Witold
LITWIN
ACM-Fellow
2001 (1st ever in France)
CENTRE DES ETUDES ET
DE RECHERCHES EN INFORMATIQUE APPLIQUEE
ACM-Fellow 2001 Award
Ceremony (Amateur Video) Play
(DSL
or Local Net & Windows Media Player and Real One Player Recommended; click
on Play button without waiting for the full file download. That
one may take some minutes otherwise)
ACM Special Event to Honor Leading European Computer Scientists. Oct 8, 2009, Paris.
Amateur video features the main part of the ceremony. Wmv or svf or, or flv, allow a minute or so before the streaming begins).
Prof Dame Wendy Hall, ACM President, in green jacket, greets European
Recipients of major research-oriented ACM Awards. ACM VP (sorry not to retain
the name) calls us by name and the original Award Citation. Sir Tony Hoare,
called, regretfully, before the video begins, for his 1980 Turing Award, starts
the “pas de deux”.
The two photos, 1 and 2, are derived from
the video and feature eight from nine honored European ACM Fellows. The picture misses the last one called in for
technical reasons. I possibly fix it in near future. Finally, guess who folks
on the final enclosed souvenirs are, P1, P2, P3, especially the
blond lady with the (Andy Warhol’s to respond to some inquires) scarf.
Address
Prof. Witold LITWIN
Lamsade
Université Paris Dauphine
Pl. du Mal. de
Lattre de Tassigny
My Virtual Computer History Museum
Unsung
Pioneers of Distributed Data Universe which We Take for Granted Today
Open
to public 24/7. Enter Here
Witold Litwin. SIR
SQL for Logical Navigation and Calculated Attribute Free Queries to Base
Tables. Lamsade Res. Rep., April 15, 2024, 15p, pdf. Submitted.
The
report expands “Manifesto for SIR SQL” with figures, additional details and
examples.
Witold
Litwin. Manifesto for SIR SQL. 9th Int. Conf. On
Information And Education Innovations (ICIEI 2024).
April 12-14, 2024, 5p, ACM Digital Library, ACM ISBN 979-8-4007-1640-9/24/04 https://doi.org/10.1145/3664934.3664948
Pdf protected by on-demand password, to respect ACM
copyrights, here
.
Witold
Litwin. Prototype Front-End for Stored and Inherited Relations On SQLite3:Demo for Supplier-Part Database. Aug. 15,
2023, 1p, pdf.
A Trusted Cloud Database System manages client-side encrypted cloud
DBs. Queries may include encryption keys. The DBS
decrypts/encrypts the data on-the-fly at the cloud. Plaintext is only in
protected run-time variables. Stored data are by default probabilistically
encrypted through AES. Any SQL queries are feasible, with negligible processing
overhead and practical storage overhead. This is a major advance over the
current alternative research proposals. We detail capabilities of a trusted
DBS. We adapt SQL to client-side key management. Queries may remain usually
almost as non-procedural as now. A prototype implementation appears easy.
Notice on Dec. 23,
2019. Browse through the recent abundant MS Docs on Azure
SQL Server with secure enclaves. Start
perhaps with the September 14, 2017 CTO’s framework statement: “Introducing
Azure confidential computing”, here.
In our humble opinion, no need to be rocket scientist to see that Azure SQL
Server is straight in the footsteps of our paper. Up to now, we could not
locate nevertheless any MS paper with the reference to. Old-time Academia
customs would imply so. If you spot one, forward the info, please.
Witold Litwin. Stored
and Inherited Relations. Lamsade Technical Report,
Mars, 2016. Revised, 33p, March 2017 snapshot archived in
Computing Research Repository (CoRR), pdf.
10p revised conference version untitled “SQL for Stored and Inherited
Relations” published at ICEIS 2019, pdf , pptx.
As the
name suggests, the next property of an SDVS is to typically use a large cloud.
