Irv Traiger: Leonard had ordered all of us to pick a name for this project. We just sort of shrugged off, "It's not important." He said, "It's important in terms of recognition to have a name." We would make attempts at coming up with a name over weeks. One was Rufus, which was Franco's dog.
Franco Putzolu: Rufus would have been a better name. It stands for Relational User Friendly Universal System.
Mike Blasgen: It would have been a better name.
C. Mohan: Later we actually had a project named Rufus. Kurt Shoens' ...
Irv Traiger: It was really hard.
Mike Blasgen: So it was named roughly at the end of 1974?
Irv Traiger: Don't remember.
Tom Price: Was that the time that Leonard made you guys all work on Christmas Eve? I heard a story once that he wouldn't let anybody off on Christmas Eve?
Irv Traiger: I think that was back in Yorktown.
Don Chamberlin: That was in Yorktown; yeah, I remember that. [laughter] This was the Friday before Christmas and the lab had some kind of a party with cookies and Santa Claus and music and everything down in the cafeteria and Leonard wanted to have some kind of technical meeting right through the whole thing. Leonard expected a lot of his people, but he also treated them well.
Mike Blasgen: Leonard was quite a character: a lot of fire and brimstone and vim and vigor and all those pairs of words. I remember probably in 1975 we went off to the beach at Pajaro Dunes and Leonard stood up and said, "OK, what are all the bad things that are going on in the department? What are all the bad things I'm doing?" And he made everybody say them. Everybody complained. And he wrote down this list of complaints. He didn't say anything. He just wrote down complaints. And then he said, "OK, shut up," and he talked for two hours without a break telling us, basically, everything we were complaining about was not correct. [laughter]
Management by consensus: I have decided; you concede. [laughter] It was so amazing; it was completely oblivious to him that he was doing this. It worked; it worked very well for him. In case you don't know, he's the Chief Operating Officer of Cadence. Cadence's number one customer is IBM. They sell electronic design tools for laying out circuits on chips.
By the way, this System R thing of course makes me put this [cartoon] up. I don't know when this picture was drawn; this is my favorite chart. This is a rabbit and a beaver talking, and behind them you can see Hoover Dam. The beaver is saying to the rabbit, "I didn't actually build it, but it was based on my idea." [laughter] So this little beaver is System R, because I don't think there is much code of System R left around; a little bit in SQL/DS I guess.
C. Mohan: Quite a bit, actually, especially the RSS.
Mike Blasgen: All right, the index component is still alive. [laughter] That's what I wrote, and the index component is still in the product, SQL/DS.
C. Mohan: All the shadow-paging stuff is there.
Mike Blasgen: Oh, the shadow page's still there? That's Raymond Lorie's stuff.
C. Mohan: Record management, all that stuff's still there.
Bruce Lindsay: Storage pool.
Brad Wade: Like to know if anybody can still understand it.
Pat Selinger: Mohan still reads it.
Mike Blasgen: You don't have to understand it; it just has to produce revenue and profit. It's a successful product today.
???: It supports a lot of us.
Mike Blasgen: Right. So ...
Brad Wade: Before we leave naming, there was also the RDS and the RSS names. Of course Don was manager of the RDS before it was called RDS; Irv was manager of the RSS before it was called RSS. And they were carpooling and they came in one day and said, "OK, here are your names: Don and Irv, Data Organized Naturally, and I forget what Irv was for: Intermediate or Interactive Relational ...
Mike Blasgen: Intermediate Retrieval Vehicle? How about that? Sounds good. No, there was the Peterlee Relational Test Vehicle, so V was already established as an acceptable term in Relational terminology. So it's just a question of putting the Vehicle in there somewhere.
So how about what sort of happened with System R. Irv and Don were the managers of the project. Why don't one of you volunteer to take us through the System R history?
Don Chamberlin: I think it's going to need both of us to do this. I'll give it a start.
This shouldn't be a monologue; please stand up and help me out here. As Irv said, there was a long period after Frank arrived in California when we had a lot of meetings and a lot of discussions and task forces and tried to organize an approach to take to this business. Interestingly enough, Ted Codd didn't participate in that as much as you might expect. He got off into natural language processing and wrote a very large APL program called Rendezvous, . He really didn't get involved in the nuts and bolts of System R very much. I think he may have wanted to maintain a certain distance from it in case we didn't get it right. Which I think he would probably say we didn't.
Mike Blasgen: Oh, he has said that, many times.
Don Chamberlin: What came out of this was we got organized into two groups, a higher-level group which ultimately was called the RDS and which was interested mainly in language issues, and a lower-level group called the Research Storage System, which was interested more in physical data management issues. I can talk mainly about what was happening in the top half of the project in those days and I'm hoping that Irv and maybe some of the rest of you - Jim - will talk about what was happening in the bottom half.
What really happened in the early days was Irv's group began developing a new data management interface, with support for indexes, locking, logging, concurrency and transactions, and all those kinds of things. Meanwhile the language folks wanted to build a prototype of their language and they needed a base to build it on, and the RSS wasn't ready. The only thing we could get our hands on was something that Raymond Lorie had built at the Cambridge Scientific Center called XRM. So we built a prototype of our language on top of XRM in the early days; we called it Phase Zero 27. Brad has a wonderful tape which many of you saw last night that represents a complete working prototype of SEQUEL in 1976 I believe, complete with integrity assertions, which have just now made it into the product twenty years later. [laughter] And we demonstrated that, or at least showed the tape, at the SIGMOD conference in, was it 1976?
Brad Wade: 1976.
Don Chamberlin: Hopefully today we'll get a chance to see that tape again. It's a wonderful tape; you get to see Brad with a handlebar mustache. Good stuff.
Franco Putzolu: Don, did you have a customer in New England?
Don Chamberlin: Yes, as a matter of fact, that was the most important outcome of the Phase Zero work, I think in my opinion. That's a kind of interesting story. Back in those days, there were a lot of problems with fuel shortages; OPEC had just raised the price of oil and the gasoline companies were hoarding it and there were lines at the gas stations. The MIT Sloan School of Management had some kind of a plan in New England where they got a grant to build something called the New England Energy Management Information System, or NEMIS, and they needed a database to keep track of how full the oil tanks were and things like that. So the Cambridge Scientific Center was kind of tight with San Jose Research, and they got their hands on this Phase Zero prototype and worked on it with the Sloan School of Management on this energy management system, but anyway, one of the students at MIT who was involved with this was somebody named Bob Selinger. And Bob, didn't you kind of get your fingers into Phase Zero and use it a little bit for something? As a result of this, Bob came out to San Jose as a summer student, because of the experience that he'd had with the Phase Zero prototype. When he came to San Jose, he met someone named Pat Griffiths. That's how Bob came to IBM.
So I think the most important outcome of the Phase Zero prototype was ... [laughter]
Pat Selinger: Did the energy management system ever get used? [???]
