1. David Walker Princeton University Computer Science Pads: Simplified Data Processing For Scientists

2. 2 Computer Science in the 21st Century One part computation to determine the answer to your problem. One part communication to tell someone about it.

3. Who: actress Jennifer Aniston and actor Brad Pitt When: July 29, 2000 Where: The nuptials took place on the grounds of TV producer Marcy Carsey's Malibu estate The Ceremony: As the sun sank low in the California sky, two hundred assembled guests watched as John Aniston, known to daytime television fans for his work on Days of Our Lives, walked his daughter down the aisle. Shielded by a flower-bedecked canopy, the bride and groom were able to say....

4. 4

5. 5 Our Common Communication Infrastructure Behind the scenes, much of this information is represented in standardized data formats Standardized data formats: – Web pages in HTML – Pictures in JPEG – Movies in MPEG – “Universal” information format XML – Standard relational database formats A plethora of data processing tools: – Visualizers (Browsers Display JPEG, HTML,...) – Query languages allow users extract information (SQL, XQuery) – Programmers get easy access through standard libraries Java XML libraries --- JAXP – Many applications handle it natively and convert back and forth MS Word

6. 6 Ad Hoc Data Massive amounts of data are stored in XML, HTML or relational databases but there’s even more data that isn’t An ad hoc data format is any nonstandard data format for which convenient parsing, querying, visualizing, transformation tools are not available – ad hoc data is everywhere.

7. 7 Ad Hoc data from www.investors.com Date: 3/21/2005 1:00PM PACIFIC Investor's Business Daily ® Stock List Name: DAVE Stock Company Price Price Volume EPS RS Symbol Name Price Change % Change % Change Rating Rating AET Aetna Inc 73.68 -0.22 0% 31% 64 93 GE General Electric Co 36.01 0.13 0% -8% 59 56 HD Home Depot Inc 37.99 -0.89 -2% 63% 84 38 IBM Intl Business Machines 89.51 0.23 0% -13% 66 35 INTC Intel Corp 23.50 0.09 0% -47% 39 33 Data provided by William O'Neil + Co., Inc. © 2005. All Rights Reserved. Investor's Business Daily is a registered trademark of Investor's Business Daily, Inc. Reproduction or redistribution other than for personal use is prohibited. All prices are delayed at least 20 minutes.

8. 8 Ad Hoc data from www.geneontology.org !autogenerated-by: DAG-Edit version 1.419 rev 3 !saved-by: gocvs !date: Fri Mar 18 21:00:28 PST 2005 !version: $Revision: 3.223 $ !type: % is_a is a !type: < part_of part of !type: ^ inverse_of inverse of !type: | disjoint_from disjoint from $Gene_Ontology ; GO:0003673 <biological_process ; GO:0008150 %behavior ; GO:0007610 ; synonym:behaviour %adult behavior ; GO:0030534 ; synonym:adult behaviour %adult feeding behavior ; GO:0008343 ; synonym:adult feeding behaviour % feeding behavior ; GO:0007631 %adult locomotory behavior ; GO:0008344 ;...

9. 9 Ad Hoc Data From Steve Kleinstein (Immune Response Simulation Data) 0812583260 (~6:0:0:0:0~1:0:0:0:1,1:1:0:0:0) 13772160 (~6:0:0:0:0~1:1:0:0:0) 273762150 (~5:0:0:0:0~1:1:0:0:0) 351654320 (~2:0:0:0:0~1:1:0:0:0,1:1:0:0:0,1:0:0:1:0) 4816122110 (~1:0:0:0:0~1:0:0:1:0) 55271845134 (~13:0:0:0:0~2:0:0:0:1,1:0:0:1:0,2:0:0:1:0) 665051050 5:0:0:0:0....