The final distinguish property is that every SDVS search carries a user-defined
time limit, say R over the search time at every of its cloud nodes. The overall
worst search time is then R usually as well. To meet R, a node that gets a
search over some records, first estimates its throughput. If the calculus shows
the node unable to meet R, it splits the key set. The split sends a part of the
set, say P, basically a half, to another node, for processing of the virtual
records with keys within P. The partitioning is recursive till every node
estimates itself able to respects R. Presently, the hash partitioning and the
range partitioning of virtual records have been studied.
Use of
SDVSs allows the user defined worst-case time limit on the requested key
recovery. Smaller R usually requires a larger cloud, evidently. The cloud can
be heterogeneous, i.e. with nodes having different throughputs. One may also
tune the cloud size as well as the average R. One may finally make SDVS
reliable, i.e., protected against node failures. See the papers for more.
Beyond key recovery, SDVSs constitute a new tool for
the operational research. They appear 1st to allow some traditionally hard
calculations to be completed within the user-defined practical time limits. The
condition is the availability of a sufficiently large cloud, to spread the
calculations enough. One may cite the knapsack problems, the travelling
salesman problems, the combinatorial enumerations, e.g., of all the
permutations… It is an open research field at present.
We have proposed this concept with my dear co-author in my BDA 2010
invited talk, in Toulouse, The intention was to
facilitate the knapsack problems calculations over relational DBs. A k-way knapsack q-join over numerical columns r1…rk selects every tuple (t1…tk) with
t1+…+tk ≤ C, where C is a user defined constant. A practical query may
also carry the Order By and Top K clauses. For instance, someone may search for
ten sets of five real estate properties to buy from a listing of a thousand
with the cumulated price tag as close as possible but not over 1M$. We showed
ways to make knapsack joins more efficient than through the naïve nested loop.
That one was and possibly still is, the only technique used by DBMSs at the
time of writing. One may easily see that our example is then yet only a dream.
My talk opened the conference. A 2nd invited talk closed it. It was
by a nice researcher from Yahoo, New York. Amazingly, it turned out that she
was investigating similar issues. The goal of her work was indeed a
hypothetical Yahoo assistant for folks searching for a collection of IPhone’s
accessories fitting best a given budget. I did not hear since about anything
commercial, at Yahoo or elsewhere that would follow up her proposals or ours.
SDDSs are a class of data structures we have proposed
with M-A Neimat and D. Schneider at HP Labs, Palo
Alto, in 1992. This concept and the subsequent outcomes were apparently one “leg”
of my ACM Fellow Award. SDDSs were intended for multicomputers.
This naming was coined by A. Tanebaum, in his famous
book. It designated for networks of share-nothing processors (nodes) linked
through high-speed network, wide-area or local, or a bus. We published the 1st
SDDS termed LH* at ACM-SIGMOD 93. Soon after, HP Labs patented it. An SDDS can
span over potentially any number of nodes, scaling up or down, i.e., on more or
less nodes, transparently for the application. SDDS files can reach sizes and
access performance impossible for more traditional file structures. For best
access performance, an SDDS file should reside for processing in the
distributed RAM. Performance analysis has demonstrated experimentally the
search speed of up to 300 times faster than to a local (magnetic) disk. The
capabilities of SDDSs open new perspectives for many applications. Today, in
2016, many SDDSs have been proposed. Some are in everyday use by billions (see
below). There are many papers on the subject with thousands of references to
(see Google). SDDSs are especially investigated and now operationally built as
the infrastructure for P2P systems, grids and clouds: public, private,
hybrid.... CERIA prototypes focused mainly on so-called today RAM Clouds and NoSQL (cloud) databases, except for SD-SQL Server (Scalable
Distributed SQL Server), the 1st ever prototype Scalable Dsitributed
relational DBMS. Google for the research papers or see the one below. See also
the page of S. Soror or her video demo. The RAM Cloud orientation is in part
due to 1992 visionary lecture at Berkeley by Jim Gray. See my MS Redmond 2001
talk below or my SDDS class (in French) where one (almost) original slide of
Jim Gray on “distance to data” is present. Notice also the recent RAMCloud project headed by John Ousterhout (Stanford),
although its papers, surprisingly, do not cite any of ours. For further info on
SDDSs – see our classes Cloud Databases.
