Is there a perfect data-parallel programming language? (Experiments with Morel and Apache Calcite)

Is there a perfect
data-parallel
programming language?
Experiments with Morel and Apache Calcite
Julian Hyde @julianhyde (Looker/Google)
March 7th, 2022

Agenda
Programming language
types (and how they might be
unified)
Data parallel
(e.g. Flume,
Spark)
Relational
(e.g. SQL)
Functional
(e.g. Haskell,
Standard ML,
Morel)
Deductive
(e.g. Datalog)

Agenda
1. Data parallel
2. Functional
3. Functional + Relational
4. Functional + Data parallel
5. Functional + Deductive
Data parallel
(e.g. Flume,
Spark)
Relational
(e.g. SQL)
Functional
(e.g. Haskell,
Standard ML,
Morel)
Deductive
(e.g. Datalog)

1. Data parallel
Data parallel
(e.g. Flume,
Spark)
Relational
(e.g. SQL)
Functional
(e.g. Haskell,
Standard ML,
Morel)
Deductive
(e.g. Datalog)
Explained via two data parallel
problems (text indexing and
WordCount)

. .
.
Text files
(input)
Index files
(output)
Building a document index
a b c x y z

x y z
split split split
reduce reduce
reduce reduce
Text files
(input)
Index files
(output)
Building a document index
a b c

yellow: 12,
zebra: 9
WordCount
split split split
reduce reduce
reduce reduce
Text files
(input)
Word counts
(output)
aardvark: 3,
badger: 7
... ...

Data-parallel programming
Batch processing
Input is a large, immutable data sets
Processing can be processed in parallel pipelines
Significant chance of hardware failure during execution
Related programming models:
● Stream and incremental processing
● Deductive and graph programming

Data-parallel
framework
(e.g.
MapReduce,
FlumeJava,
Apache Spark)
(*) You write two small functions
fun wc_mapper line =
List.map (fn w => (w, 1)) (split line);
fun wc_reducer (key, values) =
List.foldl (fn (x, y) => x + y) 0 values;
(*) The framework provides mapReduce
fun mapReduce mapper reducer list = ...;
(*) Combine them to build your program
fun wordCount list = mapReduce wc_mapper wc_reducer list;

Data-parallel
framework
(e.g.
MapReduce,
FlumeJava,
Apache Spark)
SQL SELECT word, COUNT(*) AS c
FROM Documents AS d,
LATERAL TABLE (split(d.text)) AS word // requires a ‘split’ UDF
GROUP BY word;

Data-parallel
framework
(e.g.
MapReduce,
FlumeJava,
Apache Spark)
SQL SELECT word, COUNT(*) AS c
FROM Documents AS d,
LATERAL TABLE (split(d.text)) AS word // requires a ‘split’ UDF
GROUP BY word;
Morel from line in lines,
word in split line (*) requires ‘split’ function - see later...
group word compute c = count

What are our options?
Extend SQL Extend a functional programming
language
We’ll need to:
● Allow functions defined in
queries
● Add relations-as-values
● Add functions-as-values
● Modernize the type system
(adding type variables, function
types, algebraic types)
● Write an optimizing compiler
We’ll need to:
● Add syntax for relational operations
● Map onto external data
● Write a query optimizer
Nice stuff we get for free:
● Algebraic types
● Pattern matching
● Inline function and value declarations

Morel is a functional programming language. It is
derived from Standard ML, and is extended with
list comprehensions and other relational
operators. Like Standard ML, Morel has
parametric and algebraic data types with
Hindley-Milner type inference. Morel is
implemented in Java, and is optimized and
executed using a combination of techniques from
functional programming language compilers and
database query optimizers.