10. 10 Ad Hoc Data in Chemistry O=C([C@@H]2OC(C)=O)[C@@]3(C)[C@]([C@](CO4) (OC(C)=O)[C@H]4C[C@@H]3O)([H])[C@H] (OC(C7=CC=CC=C7)=O)[C@@]1(O)[C@@](C)(C)C2=C(C) [C@@H](OC([C@H](O)[C@@H](NC(C6=CC=CC=C6)=O) C5=CC=CC=C5)=O)C1

11. 11 Ad Hoc Data from Web Server Logs (CLF) 207.136.97.49 - - [15/Oct/1997:18:46:51 -0700] "GET /tk/p.txt HTTP/1.0" 200 30 tj62.aol.com - - [16/Oct/1997:14:32:22 -0700] "POST /scpt/dd@grp.org/confirm HTTP/1.0" 200 941

12. 12 Ad Hoc Data: DNS packets 00000000: 9192 d8fb 8480 0001 05d8 0000 0000 0872...............r 00000010: 6573 6561 7263 6803 6174 7403 636f 6d00 esearch.att.com. 00000020: 00fc 0001 c00c 0006 0001 0000 0e10 0027...............' 00000030: 036e 7331 c00c 0a68 6f73 746d 6173 7465.ns1...hostmaste 00000040: 72c0 0c77 64e5 4900 000e 1000 0003 8400 r..wd.I......... 00000050: 36ee 8000 000e 10c0 0c00 0f00 0100 000e 6............... 00000060: 1000 0a00 0a05 6c69 6e75 78c0 0cc0 0c00......linux..... 00000070: 0f00 0100 000e 1000 0c00 0a07 6d61 696c............mail 00000080: 6d61 6ec0 0cc0 0c00 0100 0100 000e 1000 man............. 00000090: 0487 cf1a 16c0 0c00 0200 0100 000e 1000................ 000000a0: 0603 6e73 30c0 0cc0 0c00 0200 0100 000e..ns0........... 000000b0: 1000 02c0 2e03 5f67 63c0 0c00 2100 0100......_gc...!... 000000c0: 0002 5800 1d00 0000 640c c404 7068 7973..X.....d...phys 000000d0: 0872 6573 6561 7263 6803 6174 7403 636f.research.att.co

13. 13 Who uses ad hoc data? Ad hoc data sources are everywhere – containing valuable information of all kinds – everybody wants it: chemists, physicists, biologists, economists, computer scientists, network administrators,... just about anyone who writes their own programs

14. 14 The challenge of ad hoc data What can we do about ad hoc data? – how do we read it into programs? – how do we detect errors? – how do we correct errors? – how do we query it? – how do we view it? – how do we gather statistics on it? – how do we load it into a database? – how do we transform it into a standard format like XML? – how do we combine multiple ad data sources? – how do we filter, normalize and transform it? In short: how do we do all the things we take for granted when dealing with standard formats in a reliable, fault-tolerant and efficient, yet effortless way?

15. 15 Most people use C / Perl / Shell scripts But: – Writing hand-coded parsers is time consuming & error prone. – Reading and maintaining them in the face of even small format changes can be difficult. – Such programs are often incomplete, particularly with respect to errors. – Not all that efficient unless the author invests extra effort For reliable, fault-tolerant, efficient data processing, we can do better!

16. 16 Why not use traditional parsers? Overall, a very heavy-weight solution – people just do not do it – specifying a lexer and parser separately can be a barrier data specs as Lex and Yacc files are relatively complicated – lexing and parsing tools only solve a small part of the problem internal data structures built by hand printer by hand transforms by hand viewers by hand query engine by hand Error processing is fairly rigid We can do better!

17. 17 Enter Pads Pads: a system for Processing Ad hoc Data Sources Two main components: – a data description language for concise and precise specifications of ad hoc data formats and properties – a compiler that automatically generates a suite of data processing tools robust libraries for C programming – parser that flags all errors and automatically recovers – printing utilities an interface that allows users to query ad hoc data converter to XML a statistical profiler – collects stats on common values appearing in all parts of the data; records error stats visual interface & viewer (coming soon!)