SDDS-2000 : A Prototype System for Scalable Distributed Data
Structures on Windows 2000 (ppt). MS Research, Redmond & NWU, Chicago, 2001.
This talk introduced SDDSs to MS Research audience. As
usual, most folks in the room did not bother to read the papers published
before. The talk largely predated any popular industrial development we
discussed. I still wonder to what extent it influenced the future Azure Table
designers. I was invited by two old pals D. Lomet & Per-Ake
Larson. With D. Lomet we have proposed earlier the bounded disorder (BD) file
access method. See our publication list and the Web for the follow up work.
With Per-Ake we had long lasting coop over dynamic
and linear hash schemes.
On the souvenir side, after the talk I chatted with
Rick Rachid. He was in charge of MS Research
then. He noticed the SDDs potential.
But, he was rather focused on the prototype MS wristwatch. It is now called
generally a smartwatch. He proudly displayed a
Cartier-like he had. We accessed the MFST Nasdaq
quote somewhere. Well, I expressed him doubts about the bright future of the
invention, given the display size. Smart phones invasion (headed then by Nokia)
was about to begin and these were naturally much less limited on that side.
Vice versa, his nice comments, based on his experience of building his
distributed OS, did not seem sharing my
enthusiasm for the bright future of SDDSs, i.e., of the cloud infrastructure as
we say today. Well, today clouds are in use by billions as we all know (see
also below). MS got finally so convinced to clouds that their main fan there
got promoted CEO. Actually, even the name cloud was coined later by MS. Smart
watches attempt to make it as smart phone accessories. No news about MS smart
watches as yet, in Feb. 2016. Researcher life has sometimes amazing turns.
LH* : A Scalable
Distributed Data Structure (ACM-TODS) pdf
The paper expands earlier results, reported at Sigmod Conf. It presents especially two versions of LH*:
with and without so-called coordinator. The latter uses a circulating token
instead. It was never implemented beyond what our paper reports (work by J.
Levy, during internship at HPL). After reading the paper, compare LH* features
to those of DHTs (Distributed Hash Tables). Especially, compare the messaging
cost of both types of structures. The paper proposing DHTs should say something
about. But didn’t. Worse, it missed any reference to LH* work. This disregard
for honest scientific practice was surprising. At least, since it was coming
from Berkeley U. You might wish to look also at the LH* scan operation. It was
never implemented to our best knowledge. A likely reason was the lack of good
programming model. Like the popular MapReduce
model. Basic MapReduce
uses nevertheless a static hash for the data distribution. It’s well-known as
often inconvenient. Amazon’s Elastic Map/Reduce is closer to what LH* could
provide. Again, a direction for volunteers.
LH*RS a highly-available scalable distributed
data structure (ACM-TODS 2005)
Presents the prototype
designed by Rim Moussa (her Ph. D.) and the theory
basis. Builds up on earlier Ph.D. prototypes. One by
J. Karlson (Uppsala) proposing a particularly elegant
variant called LH*LH.
The other by F. Bennour, called SDDS-2002, built
specifically for Windows RAM files. See CERIA publication list (or the Web).
Notice that in last 30 years, only a handful of French DB research papers
passed the ACM-TODS publication standards. Rim’s prototype was also
demonstrated at VLDB in Toronto. Early results for LH*RS were reported also in
a Sigmod paper.
RP*
: A Family of Order Preserving Scalable Distributed Data Structures (VLDB-94) pdf
See Google’s landmark
paper (OSDI06) on Bigtable. You likely used that
without even noticing. As billions of others do every day, to get pages from
Google. Appreciate yourself the similarity/differences between the BigTable structure and RP* schemes in our paper. In the
same spirit, take a peek on Azure Table, Mongo DB, PNUTS and Hive.
Needless to hide
nevertheless our opinion that all these great products root in one of RP*
schemes, called RP*S in our paper. None of their documentation we’re aware of
refers to it however. Unlike the honest scientific practice requires. I asked
Google guys why, when I gave my LH*RSP2P Google Talk, in Google Tech Talks
Series, 2007). All but one did not respond. The nice one, who, by the way, is
by himself among top researchers there, said: sorry,
we did not spot the paper. But we’ll cite it from now on. We’re systems guys.