Evolution of functional
languages
Standard
ML
OCaml
Haskell
Scala
ML
Java
F#
C#
Lisp
Scheme
Clojure
1958
1975
1973
1983
1995
1995
1990
2000
2004
2005
2007

Extend Standard ML
Standard ML is strongly typed, and
can very frequently deduce every
type in a program.
Standard ML has record types.
(These are important for queries.)
Haven’t decided whether Morel is
eager (like Standard ML) or lazy
(like Haskell and SQL)
Haskell’s type system is more
powerful. So, Haskell programs
often have explicit types.
Standard
ML
Haskell
Scala
ML
Java
F#
C#
Morel
SQL
Lisp
Scheme
Clojure
OCaml
1975
2019

Morel themes
Early stage language
The target audience is SQL users
Less emphasis on abstract algebra
Program compilation, query optimization
Functional programming in-the-small, and in-the-large

2. Functional
Data parallel
(e.g. Flume,
Spark)
Relational
(e.g. SQL)
Functional
(e.g. Haskell,
Standard ML,
Morel)
Deductive
(e.g. Datalog)
A quick introduction to
Standard ML
All of the following examples work in both Standard ML
and Morel.

Standard ML: values and types
- "Hello, world!";
val it = "Hello, world!" : string
- 1 + 2;
val it = 3 : int
- ~1.5;
val it = ~1.5 : real
- [1, 1, 2, 3, 5];
val it = [1,1,2,3,5] : int list
- fn i => i mod 2 = 1;
val it = fn : int -> bool
- (1, "a");
val it = (1,"a") : int * string
- {name = "Fred", empno = 100};
val it = {empno=100,name="Fred"} : {empno:int, name:string}

Standard ML: variables and functions
- val x = 1;
val x = 1 : int
- val isOdd = fn i => i mod 2 = 1;
val isOdd = fn : int -> bool
- fun isOdd i = i mod 2 = 0;
val isOdd = fn : int -> bool
- isOdd x;
val it = true : bool
- let
= val x = 6
= fun isOdd i = i mod 2 = 1
= in
= isOdd x
= end;
val it = false : bool
val assigns a value to a variable.
fun declares a function.
● fun is syntactic sugar for
val ... = fn ... => ...
let allows you to make several declarations
before evaluating an expression.

- datatype 'a tree =
= EMPTY
= | LEAF of 'a
= | NODE of ('a * 'a tree * 'a tree);
Algebraic data types, case, and recursion
datatype declares an algebraic data type. 'a Is a
type variable and therefore 'a tree is a
polymorphic type.

= EMPTY
= | LEAF of 'a
- val t =
= NODE (1, LEAF 2, NODE (3, EMPTY, LEAF 7));
EMPTY
polymorphic type.
Define an instance of tree using its constructors
NODE, LEAF and EMPTY.
LEAF
2
NODE
1
NODE
3
LEAF
7

= EMPTY
= | LEAF of 'a
- val t =
- val rec sumTree = fn t =>
= case t of EMPTY => 0
= | LEAF i => i
= | NODE (i, l, r) =>
= i + sumTree l + sumTree r;
val sumTree = fn : int tree -> int
- sumTree t;
val it = 13 : int
polymorphic type.
case matches patterns,
deconstructing data types, and
binding variables as it goes.
sumTree use case and calls
itself recursively.
EMPTY
LEAF
2
NODE
1
NODE
3
LEAF
7

= EMPTY
= | LEAF of 'a
- val t =
- val rec sumTree = fn t =>
= case t of EMPTY => 0
= | LEAF i => i
= | NODE (i, l, r) =>
- sumTree t;
val it = 13 : int
- fun sumTree EMPTY = 0
= | sumTree (LEAF i) = i
= | sumTree (NODE (i, l, r)) =
polymorphic type.
case matches patterns,
deconstructing data types, and
binding variables as it goes.
sumTree use case and calls
itself recursively.
EMPTY
LEAF
2
NODE
1
NODE
3
LEAF
7
fun is a shorthand for
case and val rec.