Bob Selinger: There were databases on it. I'm not sure they were widely used. Actually they used it as a database for building designs. They kept track of square footage, number of windows, and then they had some FORTRAN programs that ran on top of it. It bridged FORTRAN into, I think PL/1, to extract the data. It was pretty hokey.
Don Chamberlin: So what this language group wanted to do when we first got organized: we had started from this background of SQUARE, but we weren't very satisfied with it for several reasons. First of all, you couldn't type it on a keyboard because it had a lot of funny subscripts in it. So we began saying we'll adapt the SQUARE ideas to a more English keyword approach which is easier to type, because it was based on English structures. We called it Structured English Query Language and used the acronym SEQUEL for it. And we got to working on building a SEQUEL prototype on top of Raymond Lorie's access method called XRM.
At the time, we wanted to find out if this syntax was good for anything or not, so we had a linguist on our staff, for reasons that are kind of obscure. Her name was Phyllis Reisner, and what she liked to do was human-factors experiments. So she went down to San Jose State and recruited a bunch of San Jose State students to teach them the SEQUEL language and see if they could learn it. She did this for several months and wrote a paper about it, and gained recognition in the human-factors community for her work., 31 I'm not sure if the results were very conclusive; it turned out that sure enough if you worked hard enough, you could teach SEQUEL to college students. [laughter] Most of the mistakes they made didn't really have anything to do with syntax. They made lots of mistakes - they wouldn't capitalize correctly, and things like that.
Looking back on it, I don't think the problem we thought we were solving was where we had the most impact. What we thought we were doing was making it possible for non-programmers to interact with databases. We thought that this was going to open up access to data to a whole new class of people who could do things that were never possible before because they didn't know how to program. This was before the days of graphical user interfaces which ultimately did make that sort of a revolution, and we didn't know anything about that, and so I don't think we impacted the world as much as we hoped we were going to in terms of making data accessible to non-programmers. It kind of took Apple to do that. The problem that we didn't think we were working on at all - at least, we didn't pay any attention to it - was how to embed query languages into host languages, or how to make a language that would serve as an interchange medium between different systems - those are the ways in which SQL ultimately turned out to be very successful, rather than as an end-user language for ad hoc users. So I think the problem that we solved wasn't really the problem that we thought we were solving at the time.
Anyway, we were working on this language, and we adapted it from SQUARE and turned it into English and then we started adding a bunch of things to it like GROUP BY that didn't really come out of the SQUARE heritage at all. So you couldn't really say it had much to do with SQUARE before we were done. Ray and I wrote some papers about this language in 1974. We wrote two papers: one on SEQUEL/DML and one on SEQUEL/DDL. We were cooperating very closely on this. The DML paper's authors were Chamberlin and Boyce; the DDL paper's authors were Boyce and Chamberlin, for no special reason; we just sort of split it up. We wanted to go to Stockholm that year because it was the year of the IFIP Congress in Stockholm. I had a ticket to Stockholm because of some work I'd done in Yorktown, so Ray submitted the DDL paper to the IFIP Congress in Stockholm, and the DML paper we submitted to SIGMOD. This is the cover page of the SEQUEL/DML paper. It was 24 pages long. These were twin papers in our original estimation. We wrote them together and thought they were of comparable value and impact. But what happened to them was quite different. The DDL paper got rejected by the IFIP Congress; Ray didn't get to go to Stockholm. I still have that paper in my drawer; it's never been published. The DML paper did get accepted at SIGMOD. Several years later I got a call from a guy named Larry Ellison who'd read that paper; he basically used some of the ideas from that paper to good advantage. [laughter] The latest incarnation of these ideas is longer than 24 pages long; it's the ISO standard for the SQL language, which was just described last week at SIGMOD by Nelson Mattos. It's now about 1600 pages.
Jim Gray: It's two large binders over there [on the artifact table].
Mario Schkolnick: Don, I remember you used to tell that Larry Ellison had called you and asked for the error codes; what error codes would IBM be using? He wanted to be compatible.
Don Chamberlin: Larry called up. Larry's company in those days was not called Oracle. His company's gone through two changes of name. The original name was Software Development Laboratories. He had heard about the System R prototype and he wanted to make sure that his product was fully compatible with it, right down to the error code values. We went and asked Frank, "Can we give our error codes to this guy Ellison and Frank said, "No - those are IBM Confidential."
Franco Putzolu: That was the only part that was confidential.
Mike Blasgen: You know that whole thing is sort of interesting. When we submitted the TODS paper, one of the referees said that we ought to include the SEQUEL BNF, which we did, but it wasn't in the paper that we originally submitted. Its inclusion was insisted on by a reviewer and demanded by the editor, and so we put it in even though we thought it was kind of ... whatever. I think the common wisdom in the world for many years was that we shouldn't have done that; we should not have put it in because that was sort of too much detail, made it too easy for copycats to copy it. I'm not sure this is correct, but ...
Jim Gray: What was it that you put in?
Mike Blasgen: BNF - the syntax. No, the semantics were in the paper; that wasn't changed - we always described it. But somehow the details of the syntax ... Leonard, for example, many years later felt that was a big mistake; we never should have done it.
Franco Putzolu: Later on I thought that publishing everything was a big mistake.
Josephine Cheng: Only you should have patented it before you published.
Mike Blasgen: People should know that patents were basically prohibited. Patents at this time were prohibited by the company and the Supreme Court. Software patents.
Franco Putzolu: I remember until 1979 we were publishing everything that would come to our mind, either implemented or not implemented, or dreamed of; and then all of a sudden there was a barrier.
Mike Blasgen: Right, somehow we decided maybe we could make some money out of this thing. Actually, that's a compliment, right?
We put out a big press release in 1975 or so associated with the kicking off of this. And that was suppressed by GPD. They wouldn't let us put out the press release; do you remember that?
Irv Traiger: I don't.
Mike Blasgen: We had a bunch of paper work. Actually Sharon got involved. [laughter] My wife was the lawyer and she helped them suppress it.
Bob Yost: Do you think this would have been anywhere near as successful if IBM had just held it inside? I don't think so. I don't think it would have gone anywhere near as far.
Franco Putzolu: Well I think the critical thing was the fact that it was adopted by SQL/DS and DB2 - not that much that it was popular in universities.
Mike Blasgen: I used to talk a lot about this. I was kind of a spokesman for System R for a long time and a lot of people inside IBM asked that question. My answer was exactly what Yost said, which is that if we had not published those papers it would have failed. Now the reason it would have failed is that IBM would have ignored it.