Do not follow up DB conferences. Well,
VLDB where the paper appeared is among main DB gatherings. DB researchers
besides, follow mainstream conferences outside the field, e.g. Usenix, Mascots...
So with all due respect, we do not fully subscribe to this excuse.
Especially that the no-cite practice became disgracefully widespread in the
last decade. Despite great tools like Google, (with its BigTable),
Scholar, DBLP, Bing… you name it. Commercial R & D world seems particularly
affected by the disease. But, the university research is no more exempt either.
As we already mentioned for DHTs.
High-Availability LH*
Schemes with Mirroring (COOPIS-96) pdf
As the title suggests, we proposed a fully replicated LH* scheme. Like
any such scheme, it is simple to implement. The drawback is the N*100% storage
cost for N replicas, as usual as well.
LH*S
: a High-availability and High-security Scalable Distributed Data
Structure (IEEE-RIDE-97) pdf
(Res. Rep. Version)
First in the series below, exploring parity codes for high-availability
SDDSs. Culminating with the LH*RS in ACM-TODS referred to above. The low
storage overhead makes such schemes potentially more attractive for a RAM
Cloud.
LH*g
: a High-availability Scalable Distributed Data Structure (CERIA & U. Linkoping Res. Rep. & IEEE TKDE Journal, Vol.
14, Issue: 4, 2002) pdf (Res. Rep. Version)
LH* Schemes with
Scalable Availability (IBM-Almaden Res. Rep.) pdf
Work patented by IBM. Republished in IEEE tutorial book
(J. Menon ed.). See the Web.
Design
Issues for Scalable Availability LH* Schemes with Record Grouping (DIMACS 99, Princeton U. 1999 & IBM-Almaden
Res. Rep.) Idem. pdf
Design Issues for
Scalable Availability LH* Schemes with Record Grouping (DIMACS 99, Princeton U. 1999,
Inv. Talk, Ms Powerpoint) ppt
LH*RS: A High-Availability Scalable Distributed Data Structure using Reed Solomon Codes pdf (CERIA Res. Rep. & ACM-SIGMOD 2000, Dallas).
Scalable Distributed Data Structures for
High-Performance Computing ppt (Aurora-2000,
1st Europ. Conf. on High. Perf. Comp., Vienna 2000, Inv. Talk, Ms
Powerpoint).
LH*RSP2P:
A Scalable Distributed Data Structure for P2P Environment. Conf. Intl. sur les NOuvelles TEchnologies de la REpartition
(NOTERE 2008), Lyon June 2008 (Keynote).
pdf.
This scheme is remarkable by its one message forwarding cost for the
worst case key-based addressing in an SDDS over a P2P cloud. That is, where
every node is both client and server. Like, e.g., seti@home
cloud. It is the least possible such cost for these clouds. We recall that LH*
in general is the only to provide the least such cost, of two messages, for any
SDDS over a client-server cloud, i.e., where all nodes are servers only. To
compare, the same cost for DHT is O (log N) for N-node cloud. The only “competitor” to LH*RSP2P is the
scheme proposed by B. Liskov & al. That one has
however in addition a specific thought more sporadic messaging cost to propagate
the file scalability.
The paper reports on initial results of Ph.D. research by Y. Hanafi.
See the Thesis & his journal paper for more.
An
Overview of a Scalable Distributed Database System SD-SQL Server. British Natl. Conf on Data Bases (BNCOD-06),
Dublin, Juillet 2006. Keynote.
This was 1st in the series devoted to SD-SQL Server: a scalable
distributed cloud DBMS. The system was the subject of the Ph.D. by S. Sahri. We
have presented SD-SQL Server twice to SQL Server R & D team at MS Redmond.
The talks seemed well-received (as well as at BNCOD). However, the transparent
table scale up of
SD-SQL Server still does not have the industrial counterpart (Feb. 2016). Neither in SQL Server nor in MS Azure SQL Database, nor elsewhere.