Relations and higher-order functions
Relations are lists of records.
A higher-order function is a function whose
arguments or return value are functions.
Common higher-order functions that act on
lists:
● List.filter – equivalent to the Filter
relational operator (SQL WHERE)
● List.map – equivalent to Project
relational operator (SQL SELECT)
● flatMap – as List.map, but may
produce several output elements per
input element
#deptno is a system-generated function that
accesses the deptno field of a record.
- val emps = [
= {id = 100, name = "Fred", deptno = 10},
= {id = 101, name = "Velma", deptno = 20},
= {id = 102, name = "Shaggy", deptno = 30},
= {id = 103, name = "Scooby", deptno = 30}];
- List.filter (fn e => #deptno e = 30) emps;
val it = [{deptno=30,id=102,name="Shaggy"},
{deptno=30,id=103,name="Scooby"}]
: {deptno:int, id:int, name:string} list
SELECT *
FROM emps AS e
WHERE e.deptno = 30
Equivalent SQL

Implementing Join using higher-order functions
List.map function is equivalent to
Project relational operator (SQL
SELECT)
We also define a flatMap function.
- val depts = [
= {deptno = 10, name = "Sales"},
= {deptno = 20, name = "Marketing"},
= {deptno = 30, name = "R&D"}];
- fun flatMap f xs = List.concat (List.map f xs);
val flatMap =
fn : ('a -> 'b list) -> 'a list -> 'b list
- List.map
= (fn (d, e) => {deptno = #deptno d, name = #name e})
= (List.filter
= (fn (d, e) => #deptno d = #deptno e)
= (flatMap
= (fn e => (List.map (fn d => (d, e)) depts))
= emps));
val it =
[{deptno=10,name="Fred"}, {deptno=20,name="Velma"},
{deptno=30,name="Shaggy"}, {deptno=30,name="Scooby"}]
: {deptno:int, name:string} list
SELECT d.deptno, e.name
FROM emps AS e,
depts AS d
WHERE d.deptno = e.deptno
Equivalent SQL

3. Functional +
Relational
Data parallel
(e.g. Flume,
Spark)
Relational
(e.g. SQL)
Functional
(e.g. Haskell,
Standard ML,
Morel)
Deductive
(e.g. Datalog)
Morel

List.map
(fn (d, e) => {deptno = #deptno d, name = #name e})
(List.filter
(fn (d, e) => #deptno d = #deptno e)
(flatMap
(fn e => (List.map (fn d => (d, e)) depts))
emps));
Implementing Join in Morel using from
Morel extensions to Standard ML:
● from operator creates a list
comprehension
● x.field is shorthand for
#field x
● {#field} is shorthand for
{field = #field x}
FROM emps AS e,
depts AS d
Equivalent SQL

List.map
(List.filter
(flatMap
emps));
from e in emps,
d in depts
where #deptno e = #deptno d
yield {deptno = #deptno d, name = #name e};
comprehension
#field x
{field = #field x}
FROM emps AS e,
depts AS d
Equivalent SQL

List.map
(List.filter
(flatMap
emps));
from e in emps,
d in depts
where #deptno e = #deptno d
yield {deptno = #deptno d, name = #name e};
from e in emps,
d in depts
where e.deptno = d.deptno
yield {d.deptno, e.name};
comprehension
#field x
{field = #field x}
FROM emps AS e,
depts AS d
Equivalent SQL

WordCount
let
fun split0 [] word words = word :: words
| split0 (#" " :: s) word words = split0 s "" (word :: words)
| split0 (c :: s) word words = split0 s (word ^ (String.str c)) words
fun split = List.rev (split0 (String.explode s) "" [])
in
from line in lines,
word in split line
group word compute c = count
end;

