're passionate about and and you will you know either you won't starve. You may not make a lot of money, but I think chances are you'll be happier than if you do something you're not passionate about.