Mike Blasgen: No, it's clear that if you could change history and not publish all those papers and know that you were getting SQL/DS and DB2 out, then we would have been better off not to have published the early papers. But I'm convinced that the only reason that anybody cared ... well, Jolls will say something maybe about this. Actually it's too early for your time; your time will come. But I'm convinced that publication was the right thing to do. I know a lot about this because I worked on RISC. I was the manager of the 801 project, too. The 801 project did not publish anything, and it was much harder to get it out. It was much harder to get IBM to do something about it. We had to transfer it to Sun. SPARC was the first highly popular RISC, and it was only after Sun went to RISC that we could wake IBM up to the opportunity here.
Tom Price: Was it only after Ellison started doing Oracle that DB2 ...
Mike Blasgen: No, Ellison was not a factor in SQL/DS and I don't know about DB2.
C. Mohan: No, I was told that SQL/DS came out after Oracle came out.
Mike Blasgen: Oh, that's true, but that just shows how long it takes IBM to do something.
Irv Traiger: So thinking back to the task force days of System R, which wasn't named System R yet, there was this notion of getting the Phase Zero prototype going, that Don talked about. It was understood that GAMMA-0 and XRM and other systems might not be the right platform. They all had a funny characteristic - all of them. None of them stored the values in the tuple. They all stored 32-bit things that would point to the values. This was in the days of small disks and small memory. The concept was that if somebody was a programmer or lived in Poughkeepsie, you didn't want to have to store "programmer" or "Poughkeepsie" more than once. You'd have these classes of names of things, like names of cities, or names of job titles, or things like that - people's names. You'd store pointers to an element in that class for these variable-length strings. All of them did this; all of them. RAM was binary; a tuple-id, and a pointer to a thing, and a pointer to another thing. If it fit, great, but very, very few things fit. It became clear pretty early that, what if you're just going after one tuple? You know, "Tell me about Mohan, in the Employee file". The overhead would be incredible because you'd be chasing this pointer and chasing that pointer, so why not just store the stuff right there, which was being done anyway in VSAM and IMS and DBTG.
So we came to that realization pretty quickly, and then, again in task force mode, which can kind of wear you down after a while, we came to this other notion of an intermediate level called the SLI: the System Logical Interface. This would be set-oriented query, but I think only on one index and one field and one table. Somehow SEQUEL would translate down to these smaller set-oriented things and paste together complex queries. This idea was something that my group, which was just getting going, was going to work on while Don and Ray and Paul Fehder and Morton Astrahan worked on the Phase Zero prototype. But none of us really liked this SLI thing, so that kind of petered out.
Something else was going on around then that helped it peter out: I got a kind of co-conspirator, Franco. Franco was brought out from Yorktown as part of this Leonard Liu package deal with Gomory, on who would come out. He was not supposed to work on this System R stuff. Ed Altman was one of the principals in the Mike Senko department on the DIAM project, and he was becoming a second-line of various other groups in Computer Science. I think Franco was brought out to do a physical database design tool with Altman ...
Franco Putzolu: Yes, something like that; I never figured it out.
Irv Traiger: ... maybe C.P. Wang was heading it, who had come from that Senko group. It was very delicate how Leonard would balance the skills across the Altman bunch and the Frank King bunch, because he really didn't want to look like he was taking advantage of the old DIAM people and favoring Frank King. So some of the strong people who came out were directed to the Altman side, including Franco. One afternoon, Leonard said to me I should go talk to Franco, and I didn't know why he wanted me to do this; he was just being kind of coy. And it was clear that Franco was a very, very perceptive guy. He understood what database people were doing back then and he really cared about applications. He would read these weird little papers on ...
Franco Putzolu: I studied IMS, among other things; I even installed IMS in Yorktown once.
Irv Traiger: So it occurred to me that this SLI thing really was a bad idea, and we just needed somebody with a bit more practical insight, so I talked to Frank King and said we ought to get Franco. But Frank King didn't want to touch this situation because of this balance of power thing. We somehow made it happen.
Kapali Eswaran was hired in around this time, and I believe he was maybe reporting to me, but helping these folks in the Phase Zero prototype, putting in consistency constraints and triggers, which as noted before have only recently made it into the IBM product line. There were other things going on, too. We were working on concurrency, trying to add that, because none of these early systems had concurrency. If they did, it was by accident. I had done some early stuff in GAMMA-0, Jim Gray was very interested and he was doing some things, and Raymond came up with this gleam of this idea of what became called predicate locks, where since you're querying sets, why not lock sets: the most natural thing you could imagine. And that was consistent with what we had finally after a struggle figured out about authorization and views. Instead of authorizing columns of a table, just make it a view of those columns and authorize that. Kapali heard about some of these predicate things, and he went off and worked on predicate locking as well, and we began to also understand that transactions were like logical units, like all or nothing.
Bruce Lindsay: There's a great line in this paper I found in Jim's box about predicate locking. Little paragraph says, "The overhead and complexity of constructing the predicate, testing the predicate, and scheduling the predicate terrifies both Morton and Franco. It merely scares the rest of us." [laughter]
Irv Traiger: There was one short period where we thought that predicate locks were the right approach and, although we weren't saying it, that would give you this notion of a serial schedule, you know, logically equivalent to a single-user system. But it wasn't real crisp yet. I remember another afternoon, I was sitting I think in Jim Gray's bean-bag chair. He had this small office and this huge bean-bag chair, and a regular chair. We were talking about what all of this meant. Marc Auslander was visiting you that day. He was just kind of wandering around, sort of looking over our shoulder, and suddenly Jim began to better understand what serial schedule meant, why it was important, and why maybe predicate locks had nothing to do with that. But they sort of helped us to get there. He referred to a paper by Donovan that had to do with something like serializability on cache operations. This was kind of a stretch, but that's where consistency and serializability happened, as I recall. It was there. Does that ring a bell at all, Jim?
Jim Gray: Yes. Also Karp & Miller's asynchronous communication. They were interested in determinacy and we were just looking for consistency. We wanted an answer, they wanted the right answer.
Irv Traiger: That's where that happened. Versioning was ... shadows were a very strong notion in XRM, as I recall, which Raymond brought to us from Cambridge. He had transferred out pretty early in the cycle. So is there anything else I can think of? Recovery, logging, we wanted to go for tuple-level locking, so we realized that shadows weren't going to cut it, because they were at a page level, so we introduced this intricate combination of record-level locking and logging and shadows, which actually works.
Franco Putzolu: Kind of. [laughter]
Irv Traiger: One of the very strong notions is that we wanted to support all the data models, so the RSS had links, basically pointer chains, and the idea was to support hierarchies, networks, and relational. Over time we gave up on this noble ambition, but it got resurrected again at Santa Teresa. It was very interesting, it was more this not knowing where data was going to go, wanting to build the universal low-level thing, whatever might be on top of you.
Franco Putzolu: Yeah, I would say that there were two really important points. One is that RSS just existed. We accept as a given that you have to split the system into a low-level component and a high-level component. But only the systems with System R ancestry have this clear split. Many of the other systems don't, and I think suffer from it.