WordCount
- let
= fun split0 [] word words = word :: words
= | split0 (#" " :: s) word words = split0 s "" (word :: words)
= | split0 (c :: s) word words = split0 s (word ^ (String.str c)) words
= fun split = List.rev (split0 (String.explode s) "" [])
= in
= from line in lines,
= word in split line
= group word compute c = count
= end;
val wordCount = fn : string list -> {c:int, word:string} list
- wordCount ["a skunk sat on a stump",
= "and thunk the stump stunk",
= "but the stump thunk the skunk stunk"];
val it =
[{c=2,word="a"},{c=3,word="the"},{c=1,word="but"},
{c=1,word="sat"},{c=1,word="and"},{c=2,word="stunk"},
{c=3,word="stump"},{c=1,word="on"},{c=2,word="thunk"},
{c=2,word="skunk"}] : {c:int, word:string} list

- let
= in
= end;
val it =
WordCount
Functional
programming
“in the small”

WordCount
- let
= in
= end;
val it =
Functional
programming
“in the large”

Functions as views, functions as values
- fun emps2 () =
= from e in emps
= yield {e.id,
= e.name,
= e.deptno,
= comp = fn revenue => case e.deptno of
= 30 => e.id + revenue / 2
= | _ => e.id};
val emps2 = fn : unit -> {comp:int -> int,
deptno:int, id:int, name:string} list

emps2 is a function
that returns a
collection
- fun emps2 () =
= from e in emps
= yield {e.id,
= e.name,
= e.deptno,
= 30 => e.id + revenue / 2
= | _ => e.id};

- fun emps2 () =
= from e in emps
= yield {e.id,
= e.name,
= e.deptno,
= 30 => e.id + revenue / 2
= | _ => e.id};
The comp field is a
function value (in
fact, it’s a closure)
emps2 is a function
that returns a
collection

- fun emps2 () =
= from e in emps
= yield {e.id,
= e.name,
= e.deptno,
= 30 => e.id + revenue / 2
= | _ => e.id};
- fun emps2 () =
= from e in emps
= yield {e.id,
= e.name,
= e.deptno,
= 30 => e.id + revenue / 2
= | _ => e.id};
- from e in emps2 ()
= yield {e.name, e.id, c = e.comp 1000};
The comp field is a
function value (in
fact, it’s a closure)
emps2 is a function
that returns a
collection

- fun emps2 () =
= from e in emps
= yield {e.id,
= e.name,
= e.deptno,
= 30 => e.id + revenue / 2
= | _ => e.id};
- from e in emps2 ()
= yield {e.name, e.id, c = e.comp 1000};
val it =
[{c=100,id=100,name="Fred"},
{c=101,id=101,name="Velma"},
{c=602,id=102,name="Shaggy"},
{c=603,id=103,name="Scooby"}]
: {c:int, id:int, name:string} list

Chaining relational operators
- from e in emps
= order e.deptno, e.id desc
= yield {e.name, nameLength = String.size e.name, e.id, e.deptno}
= where nameLength > 4
= group deptno compute c = count, s = sum of nameLength
= where s > 10
= yield c + s;
val it = [14] : int list

Chaining relational operators - step 1
- from e in emps;
val it =
[{deptno=10,id=100,name="Fred"},
{deptno=20,id=101,name="Velma"},
{deptno=30,id=102,name="Shaggy"},
{deptno=30,id=103,name="Scooby"}]

- from e in emps
= order e.deptno, e.id desc;
val it =
[{deptno=10,id=100,name="Fred"},
{deptno=20,id=101,name="Velma"},
{deptno=30,id=103,name="Scooby"},
{deptno=30,id=102,name="Shaggy"}]

- from e in emps
= yield {e.name, nameLength = String.size e.name, e.id, e.deptno};
val it =
[{deptno=10,id=100,name="Fred",nameLength=4},
{deptno=20,id=101,name="Velma",nameLength=5},
{deptno=30,id=103,name="Scooby",nameLength=6},
{deptno=30,id=102,name="Shaggy",nameLength=6}]
: {deptno:int, id:int, name:string, nameLength:int} list

- from e in emps
= where nameLength > 4;
val it =
[{deptno=20,id=101,name="Velma",nameLength=5},
{deptno=30,id=103,name="Scooby",nameLength=6},
{deptno=30,id=102,name="Shaggy",nameLength=6}]
: {deptno:int, id:int, name:string, nameLength:int} list