The popular manual approach instead, by so-called industrially sharding, e.g., in Azure, is comparatively in the Middle
Age.
Multidatabase Systems
MDBSs are systems for management of data in multiple autonomous
databases. Work on this concept was the other “leg” of my ACM Fellow Award. We
proposed it in 1980, within SIRIUS project at INRIA. The intention was to
generalize adequately the database model, i.e. the concept and architecture of
a DBS (Database Management System). Our “root” was the concept of a multi(data)base defined as “DB set of DBs”. Quite popular
today, e.g., see the architecture of SQL Server and more at the Web. Our first publication on the topic was at the
2-nd ACM Annual Computer Exposition in Lafayette, LA. See our publication list
for the reference. To my happiness, the talk was well received. Perhaps with
the exception of the alligator that run away when I
stepped upon on the university garden pathway in the dark. A major benefit was
also the acquaintance with a dear friend since, Peter Scheuermann, Prof.
Emeritus, NWU.
First prototype MDBSs, named MRDSM and SYSIDORE were
designed at INRIA and presented at 1st Intl. Symp. on Distributed Databases in Paris, March 1980. One can still (June 09) see the short
description of MRDSM as one of pioneering CS tools at the Multics
site maintained at MIT.
A central capability of an MDBS is the multidatabase definition and manipulation language, e.g.
our MSQL for the relational data. This capability lacked to DBSs. Today,
research on the multidatabase systems is among most
popular topics. Most of DBSs silently became MDBSs, e.g., MsAccess,
SQL-Server, Sybase, Interbase, Oracle, MySQL, even
DB2 (the latest among the converted ones). Postgres resisted
up to now (Feb. 2016). The industrial multidatabase
capabilities are still rather limited as compared to those known at the
research level. See, e.g., the capabilities proposed by MSQL language. See our
journal papers on it or M. Rusinkiewicz & al
book. Many research issues remain open. Techniques for MDBSs apply also
generally to Meta-search engines on Internet. E.g., Rob's SearchEngineZ interfacing dozens of search engines. Or the MetaCrawler that searches Google,
Yahoo, Bing an others. Also - EDreams, Google
Shopping, Comparateur Obsèques i.e., Funeral Cost Comparator (I get this and
similar friendly spam daily since my retirement started) …
There is a very large literature on MDBSs. E.g., Google just displayed the estimate of 237K entries for my
“multidatabase” keyword. If interested & French
understanding, start perhaps with my MDBS classes. Also see the Virtual Museum
above for notable grey literature, witnessing the early effort y its authors who should not get forgotten.
Today, the utility of multidatabase
systems is obvious. It wasn’t so, by far, when we started to talk about the
idea. The dominant trend was that of a single (integrated) DB. Possibly very large whenever needed. This was actually the
motivation for the name of the well-known VLDB conference series. Also, there
was a paper by Tom Steel, I believe, conjecturing the construction of a DBMS
managing worldwide data as a (single) DB in a couple of years. Besides, we had
heat discussions at DB conferences. To illustrate how dominant the single DB
idea was, to aim at by DBMS designers, let me recall one anecdote. I was on a plane to Cannes with one of the DB
arena friends now, to attend VLDB-81 there. He was already and still is, among top
researches in our field. When I asked for opinion about the idea, his response
was straight: “Why should we bother about managing multiple DBs if we can
manage only one”. That was the end of
the discussion, for the flight and until today.
4.Other Research
Our research in 70-80-ties on virtual and especially
linear hashing is sufficiently known to not be further recalled. Perhaps,
except for an anecdote. Our 1st paper ever on it was a Note in highly
prestigious “Comptes Rendus”
of French Académie des Sciènces.
To my stunning surprise, the Académie changed my
original wording “Hash-Coding” in the title into “Codage
Découpé”. Everyone concerned in France knew the
meaning of the former, as there was no official French substitute. In contrast,
no one knew the latter. The Académie just coined it
because of my Note, in their role of the Guardian of Purity of French language.