Irv Traiger: Yes, it wasn't clear that you should split it, or how to split it. As I mentioned, we struggled ourselves with this SLI thing that was a kind of medium level.
Franco Putzolu: The split was done at the right level. And the other major point was the emphasis that multiple high-level subsystems would use the common low-level engine. This approach was tried by all the systems that came after System R; all these attempts failed.
Pat Selinger: I remember Franco that you had led at least one study that I recall on how to map IMS - every construct: logical deletes, and sparse indexing, and all kinds of different things - all into the RSS level.
Franco Putzolu: Yes, what a waste. [laughter]
John Nauman: But you kept doing it, Franco.
Jim Gray: So somebody can appreciate that all this was happening in the background of something called FS - Future System - and the people in FS thought that it was an incredible waste of research energy to be working on this relational database stuff. They were working on GRID or some project that was much more sophisticated and advanced than anything we were doing - it had a GUI and was wonderful.
John Nauman: Actually, it was three systems: it was supposed to support relational, and the network CODASYL stuff, and flat files - it was going do everything. But that was all dropped.
C. Mohan: But part of that came in System/38, right? They allowed file system access as well as high-level query access to the same data, which is still there in AS/400.
various: Astrahan. [laughter]
Mike Blasgen: To my pleasure, Blasgen has two, so I'm not in the noise. Chamberlin has: "Data Base System Authorizations" in Foundations of Secure Computing. It's probably not even on his CV. Anyway, that's Chamberlin, Gray, Griffiths, Mresse, Traiger, and Wade. But the answer of course is Ron Fagin, with about ten publications. The theory guys always beat us systems guys. And Frank King wrote "The phonetic encoding of word-components for the computer input of Chinese characters". It's his only publication in here. So I'll pass this around and people can look at it. This is really a great little thing. It's got a lot of good stuff in it.
We're back on track. Unless there's some opposition, at least part of this afternoon's session is going to be devoted to more of this pre-1982 material. Any comments or suggestions? Because we're not done; in fact we haven't mentioned the joint studies; we haven't mentioned the interaction with Santa Teresa. And if we're going to lunch at noon, we only have 40 minutes. We have videotapes that we're going to show you. We'll show you a little snapshot of those just before we go to lunch. OK? Some of you may have seen little snapshots last night. We have videotapes from the era, from 1979, so you can see what you looked like if you were on tape.
This [begins showing System R slides] was the kind of talk that went around at the time on System R. Ease of use. See - we cover all the bases; everything important is there, right? Did we miss anything? We got PL/1, we got VM and MVS; there were no other operating systems of importance ...
Mario Schkolnick: And it's all under ease of use.
Mike Blasgen: So of course we have all the people who make more than their manager. We only had one database. Here's our advertisement for SEQUEL. Here's "Find the names and salaries of employees who make more than their manager" - right there, number four or something. "Give a $1000 raise to all programmers in Department 50" - it used to be K55, is that what it was?
C. Mohan: It's still K55.
Mike Blasgen: We compile. Now one of the things I want to talk about is Raymond's discovery of compilation. He gave a talk in the cafeteria conference room in Building 28 once where he said we can compile and it will run faster. I remember that meeting very well because we basically threw out a huge pile of code and started all over again because we realized we were doing it wrong. And we had to change all our presentations. Like the TODS paper is all wrong, as an example, because we changed the way it worked.
Here's an application program. This is what SEQUEL looked like, in case people don't remember. The dollar signs were required because we had a scanner that scanned PL/1 and we didn't want to fully parse PL/1 so we just wanted to look for illegal delimiters. So we just tokenized it and looked for illegal symbols. The dollar sign is illegal in PL/1; you can't say $DECL. This is "Let a cursor be"; OPEN; FETCH; System R codes; the codes of course are what Ellison wanted. There's the same picture as you saw in the overheads. MVS still in parentheses. That dates it, because MVS came out of parentheses in 1979. Execution. Oh yes, so we load the ... Performance ... More nuts. Oh, transactions. Why did this project turn into a research project on transactions, and locking, and logging? Actually the fundamental contributions of the work turns out to be the stuff that hardly appears in the papers, which is all the work that Jim and his colleagues did.
Oh, here's a history of INGRES, PRTV, QBE, and all the System R ... here's R* ... so that dates this. This is actually a presentation I gave in Atlanta at SHARE right after the SQL/DS announcement. So here System R begins; TODS paper in 1976; first install - that was at Pratt & Whitney in 1977. First System R install was 1977. So here's the definition of all the papers that we wrote. We called it full function, but hardly the full 1600 pages that is required today.
Jim Gray: Oh, I like that slide.
Mike Blasgen: ... Oh, here's the "Indian's salary's greater than the chief's salary". NULL. Support for NULL. I remember going to talk to Ted Codd about support for NULL. [laughter] We couldn't get together - the RSS and RDS couldn't get together on that, so finally the RSS said, "You guys decide; we'll support whatever." We supported a general NULL.
Franco Putzolu: I didn't like NULLs in those days.
Mike Blasgen: Right, you sure didn't. But we supported them; RSS supported NULLs, in a particular way. We allowed you to specify a NULL symbol, and then we would insert that wherever it was.
Jim Mehl: I still remember a memo that Franco wrote addressed to "NULL theologians". [laughter]
Mike Blasgen: The non-nerds in the room don't even know what we're talking about. If you don't know somebody's age ...
Bruce Lindsay: I just found "Revisions to the NULL Memo".
Mike Blasgen: There were great controversies about NULL. Just to give you an example, if you don't know somebody's age, but you know a lot of other things about them and you want to record it in a database but you don't want to put in their age. So you leave it blank, but now you sort, say, on age. Where do you want that person's age to come out: at the beginning or the end or the middle? This is NULL theology. There are the people who want to eat the egg from the big end and people who want to eat it from the little end. And then there are people who want to eat it from the middle.
Oh, gee, views and authorization. That's good; didn't know we had all that. Can't read that. I think I had a big screen for this presentation. Oh, accounts. I went to accounts payable. We went to ET1 - Eagle Transaction 1, which is now called Debit-Credit and now called TPC/A, but it was at that time called Eagle Transaction 1.
C. Mohan: Where did the name come from?
Mike Blasgen: Eagle was an IMS successor; it was going to do everything. And they were very worried about path lengths. So there had been something in IMS called TP1. But TP1 was more of a general characterization; ET1 was a specific program. And then Jim wrote all this stuff down in an article that he published in Datamation. It had Anonymous et al. or something like that as the author.