18. 18 The rest of the talk Introduction to ad hoc data sources (check) Pads Tools Pads Language Pads Semantics Wrap-up

19. 19 Pads Tool Generation Architecture Pads Compiler Gene Ontology description Statistical Profiler Tool gene data Profile ACE 25% BKJ 25%... XML Formatter Tool gene data Viewer Tool gene data

20. 20 Pads Tool Generation Architecture Pads Compiler Gene Ontology description Gene Ontology Generated Parser Pads Base Library Gene Ontology Statistical Profiler Glue code for statistical profile

21. 21 Pads Programmer Tools Pads Compiler Gene Ontology description Gene Ontology Generated Parser Pads Base Library Ad Hoc User Program Ad Hoc User Program in C

22. 22 The Statistical Profiler Tool for each part of a data source, profiler reports errors & most common values. from example weblog data:.length : uint32 +++++++++++++++++++++++++++++++++++++++++++ good: 53544 bad: 3824 pcnt-bad: 6.666 min: 35 max: 248591 avg: 4090.234 top 10 values out of 1000 distinct values: tracked 99.552% of values val: 3082 count: 1254 %-of-good: 2.342 val: 170 count: 1148 %-of-good: 2.144 val: 43 count: 1018 %-of-good: 1.901.....

23. 23 The Statistical Profiler Tool ad hoc data is often poorly documented or out-of-date even the documentation of weblog data from our textbook was missing some information: good: 53544 bad: 3824 pcnt-bad: 6.666 – web server sometimes return a ‘-’ instead of length of bytes, which wasn’t mentioned in the textbook data descriptions can be written in a iterative fashion – use the profiler at each stage to uncover additional information about the data and refine the description

24. Pads Language

25. 25 PADS language Based on Type Theory – in most modern programming languages, types (int, bool, struct, object...) describe program data the source of most of my research – in Pads, types describe physical data formats, semantic properties of data, and a mapping into an internal program representation (ie, a parser) Can describe ASCII, binary, and mixed data formats.

26. 26 PADS language Basic Types – Rich and extensible. – Pint8, Puint8, Pint16,... – Pstring(:term-char:) – Pstring_FW(:size:) – Pstring_ME(:reg_exp:) – Pdate,... Supports user-defined compound types to describe data source structure: – Pstruct, Parray, Punion, Ptypedef, Penum

27. 27 Example: CLF web log Common Log Format from Web Protocols and Practice. (Bala and Rexford) Fields: – IP address of remote host – Remote identity (usually ‘-’ to indicate name not collected) – Authenticated user (usually ‘-’ to indicate name not collected) – Time associated with request – Request – Response code – Content length 207.136.97.50 - - [15/Oct/1997:18:46:51 -0700] "GET /turkey/amnty1.gif HTTP/1.0" 200 3013

28. 28 Example: Pstruct Pstruct http_weblog { host client; /- Client requesting service ' '; auth_id remoteID; /- Remote identity ' '; auth_id auth; /- Name of authenticated user “ [”; Pdate(:']':) date; /- Timestamp of request “] ”; http_request request; /- Request ' '; Puint16_FW(:3:) response; /- 3-digit response code ' '; Puint32 contentLength; /- Bytes in response }; 207.136.97.50 - - [15/Oct/1997:18:46:51 -0700] "GET /turkey/amnty1.gif HTTP/1.0" 200 3013 For reading a sequence of different data elements:

29. 29 Example: Punion Punion auth_id { Pchar unavailable : unavailable == '-'; Pstring(:' ':) id; }; Union declarations allow the user to describe variations. Implementation tries branches in order. Stops when it finds a branch whose constraints are all true. 207.136.97.50 - - [15/Oct/1997:18:46:51 -0700] "GET /turkey/amnty1.gif HTTP/1.0" 200 3013

30. 30 Example: Parray Parray nIP { Puint8[4]: Psep(‘.’) && Pterm(‘ ’); }; Array declarations allow the user to specify: Size (fixed, lower-bounded, upper-bounded, unbounded.) Boolean-valued constraints Psep and Pterm predicates Array terminates upon exhausting EOF/EOR, reaching terminator, or reaching maximum size. 207.136.97.50 - - [15/Oct/1997:18:46:51 -0700] "GET /turkey/amnty1.gif HTTP/1.0" 200 3013