- from e in emps
= group deptno compute c = count, s = sum of nameLength;
val it =
[{c=1,deptno=20,s=5},
{c=2,deptno=30,s=12}]
: {c:int, deptno:int, s:int} list

- from e in emps
= where s > 10;
val it =
[{c=2,deptno=30,s=12}]
: {c:int, deptno:int, s:int} list

= from e in emps
= where s > 10
= yield c + s;
val it = [14] : int list

Morel from e in emps
order e.deptno, e.id desc
yield {e.name, nameLength = String.size e.name, e.id, e.deptno}
where nameLength > 4
group deptno compute c = count, s = sum of nameLength
where s > 10
yield c + s;
SQL (almost
equivalent)
SELECT c + s
FROM (SELECT deptno, COUNT(*) AS c, SUM(nameLength) AS s
FROM (SELECT e.name, CHAR_LENGTH(e.ename) AS nameLength, e.id, e.deptno
FROM (SELECT *
FROM emps AS e
ORDER BY e.deptno, e.id DESC))
WHERE nameLength > 4
GROUP BY deptno
HAVING s > 10)
Java (very
approximately)
for (Emp e : emps) {
String name = e.name;
int nameLength = name.length();
int id = e.id;
int deptno = e.deptno;
if (nameLength > 4) {
...

4. Functional +
Data parallel
Data parallel
(e.g. Flume,
Spark)
Relational
(e.g. SQL)
Functional
(e.g. Haskell,
Standard ML,
Morel)
Deductive
(e.g. Datalog)
Integrating Morel with Apache
Calcite

Apache Calcite
Apache Calcite Server
JDBC server
JDBC client
SQL parser &
validator
Query
planner
Adapter
Pluggable
rewrite rules
Pluggable
stats / cost
Pluggable
catalog
Physical
operators
Storage
Relational
algebra
Toolkit for writing a DBMS
Many parts are optional or
pluggable
Relational algebra is the
core

SELECT d.name, COUNT(*) AS c
FROM Emps AS e
JOIN Depts AS d USING (deptno)
WHERE e.age > 50
GROUP BY d.deptno
HAVING COUNT(*) > 5
ORDER BY c DESC
Relational algebra
Based on set theory, plus
operators: Project, Filter,
Aggregate, Union, Join, Sort
Calcite provides:
● SQL to relational algebra
● Query planner
● Physical operators to execute
plan
● An adapter system to make
external data sources look
like tables
Scan [Emps] Scan [Depts]
Join [e.deptno = d.deptno]
Filter [e.age > 50]
Aggregate [deptno, COUNT(*) AS c]
Filter [c > 5]
Project [name, c]
Sort [c DESC]

Algebraic rewrite
Scan [Emps] Scan [Depts]
Join [e.deptno = d.deptno]
Filter [e.age > 50]
Aggregate [deptno, COUNT(*) AS c]
Filter [c > 5]
Project [name, c]
Sort [c DESC]
Calcite optimizes queries by
applying rewrite rules that
preserve semantics.
Planner uses dynamic
programming, seeking the
lowest total cost.
SELECT d.name, COUNT(*) AS c
FROM (SELECT * FROM Emps
WHERE e.age > 50) AS e
JOIN Depts AS d USING (deptno)
GROUP BY d.deptno
HAVING COUNT(*) > 5
ORDER BY c DESC