Jim Gray: Actually in about 1972, General Automation beat IMS at the Bank of America for the automated teller system, and we saw the attack of the killer minis that were going to wipe out all the mainframes. We had maybe three or four years before the company was going to go out of business. And that was some of the stuff that was driving FS; some of the stuff that was driving ...; we thought the computer industry was going to change, really dramatically with the minis. And the benchmark that B of A used was canonized as TP 1, 2, 3, 4, 5, 6, 7. TP7 was a minibatch; TP1 was a very, very light debit-credit transaction; and we ended up just using TP7 and TP1. IMS Fast Path was built to run TP1 well and that was the whole reason for doing IMS Fast Path. TP1 was recoded - it was a set of IMS calls - it was recoded for the VSS interface, which was the early name - the Andy Heller name - for what became DB2 eventually. It was called Eagle in one of its many incarnations ...
Mike Blasgen: To call Eagle the same as DB2 is misrepresenting history quite a bit. Eagle wasn't the focus on relational; it was kind of an IMS successor.
Franco Putzolu: Well, we all came from amoebas and ...
Mike Blasgen: Oh, I see, it's all the DNA; a tree is the same as a human, right? With a few changes in the genes. Well, Eagle was a tree.
C. Mohan: Wasn't the data manager the same?
Josephine Cheng: Yes.
Jim Gray: It was the same project.
Mike Blasgen: OK, so Eagle Transaction rewrote ...
Jim Gray: They rewrote in VSS terms and ...
Mike Blasgen: Well, I believe it's possible, likely, that I'm the first person ever to write ET1 in relational. Because I decided we had to have it and everybody else was busy. At that time I was a second-line manager. So I wrote it. And I remember taking some liberties with it.
Tom Price: Everyone who's ever written it has probably taken some liberties.
Mike Blasgen: ... It might well have been Bob Selinger that helped, I mean I know you were working on tracing. Did you wind up tracing ET1?
Bob Selinger: Well, we called it SR1; it was the relational equivalent of ET1.
Mike Blasgen: All right, the details, which we can talk about at lunch, had to do with whether you had to keep the account records sorted by account. Or whether you can just insert the records into the table and then, at the time you print the bill, sort them then. And I argued for sorting then, because relational doesn't really have a notion of sorting order anyway. But then there was this argument about whether it was equivalent or not, because IMS was keeping it, if you will, more organized.
So what have we got? OK, you get the idea. There was this presentation that we took everywhere. This went to six GUIDEs and thirteen SHAREs and database conferences all over the world. I have a proceedings of one of the last of the biggies, which was in 1983, at Wang Institute, the Eastern Computer Science Colloquium or something, and Jim and me and Mike Stonebraker and Ted Codd - oh, gee, it's just a long list of neat names ... Non-procedural, lots of optimization. This was one of my favorite talks. It talked about how there were many ways to do these things and how you would need an optimizer to make your choice. ... I love this: avoids exponential growth of optimization time, and N-way join by pruning. Aren't we good. And here's ...
Bruce Lindsay: Eight megabyte database; whoa!
Mike Blasgen: We had an eight megabyte database! [laughter] ... OK, so you get the idea.
So I'll put in some of my own reminiscences here for a minute. At one point, Irv and Frank decided to change the management structure, and in something that I've never seen happen before, Irv worked for me and I became Irv's manager. We switched places. So I became the manager of the RSS and Irv became incredibly productive. [laughter] And I became incredibly non-productive. I spent a lot of time making those slides. And then Leonard was offered a job back working for Bob Evans in New York, and that was onward and upward. He became a director at that point. Frank was named the manager of the Computer Science Department. That left open the job of manager of database systems, which Frank had been, so I took that job. That was in, like, the end of 1977 probably. And I stayed in that job until July of 1979, when my wife was offered a job in Washington DC, and I thought that I should follow her, rather than be left behind and not have a marriage. So we moved to Washington, and Robin Williams succeeded me in the job of manager of Database Systems. And then shortly thereafter, Frank King left, also to go back to work for Bob Evans, and Abe Peled came. And then the department changed quite a bit; it grew, and it became much more diverse. We were just talking a minute ago about the fact that as System R became more successful, it accreted more. I remember Don Slutz, for example, was not working on System R in 1974 or something. But you joined, what the RDS in ...?
Don Slutz: 1975.
Mike Blasgen: ... 1975, right. You heard the story about Franco in 1974. I mean he was grabbed out of some other thing, and then Don Slutz. I think that by the time of the end, if you count, I mean Juan Rodriguez-Rosell was a performance analysis guy interested in whatever - operating systems performance - but we converted him into a System R performance person. And so by the time we were done we might have had as much as half the department working on - not necessarily System R, but things related to System R. Mario and Paolo Tiberio were working on a tool to do database design and that was a database design tool that was keyed off System R. Several of the projects that had preceded System R, sort of like EXPRESS, sort of became smaller projects; some of the people left.
I tell you, my artifact [the CS brochure] is winning. Because during the time I've been talking, it's gotten to the third person, and he's whispering, "Look at this!"
The joint studies were brilliant in the sense of forcing IBM to move; I'm not sure it was so important for what we learned. Frank had the idea of doing joint studies; actually, I think this was Ralph Gomory's idea, or Leonard's, or somebody's; Ralph's, I think, as much as anybody. So we established a relationship initially with Pratt & Whitney Aircraft, then with Upjohn Pharmaceuticals in Kalamazoo, Michigan - Pratt & Whitney Aircraft is in East Hartford, Connecticut. And finally, after about a one-year delay, a major relationship with Boeing Aircraft, in Seattle, Washington. We have the reports; in fact I have copies of reports and there are copies up there [on the artifact table]. I think Bob Yost brought every single report we got. It's not quite the SQL standard - I mean nothing is that big - but it's a lot of paperwork that was generated. I was thinking that maybe Don could talk about Pratt & Whitney and Upjohn. Are you the best person to talk about that?
Don Chamberlin: I'm one such person, and I'm sure there are others, too.
Pat Selinger: Jim Mehl was the chief contact for Boeing as I recall.
Don Chamberlin: Jim, do you want to tell us about Boeing?
Mike Blasgen: Well, historically, it came third.
Don Chamberlin: All right. The initial joint study was, as Mike says, at Pratt & Whitney in Hartford. I'm not sure any of these joint studies really exercised all of the neat stuff that we had to offer, like concurrency and transactions and locking and different degrees of consistency and all those different things - they really didn't care about any of that stuff. The applications initially in Pratt & Whitney - they were a manufacturer of jet engines, and they used System R for inventory control of their parts and supplies that they were using to build jet engines. The Upjohn Drug Company in Kalamazoo used System R to store the results of clinical experiments that they were using in support of their drug applications to the FDA. The thing that I remembered the most about these joint studies is that we got a lot of trips to beautiful Hartford, Connecticut and Kalamazoo, Michigan, and that was kind of neat. We got factory tours; we got a tour of the jet engine factory in Hartford; we got to go through all the machine shops and watch them building jet engines - that was kind of neat. They told us how at every stage of building the engine they would weigh it very carefully to see if they had left any extra parts inside.