To make the link, they left however the “hash-coding” in the abstract. I
wondered why they did not use therefore the existing, although not yet used for
the computers, word “hachage”. The response was: no
concept of coding in that one. Anyhow, the story seems unique & I’ve never
ever seen the term “codage découpé”
applied since.
Now, let me recall some topics I also investigated 40-30
years ago. I believe they merit more attention or at least citations they’ve
got since.
Structural
Numbers for Graphs & Electronic Circuits pdf
The concept was proposed as algebra for formal analysis of electrical circuit to Chinese Academy of Sciences by K.T. Wang in 1934. Had some follow up there. That dried out during the Chinese tumult. The topic got rediscovered in 1950 in US by R.J. Duffin and followed up by a few, especially S. Bellert in 1962. Who became in 1969 my Magister Eng. Thesis advisor at Warsaw Polytechnic School. I continued the research at INRIA, ending up with the above pointed to INRIA Scientific Note 7 in 1972.
These were nicely edited, highly selective
publications, at most one per month. The caveat was they had to be in French.
A structural number (SN) basically represents a graph. Any graph,
not only an electrical circuit. Algebraic operations operate upon, in a
particularly simple way. Enumerate all the circuits, trees, merge/cut edges…
The Note clarifies some aspects of SNs definition and manipulation, especially
for, so-called, generalized SNs. The latter apply to hypergraphs.
It also proposes the first to my knowledge implementation of the algebraic
operations on NSs, extending to hypergraphs as well.
We apply our method to generate electric circuit parameter formulae, especially
for impedance. The Thesis described also my prototype. Not the Note. The
prototype was indeed already little outdated. In particular, it was still
stored on IBM punched cards.
The graph manipulations characterize many domains. Our results were
a tip of an iceberg, aiming at our specific application. Nevertheless, they
were also a generic graph management tool. As such, SNs seemed competitive to
the usual stuff, at least. Despite this potential, about nothing happened
since. I could not spot any citation. One likely reason: the Note is invisible
to main search engines. I guess publishing in French did not help. Also, the keyword
“structural number” seems highly popular with Civil Engineering in a different
sense. These references appear naturally first.
More generally, I could spot only a handful of related papers. One was a 1972 article from India on active electrical circuits. Then, a couple of Polish papers appeared in 80-90s. The authors applied manually the SNs to the mechanical engineering. So, a perspective of interesting work apparently still remains.
Digital
Processing of Phonocardiograms
Phonocardiogram (PCG) is the recording of heart sound
(murmur). The manual signal interpretation by the physician may diagnose grave
illnesses. Like for ECG. However, the former seems less popular then the
latter. Probably since the manual processing is lengthily. My goal was to help
it through digital filtering & automatic diagnosis. Using what was then
called pattern recognition & now machine learning, usually.
My work
constituted the Doctorat en Informatique in
INRIA. It was published later in renowned Medical Informatics journals (search
the publication list pointed to below). I demonstrated the prototype validating
both goals. The signal processing introduced to PCG processing the
interactive vocoding and FFT-based digital
filtering (FFT means Fast Fourier Transform, we recall). The automatic
classification through clustering used Benzecri’s factorial
analysis. I applied the apparatus to two dangerous heart anomalies: aortic valvular stenosis and subaortic idiopathic stenosis. The diagnosis took a
split of a second, instead of hours of strain to the physician eyes.
My English paper has 3 citations at Scholar. One of
those is also at MS Academic. I thought this result regretfully showing about
no interest in the field. I was dead wrong. While doing this update, found
dozens of follow ups. Including, US, Japan, an entire book & a 1995 survey
at NIH site, cited itself over 200 times. The effort even included some French
work not far from Paris around 2012. All folks continue along the path of vocoding & FFT filtering towards some
classification algorithm. Why all but three avoided then citing us? Feel free
to question the authors. In the meantime, the rules of scientific
honesty we’re raised with seem again the past.
Still, it did not seem that the bulk made it into the
commercial world. Perhaps the current interest in Internet of Things will help,
e.g., with distant patient monitoring. On anecdotic side, in early 80ties I got
a letter from HP. Kindly informing me about a brand new interactive FFT-based
signal filtering & analysis hardware & software box. The letter was
allegedly informing me given my work in the domain. This was no more for over
ten years. I understood HP somehow acknowledged the
parenthood. Anyhow, perhaps HP still sells the babies. Clustering and
more data analysis would be a useful addition if not done already.