Integration with Apache Calcite - schemas
Expose Calcite schemas as named
records, each field of which is a table.
A default Morel connection has
variables foodmart and scott
(connections to hsqldb via Calcite’s
JDBC adapter).
Connections to other Calcite
adapters (Apache Cassandra, Druid,
Kafka, Pig, Redis) are also possible.
- foodmart;
val it = {account=<relation>, currency=<relation>,
customer=<relation>, ...} : ...;
- scott;
val it = {bonus=<relation>, dept=<relation>,
emp=<relation>, salgrade=<relation>} : ...;
- scott.dept;
val it =
[{deptno=10,dname="ACCOUNTING",loc="NEW YORK"},
{deptno=20,dname="RESEARCH",loc="DALLAS"},
{deptno=30,dname="SALES",loc="CHICAGO"},
{deptno=40,dname="OPERATIONS",loc="BOSTON"}]
: {deptno:int, dname:string, loc:string} list
- from d in scott.dept
= where notExists (from e in scott.emp
= where e.deptno = d.deptno
= andalso e.job = "CLERK");
val it =
[{deptno=40,dname="OPERATIONS",loc="BOSTON"}]

- Sys.set ("hybrid", true);
val it = () : unit
val it = [{deptno=40,dname="OPERATIONS",loc="BOSTON"}]
Integration with Apache Calcite - relational algebra
In “hybrid” mode, Morel’s compiler identifies
sections of Morel programs that are
relational operations and converts them to
Calcite relational algebra.
Calcite can them optimize these and
execute them on their native systems.

- Sys.set ("hybrid", true);
val it = () : unit
val it = [{deptno=40,dname="OPERATIONS",loc="BOSTON"}]
- Sys.plan();
val it = "calcite(plan
LogicalProject(deptno=[$0], dname=[$1], loc=[$2])
LogicalFilter(condition=[IS NULL($4)])
LogicalJoin(condition=[=($0, $3)], joinType=[left])
LogicalProject(deptno=[$0], dname=[$1], loc=[$2])
JdbcTableScan(table=[[scott, DEPT]])
LogicalProject(deptno=[$0], $f1=[true])
LogicalAggregate(group=[{0}])
LogicalProject(deptno=[$1])
LogicalFilter(condition=[AND(=($5, 'CLERK'), IS NOT NULL($1))])
LogicalProject(comm=[$6], deptno=[$7], empno=[$0],ename=[$1], hiredate=[$4], job=[$2]
JdbcTableScan(table=[[scott, EMP]]))" : string
Integration with Apache Calcite - relational algebra
In “hybrid” mode, Morel’s compiler identifies
sections of Morel programs that are
relational operations and converts them to
Calcite relational algebra.
Calcite can them optimize these and
execute them on their native systems.

Optimization
Relational query optimization
● Applies to relational operators (~10 core
operators + UDFs) not general-purpose
code
● Transformation rules that match patterns
(e.g. Filter on Project)
● Decisions based on cost & statistics
● Hybrid plans - choose target engine, sort
order
● Materialized views
● Macro-level optimizations
Functional program optimization
● Applies to typed lambda calculus
● Inline constant lambda expressions
● Eliminate dead code
● Defend against mutually recursive groups
● Careful to not inline expressions that are
used more than once (increasing code
size) or which bind closures more often
(increasing running time)
● Micro-level optimizations
[GM] “The Volcano Optimizer Generator: Extensibility and Efficient Search” (Graefe, McKenna 1993)
[PJM] “Secrets of the Glasgow Haskell Compiler inliner” (Peyton Jones, Marlow 2002)

WordCount Data parallel Relational
(Apache Calcite)
Functional
(Morel)
Concisely specified in a functional
programming language (Morel)
Translated into relational algebra (Apache
Calcite)
Executed in the data-parallel framework of
your choice
from line in lines,
word in split line
group word compute c = count;

5. Functional +
Deductive
Data parallel
(e.g. Flume,
Spark)
Relational
(e.g. SQL)
Functional
(e.g. Haskell,
Standard ML,
Morel)
Deductive
(e.g. Datalog)
An elegant way to recursively
define relations in a functional
programming language*
*Designed but not yet implemented