Tom Price: I remember when we went to Pratt & Whitney the first time, we showed them all the system mods that they needed to put on their VM systems so that they could run it, and they already had system mods on all those same lines - local mods. It was a real mess.
Don Chamberlin: I remember in particular a wonderful trip to Kalamazoo to visit the Upjohn people. They had a place that they called their homestead, which was a beautiful Victorian mansion on a huge plot of land outside Kalamazoo. It had a pond and a greenhouse and all sorts of very wonderful accommodations that they kept there for visitors. Tandem bicycles parked around for people to ride around and have a good time. They put us in a room where there was one whole wall completely filled with different kinds of liquor. We asked them if we could take home anything that we didn't drink. That was a nice trip. [laughter]
A bunch of things were happening at about this time that I think we ought to mention just in passing. One was that we had to change the name of our language from SEQUEL to SQL. And the reason that we had to do that was because of a legal challenge that came from a lawyer. Mike, you probably can help me out with this. I believe it was from the Hawker Siddeley Aircraft Company in Great Britain, that said SEQUEL was their registered trademark. We never found out what kind of an aircraft a SEQUEL was, but they said we couldn't use their name anymore, so we had to figure out what to do about that. I think I was the one who condensed all the vowels out of SEQUEL to turn it into SQL, based on the pattern of APL and languages that had three-lettered names that end in L. So that was how that happened.
A couple of other interesting things happened about that time, too. Our famous paper that got published in TODS: the second issue of TODS had a paper on System R. And there's a story about that, too. I want to prove something to you by showing you a foil here. If you've ever seen a reference to that paper - that TODS paper - it says the title of it - this is the famous fourteen-author paper; everybody that had ever attended any kind of a System R meeting was included as an author of this paper - so this is the cover page of the manuscript that we sent to TODS for the fourteen-author paper. If you've ever seen a reference to this paper, its title as it got published was "System R: Relational Approach to Database Management" - it didn't have the "A" in it. When we wrote the paper, it said, "System R: A Relational Approach to Database Management"; when it got published, the "A" went away. And the reason for that is because when the galley proofs came back from TODS, they sent them back to us for proof-reading, and all of the fourteen authors were alphabetical, and Astrahan of course was first, so a lot of our papers are Astrahan et al. The penalty that Morton had to pay for that was, he had to proof-read the galleys. So I gave the galleys to Morton, and this was a pretty long paper, he had a lot of proof-reading to do, and he was pretty busy. So he did a pretty good job of proof-reading, but he didn't proof-read the title. So that's what happened to the "A".
Another thing that happened at about that time was that some technical problems came up that got dealt with and solved, and I thought they were kind of interesting and somebody ought to talk about it. I think somebody ought to talk about the Halloween problem. Pat, you had a lot to do with the Halloween problem; do you want to talk about it?
Pat Selinger: It happened in I guess, was it 1976?
Don Chamberlin: 1976 or 1977.
Pat Selinger: I'm having a little trouble remembering this, but we had exercised the "person who earns more than their manager" query to death, and finally got to the point where the optimizer was choosing indexes sometimes to implement this query and it happened to think that the Salary index was a pretty good index to select for this. And having selected the Salary index for the first time in us testing out the optimizer, we ended up discovering that this query didn't stop. Because we were using the Salary index to go after the Employee table and we were also updating it, and Don Chamberlin kept getting more and more raises. Which made him very happy, but it made us optimizer folks a little bit uncomfortable. So Morton and I sat down and discovered this and analyzed what was going on, and came to one of your RDS meetings and it happened to be on Halloween. So we ended up telling the group about this and consulting the general wisdom to figure out what in the world we ought to be doing about this thing. As we talked about it, it came to be known as the Halloween problem. And I think it's still kept that title to this day.
Don Chamberlin: It's famous in the industry - everybody knows the Halloween problem. And it happened to be discovered on Halloween. The query was ... they had submitted a statement that was supposed to give a ten percent raise to every employee who earned less than $25,000. This query ran successfully, with no errors, and if you examined the results, all the employees in the database earned $25,000, because it kept giving them a raise until they reached that level. So that was how the Halloween problem got its name.
Pat Selinger: An interesting footnote is that we just discovered another one of these as sort of a variation on that, in the latest work that we did having to do with referential integrity and things like that, where the referential integrity relationships were going to trigger off the same kind of non-stop behavior.
Mike Blasgen: It's interesting because all these odd-ball things had names: there were phantoms, and there were other things, and those had to do with names that were somehow representative of what you were observing, right? So the phantom was because it was something that was sort of there, but not there; the name was descriptive. And this was called the Halloween problem not because it surprises you, or it's spooky, or trick-or-treat or anything; this is because it happened to be discovered on Halloween day. But I think most people think it's the other; I think most people think it's called Halloween because it's so surprising. But it's not.
Don Chamberlin: Here are a couple of more artifacts from the joint-study days. I wrote most of the manuals for the users at our various joint studies to use and we designed a nice logo. And Jean Chen helped us make these nice binders that a lot of you still have. When we went to Upjohn, like other places, they gave us a factory tour. We got a factory tour of Boeing; we got to see them putting together 747's and at Upjohn we got to see them making vitamin pills. They would give the vitamin pills away for free in the cafeteria and they all gave us this nice sign. This sign says, "Keep the quality up. W.E. Upjohn." And here's a picture of a guy squashing a pill with his thumb. It says, "Upjohn: originator of friable pills." "Friable" means sort of squashable. So W.E. Upjohn was the originator of friable pills and that's the heritage of the Upjohn Company, and that was apparently what he said. You know, Thomas Watson has a lot of famous sayings, you know, "Respect the individual" and stuff. What they say at Upjohn is "Keep the quality up." That was their slogan.
In talking about these users, we always remember Upjohn and Boeing and Pratt & Whitney, but I just wanted to mention the fact that we had a lot of other users, mainly inside of IBM, as well. IBM Owego was using System R to build attack helicopters I think. There were several groups at STL that were using System R for different things; there was a guy there named Gary Haas. Gary and Frank Nargi??? - do you remember those guys?
Bruce Lindsay: SREDIT.
Don Chamberlin: Yes, they built a sort of early GUI on top of System R, called SREDIT. At [IBM] Poughkeepsie, there were some folks using System R in a kind of a design-automation system. At Yorktown, there was a guy named Fred Damerau who was doing natural-language queries. He had a project named REQUEST that was based on System R. Most of the IBM Scientific Centers had System R installed. Alex Hurwitz installed it at Los Angeles; Yoichi Takao installed it at Tokyo; Jean-Jacques Daudenarde installed it at Paris.
Jean-Jacques Daudenarde: We also had a joint study with a company who was developing some helicopter design based on two-column relations in System R; exclusively two column relations.