Value Dates for Concurrency Control
A value date is the concept loved by banks, in routine
use. It is the time when a transaction must commits, neither earlier nor later.
For instance, your check deposit through an ATM becomes usually valid precisely
at 6 pm same business day. That is the time when the credit shows up at your
account and you can start spending it. The calculus of the interest the deposit
perhaps brings starts at the value date as well.
Our papers on this topic adapted the concept to the DB
transaction management. It looked particularly attractive for large distributed
ones. Including those reaching planetary scale, e.g., in Google’s Spanner. As
for banks, the value date was giving a predictable time for every
sub-transaction to end up and even to commit synchronously although in
isolation. The mechanism we still believe better than the popular eventual
consistency, without any such limit. Here is one of our later papers about,
from 2001. We shamefully forgot where it was published (pdf).
The value date idea caught attention of DB research and had a follow up. Not enough for the concept to enter the DB practice, to our knowledge. We believe, it still might and should. One reason was perhaps confusion with so called real-time transactions, proposed by our old pal Hector Garcia Molina & al. Hector unfortunately passed away already. That transaction also had a time limit, but was also obviously wished to commit at earliest. We had the pleasure to match both ideas as early as in 1988, when Hector invited me for the talk at Princeton. It seems a number of folks working later on transactions, missed this important difference.
For anecdote, I had a chance to have outstanding
co-authors. My gentle initial, H. Tirri, became later
the boss of Nokia Research, Palo Alto. Another friend, Y. Breitbart,
also ACM Fellow, regretfully already passed away as well. Finally, Avi Silberschatz, Prof. at Yale, when happens to be back from
Israel, does not need any introduction.
Computer Life
for Internet.
This work had for goal future organization of Internet. The proposal was to model it upon our own. The result would be an Internet of Beings, in modern wording. Beings should have their own life, could ultimately be aware of their existence and perhaps even imagine ours. We could appear them as God(s). Amazingly, the construction was recursive. Ourselves, we could be Beings in computers of our God(s) and so on.
One of the papers became an invited talk at Usenix. Another - became keynote for InfoJapan
90 (Information Technology Harmonizing With Society).
The latter conference had another keynote. That is how I made acquaintance with
Steve Jobs. His exciting talk was about NEXT. Soon after, NEXT was however a
thing of the past.
I met an unexpected number of colleagues who liked it.
The idea had thus a nice follow up. In turn, I liked the cooperation with the
coauthors. Nick Roussopoulos became since an ACM
Fellow. Gio Wiederhold
retired from Stanford, but only officially. Passed away unfortunately beginning
2023. Not surprisingly, the practical side of this research is yet to come. The
astonishing part was that I could not find citations of the InfoJapan
paper. Only the initial UMIACS Research Report got up to date nine citations on
Scholar. It is unfortunately the only visible at the Web. I manage to locate
the InfoJapan Proceedings only off-line in some US
University libraries and the Congress one. Amazon Student offers the book for
incredible price of $500 (Feb. 2016). All text in the book is under the
copyrights. I guess all this explains that. I found recently my text for InfoJapan prepared for the reviewers, who accepted it
provided some improvements. Here is at least the former one (pdf).
Amazingly, the 1st level of our model recursivity, i.e., that we are perhaps ourselves in some computer(s), appeared serious to astronomers. See there and elsewhere on the Web. Understandably, the proposers do not seem aware of our work. All things considered, again I believe that a lot of interesting work still lies ahead.
Relations with
Inherited Attributes
Codd’s relational model has two types of relations. A base relation with the stored attributes only. A view has only the inherited ones, from some base relations or views. This dichotomy may appear awkward. A single concept of a relation with stored and/or inherited attributes suffices. The popular wisdom says indeed: who/what can more, can less.
We made this proposal and analyzed some consequences for an RDB definition and manipulation, in this 1992 HPL Report (pdf