31. 31 Example: User constraints int checkVersion(http_v version, method_t meth) { if ((version.major == 1) && (version.minor == 0)) return 1; if ((meth == LINK) || (meth == UNLINK)) return 0; return 1; } Pstruct http_request { '\"'; method_t meth; /- Request method ' '; Pstring(:' ':) req_uri; /- Requested uri. ' '; http_v version : checkVersion(version, meth); /- HTTP version number of request '\"'; }; 207.136.97.50 - - [15/Oct/1997:18:46:51 -0700] "GET /turkey/amnty1.gif HTTP/1.0" 200 3013

32. 32 Example: Parameterization & Dependency “Early” data often affects parsing of later data: – Lengths of sequences – Branches of switched unions To accommodate this usage, we allow PADS types to be parameterized: Pstruct packet_t (: Puint32 length:) {... Pstring_FW(: length :) payload; };

33. Pads Semantics

34. 34 Semantics: The Big Picture As a theorist, I want to be able describe the meanings (semantics) of programs and programming languages Why bother? What is the point? – communication spread ideas, techniques and algorithms often means extracting the essence of a language and reducing it to a simple set of mathematical relations – verification prove properties of implementations particularly security-relevant or safety-critical applications – generalization the mathematics brings out the central principles and invariants leads to more general, compositional, scalable solutions – it’s just fun immensely satisfying to come up with the perfect formal system where all parts compose and blend seemlessly together

35. 35 Semantics for Pads: Goals – Communication Pads descriptions can be incorporated into just about any language. ML? Java? Perl? Matlab? Language designers need a precise specification to do so – Verification In some cases, we find the implementation incomplete or making arbitrary choices (eg: error correction semantics) Every once in awhile, the implementation is outright wrong (eg: array semantics) – Generalization Semantics allows us to compare and contrast Pads with related languages & add features (eg: intersection types & overlays from PacketTypes; recursive types; more)

36. 36 Semantics for Pads: Overview Pads is large language and if we tried to formalize the whole thing right from the get-go, we wouldn’t succeed – we’d get lost in details and make mistakes – we’d be unable to structure our proofs of key properties – we wouldn’t communicate the essential elements to our fellow researchers Strategy: – pick out the key ingredients & eliminate the ugly, but unimportant details – develop an idealized version of the real language each type in our idealized version of pads represents a single, simple pure idea each type composes with all others we give a semantics to each individual construct; we get a semantics for complex objects by putting several simple individual ones together

37. 37 Semantics for Pads: Overview Part 1: Specify idealized (abstract) syntax of types T ::= True(parse nothing successfully) | False(parse nothing unsuccessfully) | {x:T | P(x)} (constrained type; parse data as T and check P) | C (arg) (parse parameterized base type; eg: string(:’ ‘:)) | T1  T2(union type; parse one or the other) | T1  T2(intersection type; parse data as both T1 and T2) |  x:T1.T2(dependent pair; parse T1, call it x, then parse T2) | T seq(arg)(sequence type; parse Ts until finding arg) | x.T (type parameterized by argument x) | T (arg)(parameterized type applied to argument) | hide T (skip data described by T; eg: absorb ‘|’ ) | spoof (arg)(parse nothing; add arg to internal representation) basics

38. 38 Semantics for Pads: Overview Part 1: Specify idealized (abstract) syntax of types T ::= True(parse nothing successfully) | False(parse nothing unsuccessfully) | {x:T | P(x)} (constrained type; parse data as T and check P) | C (arg) (parse parameterized base type; eg: string(:’ ‘:)) | T1  T2(union type; parse one or the other) | T1  T2(intersection type; parse data as both T1 and T2) |  x:T1.T2(dependent pair; parse T1, call it x, then parse T2) | T seq(arg)(sequence type; parse Ts until finding arg) | x.T (type parameterized by argument x) | T (arg)(parameterized type applied to argument) | hide T (skip data described by T; eg: absorb ‘|’ ) | spoof (arg)(parse nothing; add arg to internal representation) basics structured types