Don Chamberlin: The system wound up being installed in Madrid, Heidelberg, Rome, Cambridge; Scientific Centers all over the world.
Bob Yost: Do you have any idea when the last System R site went down? I think Upjohn used System R even after SQL/DS came out because they didn't want to pay for it. If you used System R, then you came to have it free for a long period after that. So I think they used it for a long time.
Don Chamberlin: I don't know how long that went on.
Bob Yost: They were running that on a Model 145 with one megabyte of memory. I saw that and I said, "My God, I'm trying to put DB2 in my eight-megabyte machine and it hardly fits at all."
Don Chamberlin: So in looking back on this, one of the things that I marvel at is the impact that this work had when it was really, by today's standards, very small according to certain measures. And I brought along a couple of foils that kind of show you a measure of things. This is a profile that I drew up for Frank King one time that shows the number of people that were involved in System R at various stages of its life. From 1973 to 1978, this was kind of the profile. We started out at about ten or eleven people. And this was when the Phase Zero prototypes were installed, and this was the different installations of System R at the joint study sites. As you can see, we never had twenty people in System R; probably the average was around fifteen. And after 1977, it fell off to half that. So the area under that curve was not a lot of manpower by Microsoft standards. But we had a pretty big impact from it.
Bruce Lindsay: By Santa Teresa standards.
Don Chamberlin: As far as the code size of System R is concerned, believe it or not, it wasn't very big. The RDS was mostly written in PL/1 and had 38,000 lines of PL/1, and about 9,000 lines of assembler. Now 38,000 lines of code isn't a lot; I mean, it's pretty hard to find any kind of a product that small. The RSS was written in PL/S and had another 35,000 lines of code. So add this up, maybe 80,000 lines of code and that was System R for you. It's not a lot for what we were able to accomplish with it.
This was the size of the load map. The RSS took a megabyte and the RDS took another 1.2 megabytes on top of that. That's all there was in the systems that we installed at the joint study sites. So it wasn't very big, either in terms of its code size or the people that wrote it.
I kept track of the two lists, called the bug list and the wish list, during the time that we had the joint studies going. We would have quarterly reports at each of these joint studies and they would wish for things and I would write them on the wish list; we mostly didn't implement any of them. At the end of the project, I think there were something like 160 wishes that were open. The bugs we tried to fix, though, and I've got some statistics on this. This is where the bugs came from. Over the course of the project, we had 251 bug reports. We found most of them ourselves, but Boeing found a lot, too. These were the number of bugs that were found at different joint study sites. This is what happened to them. Out of those bugs, we fixed 71%, we couldn't reproduce quite a few, and some of them we rejected, and so on. So we did our best to ...
Jim Gray: Ten percent are features, I heard.
Don Chamberlin: Yes, we declared some features. And this was the hero list. This is the people who fixed the most bugs in System R. And Wade is on the top of the hero list, and Don Slutz, Jim Mehl, Irv Traiger fixed a lot of bugs.
Mike Blasgen: The RSS didn't have any bugs. [laughter] No, it's true, the reason is because much of the RSS was written by Franco. No, it's really true; Franco never wrote a bug. Except for one, right, Bruce? Did you find one?
Bruce Lindsay: One.
Mike Blasgen: He wrote about half of RSS, and I think we found one bug. And that was after nine years.
Brad Wade: I remember index management, though, as being a trouble ... [laughter]
Franco Putzolu: How does the wish list compare with what was implemented afterwards by other systems? Did you ever look at that?
Don Chamberlin: I haven't really looked back and analyzed that. I couldn't tell you about that.
Pat Selinger: Do you have a copy?
Don Chamberlin: I have a copy at home; I didn't bring it.
Bruce Lindsay: We should send it to the SQL 3 people. [laughter]
Roger Miller: Our wish-list is more like a database.
???: It fits in eight megabytes?
Roger Miller: No, it won't fit in eight megabytes.
Don Chamberlin: Raymond, did you want to talk some more about the compiler and interpreter issue? You had a lot to do with that.
Raymond Lorie: I don't remember it. [laughter] I must say I tried to locate my foils that I used when we had a little meeting but my database system is not up to that task. I remember one meeting; I believe Don was there, and Irv, and Morton, I think, and Frank King, of course. I showed how a compiled SQL program would simply comprise a few assembler instructions on top of the RSS. It was amazingly short. It didn't stay that short, but short enough to pursue the idea. Now of course, because we started from an interpreter, we went all the way, and compiled into machine language. Later on, because I think Franco didn't like that, we came back to well prepared tables, rather than code; but of course the idea of compilation remained unchanged, because you do the optimizer once and you package the whole thing; then you invalidate the "code" if things change that impact the access strategy. So that was one of the contributions to the system. It was a good example of how you can change direction in a group and convince people to follow it. It was a good experience. It also brought me into the RDS, where I spent several years, which I enjoyed very much.
Don Chamberlin: I think that was one of the key things that made System R a success: Raymond's idea to compile rather than interpret our high-level language. Because that was the thing that was responsible for our performance, and performance was the thing we had to prove to get relational accepted. Everybody agreed it was sort of neat, but they didn't think it could perform. And Raymond was the man that made it perform.
Mike Blasgen: Actually, this is sort of the history. You know, the first FORTRAN compiler that was ever made was probably the best FORTRAN compiler in terms of producing the fewest instructions per FORTRAN statement. Because they spent so much effort to get it all just right. For example, the reason FORTRAN has a three-way branch - IF (ABC) 1,2,3 - is because the machine had a three-way branch, and that way they could generate that in a single instruction. It's like CAR and CDR; they worked very hard on performance, from the very beginning.
Let's see, in the TODS era, I gave talks at the TODS era too, and then we had pre-optimized packages - that was some idea that we had that we wouldn't have to go through optimization. But that was before compilation, correct?
Don Chamberlin: Yes.
Mike Blasgen: Before compilation we were already worried about performance, so we didn't believe that we would fully interpret, but we had this idea of pre-optimized packages, I'm not sure if it was all worked out.
Tom Price: It was like you only had to optimize it once, but then you were still interpreting.
Mike Blasgen: Right, you'd interpret the plan.
Franco Putzolu: Weren't pre-optimized packages more like views, special ...
Mike Blasgen: Oh, were they?
Pat Selinger: Yes, I believe they were just for views.
Mike Blasgen: Oh, I see it was really like combining the two, doing composition in advance. But then I remember that after Raymond's talk, and we all decided that compiler was it, we had still the question of what to do for ad hoc query. Because Raymond had not proved anything about ad hoc query; he had proven something about canned transactions, that you want to compile canned transactions. So then the question was, "What about ad hoc query?", and then there was a bunch of work that Raymond and Pat, or I don't know who was involved in that, but I know Pat was involved - where you did measurements of interpretation of ad hoc query. Then the plan became, "We'll do both an interpreter and a compiler." And that seemed like it was going to be a mess, just in terms of keeping the semantics straight, because you have two implementations of the same thing. How are you going to make sure? So then Pat was able to conclude, right, that we could do both ad hoc and canned transactions with the compiler.
Pat Selinger: I think Don was involved in this, too.
Mario Schkolnick: There were lots of measurements Morton and I were doing on the number of pages touched when doing an ad hoc query. Once we found 92 pages were hit to save touching two pages at runtime. So I remember Franco picked up the code of the optimizer and said "Well, let's figure out for a simple query how to reduce the number of pages that the optimizer was touching." So there was all that work going on ...
Franco Putzolu: Yes, Ray's proposal was good. Without it, we wouldn't be here today.
Mike Blasgen: I think there are many such things that happened, that without that happening, maybe we wouldn't be here today. But I think I agree that this was an absolutely critical one. When we did do the ET1, by the way, and we wrote down the debit-credit transaction and ran it and did all the path lengths, if you count I/O as one instruction, the path length for ET1 was about fifty thousand instructions, and that was state-of-the-art. So System R, thanks to the work of Raymond, became state-of-the-art in performance, that is, there was no database system that could run any faster than System R on canned transactions, light-weight transactions, which is what you would think would be the weak link in a relational database.
Notice it is really true; I hadn't realized this, but this project changed quite a bit from the original idea of SEQUEL, which had to with vastly broadening the audience of computers to include people looking up recipes and mechanics wanting to know how to change oil in the car that just rolled in, to ET1. It just did that in a couple of years, and we all just thought it was great. And then we were worried about FRRs and SRBs; you know we wanted to do SRB scheduling, because then we could get three instructions out of our path length. Oh, and Irv Traiger went through and did all kinds of work to get eight instructions out of this, and four instructions out of that. Do you remember all the work you did in the RSS?
Irv Traiger: It was Fast-Next.
Mike Blasgen: Yes, he did Fast-Next! FNEXT. It just cached a bunch of stuff in YTABLE1, right, was what it did. And we had cache-invalidation tricks. But if you didn't invalidate the cache, we had Fast-Next, and that took, oh, thirty instructions out of a NEXT, and that was the key to survival.
Franco Putzolu: We got really carried away with machine language.
Mike Blasgen: You think that was a mistake?
Franco Putzolu: Yes; on the other hand, recently I've seen some systems doing that.
 K. Shoens, A. Luniewski, P. Schwarz, J. Stamos, J. Thomas. "The Rufus System: Information Organization for Semi-Structured Data" Proc. VLDB, Dublin, Ireland (1993).
 E.F. Codd. "Seven Steps to Rendezvous with the Casual User" Proc. IFIP-TC2 Conference on Data Base Management, Cargese, Corsica (April 1-5, 1974) pages 179-200.
 E.F. Codd, R.S. Arnold, J-M. Cadiou, C.L. Chang, N. Roussopoulos. RENDEZVOUS Version 1: An Experimental English Language Query Formulation System for Casual Users of Relational Data Bases. IBM Research Report RJ 2144. San Jose, California (January 1978).
 RDS stands for Relational Data System.
27 M.M. Astrahan and D.D. Chamberlin. "Implementation of a structured English query language" CACM 18, 10 (October 1975), pages 580-588.
 J.J. Donovan, L.M. Gutentag, S.E. Madnick, and G.N. Smith. "An Application of a Generalized Management Information System to Energy Policy and Decision Making - The User's View" Proc. NCC, AFIPS Vol. 44 (1975) pages 681-686.
 Now Pat Selinger.
 P. Reisner, R. F. Boyce, and D.D. Chamberlin. "Human Factors Evaluation of Two Data Base Query Languages--SQUARE and SEQUEL" Proceedings of the AFIPS National Computer Conference, Anaheim, CA (May 1975) page 447.
31 P. Reisner. "Use of Psychological Experimentation as an Aid to Development of a Query Language" IEEE Transactions on Software Engineering, Vol. SE-3 (May 1977) page 218.
 D.D. Chamberlin and R.F. Boyce: "SEQUEL: A Structured English Query Language" Proc. ACM SIGMOD Workshop on Data Description, Access and Control, Ann Arbor, Michigan (May 1974) pages 249-264. Note that the reference in the TODS paper says SIGFIDET instead of SIGMOD.
 R.F. Boyce and D.D. Chamberlin, "Using a structured English query language as a data definition language" IBM Research Report RJ1318. San Jose, California (December 1973).
 Nelson Mattos. "An Overview of the Emerging Third-Generation SQL Standard" SIGMOD '95 Tutorial.
 M.M. Astrahan, M.W. Blasgen, D.D. Chamberlin, K.P. Eswaran, J.N. Gray, P.P. Griffiths, W.F. King, R.A. Lorie, P.R. McJones, J.W. Mehl, G.R. Putzolu, I.L. Traiger, B.W. Wade, and V. Watson. "System R: Relational Approach to Database Management" ACM TODS 1, 2 (June 1976) pages 97-137.
 GPD stands for General Products Division, which operated the San Jose facility hosting the research lab.
 VSAM stands for Virtual Sequential Access Method.
 Ralph E. Gomory was Director of IBM Research from 1970 to 1986.
 R.M. Karp and R.E. Miller, "Properties of a Model for Parallel Computation: Determinacy, Termination, and Queuing," SIAM JAM 14, 6 (June 1966), pages 1390-1411.
 J. Gray, P. McJones, M. Blasgen, R. Lorie, B. Lindsay, T. Price, F. Putzolu, I. Traiger, "The Recovery Manager of the System R Database Manager" ACM Computing Surveys 13, 2 (June 1981), pages 223-242.
 G. Radin and P.R. Schneider. An Architecture for an Extended Machine with Protected Addressing. IBM Poughkeepsie Laboratory Technical Report TR 00.2757 (May 21, 1976).
 R. Demillo, D. Dobkin, A. Jones, and R. Lipton, editors. Foundations of Secure Computing. Academic Press, New York (1978) pages 39-56.
 Anon et al. "A measure of transaction processing power" Datamation 31, 7 (April 1, 1985) pages 112-118.
 J.R. Good. "Experience with a large distributed banking system" IEEE Computer Society on Database Engineering 6, 2 (June 1983).
 N.C. Shu, B.C. Housel, R.W. Taylor, S.P. Ghosh and V.Y. Lum. "EXPRESS: A Data EXtraction, Processing, and REStructuring System" TODS 2, 2 (June 1977) pages 134-174.
 STL stands for Santa Teresa Laboratory.
 Pronounced Shred-it.
 R.A. Lorie and B.W. Wade. "The Compilation of a High Level Data Language" IBM Research Report RJ2598. San Jose, California (August 1979).
 Pronounced Fuh-next.