39. 39 Semantics for Pads: Overview Part 1: Specify idealized (abstract) syntax of types T ::= True(parse nothing successfully) | False(parse nothing unsuccessfully) | {x:T | P(x)} (constrained type; parse data as T and check P) | C (arg) (parse parameterized base type; eg: string(:’ ‘:)) | T1  T2(union type; parse one or the other) | T1  T2(intersection type; parse data as both T1 and T2) |  x:T1.T2(dependent pair; parse T1, call it x, then parse T2) | T seq(arg)(sequence type; parse Ts until finding arg) | x.T (type parameterized by argument x) | T (arg)(parameterized type applied to argument) | hide T (skip data described by T; eg: absorb ‘|’ ) | spoof (arg)(parse nothing; add arg to internal representation) basics structured types para- meterized types

40. 40 Semantics for Pads: Overview Part 1: Specify idealized (abstract) syntax of types T ::= True(parse nothing successfully) | False(parse nothing unsuccessfully) | {x:T | P(x)} (constrained type; parse data as T and check P) | C (arg) (parse parameterized base type; eg: string(:’ ‘:)) | T1  T2(union type; parse one or the other) | T1  T2(intersection type; parse data as both T1 and T2) |  x:T1.T2(dependent pair; parse T1, call it x, then parse T2) | T seq(arg)(sequence type; parse Ts until finding arg) | x.T (type parameterized by argument x) | T (arg)(parameterized type applied to argument) | absorb T (skip data described by T; eg: absorb ‘|’ ) | compute (arg)(parse nothing; add arg to internal representation) basics structured types para- meterized types transforms

41. 41 Semantics for Pads: Overview Part 2: Specify denotational semantics of types – in general, a denotational semantics describes one language (poorly understood) in terms of another language (better understood) – in our case, we specify the meaning of Pads types (poorly understood) in terms of the polymorphic -calculus (better understood, at least by me) semantics(T) = bits.e a parser function mapping external bits to data structures in the -calculus

42. 42 Semantics for Pads: Overview Part 3: Prove Pads has the required properties – Theorem: Parsers never generate “bad” internal representations of external data. ie, representations are well-typed in the implementation language. – Theorem: Parsers check all semantic constraints.

43. Wrap-up

44. 44 Challenges of Ad Hoc Data Revisited Data arrives “as is” – Format determined by data source, not consumers. The Pads language allows consumers to describe data in just about any format. – Often has little documentation. A Pads description can serve as documentation for data source. The statistical profiler helps analysts understand data. – Some percentage of data is “buggy.” Constraints allow consumers to express expectations about data. Parsers check for errors and say where errors are located. Ad hoc data is a rich source of information for chemists, biologists, computer scientists, if they could only get at it. – Pads generates a collection of useful tools automatically from data descriptions Pads is our answer to the challenge of ad hoc data sources.

45. 45 Related work DataScript [Back: CGSE 2002] & PacketTypes [McCann & Chandra: SIGCOMM 2000] – Primarily for networking data – Binary data formats only – Stop on first error – No value-added tools (Profiler; XML conversion; Query engine) – No semantics

46. 46 Current and Future Work Pads Language – recursion and pointers (eg: for tree- and graph-structured data) – integrated pre- and post-processing (eg: encryption, compression) – composition and reuse (via polymorphism, modules) – multi-source data integration Pads Compiler – parsing and querying optimization (eg: dealing with massive data sets) Pads Tools – new architecture for robust & reliable tool generation – application-specific customization error correction, data normalization, ignoring or rearranging components – general data transformation – visual interface for nonprogrammers Pads Applications – genomics data (with Olga Troyanskaya) – networking and telephony data (AT&T) – a great domain for interdisciplinary undergraduate research projects

47. 47 Pads Summary The overarching goal of Pads is to make understanding, querying and transforming ad hoc data an effortless task. We do so with new programming language technology based on the principles of Type Theory. AT&T Research: Kathleen Fisher Mary Fernandez Joel Gottlieb Robert Gruber (now Google) Ricardo Medel (summer intern) Princeton: Joe Kovba (UGrad) Yitzhak Mandelbaum (Grad) David Walker http://www.padsproj.org/

48. End!