# SQL
SELECT *
FROM parents
WHERE parent = ‘elrond’;
parent child
====== =======
elrond arwen
elrond elladan
elrond elrohir
# Morel
from (parent, child) in parents
where parent = “elrond”;
[(“elrond”, “arwen”),
(“elrond”, “elladan”),
(“elrond”, “elrohir”)];
# Datalog
is_parent(earendil, elrond).
is_parent(elrond, arwen).
is_parent(elrond, elladan).
is_parent(elrond, elrohir).
answer(X) :- is_parent(elrond, X).
X = arwen
X = elladan
X = elrohir
Parents (a base relation)

# SQL
CREATE VIEW ancestors AS
WITH RECURSIVE a AS (
SELECT parent AS ancestor,
child AS descendant
FROM parents
UNION ALL
SELECT a.ancestor, p.child
FROM parents AS p
JOIN a ON a.descendant = p.parent)
SELECT * FROM a;
SELECT * FROM ancestors
WHERE descendant = ‘arwen’;
ancestor descendant
======== =======
earendil arwen
elrond arwen
# Morel
fun ancestors () =
(from (x, y) in parents)
union
(from (x, y) in parents,
(y2, z) in ancestors ()
where y = y2
yield (x, z));
from (ancestor, descendant) in ancestors ()
where descendant = “arwen”;
Uncaught exception: StackOverflowError
# Datalog
is_ancestor(X, Y) :- is_parent(X, Y).
is_ancestor(X, Y) :- is_parent(X, Z),
is_ancestor(Z, Y).
answer(X) :- is_ancestor(X, arwen).
Ancestors (a recursively-defined relation)

# SQL
CREATE VIEW ancestors AS
WITH RECURSIVE a AS (
SELECT parent AS ancestor,
child AS descendant
FROM parents
UNION ALL
SELECT a.ancestor, p.child
FROM parents AS p
JOIN a ON a.descendant = p.parent)
SELECT * FROM a;
SELECT * FROM ancestors
WHERE descendant = ‘arwen’;
ancestor descendant
======== =======
earendil arwen
elrond arwen
# Morel
fun ancestors () =
(from (x, y) in parents)
union
(from (x, y) in parents,
(y2, z) in ancestors ()
where y = y2
yield (x, z));
from (ancestor, descendant) in ancestors ()
where descendant = “arwen”;
Uncaught exception: StackOverflowError
# Datalog
is_ancestor(X, Y) :- is_parent(X, Y).
is_ancestor(X, Y) :- is_parent(X, Z),
is_ancestor(Z, Y).
answer(X) :- is_ancestor(X, arwen).
Ancestors (a recursively-defined relation)
“StackOverflowError”
is not easy to fix. It’s
hard to a write a
recursive function that
iterates until a set
reaches a fixed point.

# Morel “forwards” relation
# Relation defined using algebra
- fun clerks () =
= from e in emps
= where e.job = “CLERK”;
# Query uses regular iteration
- from e in clerks,
= d in depts
= where d.deptno = e.deptno
= andalso d.loc = “DALLAS”
= yield e.name;
val it =
[“SMITH”, “ADAMS”] : string list;
# Morel “backwards” relation
# Relation defined using a predicate
- fun isClerk e =
= e.job = “CLERK”;
# Query uses a mixture of constrained
# and regular iteration
- from e suchThat isClerk e,
= d in depts
= yield e.name;
val it =
Two ways to define a relation

# Morel “forwards” relation
# Relation defined using algebra
- fun clerks () =
= from e in emps
= where e.job = “CLERK”;
# Query uses regular iteration
- from e in clerks,
= d in depts
= yield e.name;
val it =
# Morel “backwards” relation
# Relation defined using a predicate
- fun isClerk e =
= e.job = “CLERK”;
# Query uses a mixture of constrained
# and regular iteration
- from e suchThat isClerk e,
= d in depts
= yield e.name;
val it =
Two ways to define a relation
“suchThat” keyword
converts a bool
function into a relation

# Morel
- fun isAncestor (x, z) =
= (x, z) elem parents
= orelse exists (
= from y suchThat isAncestor (x, y)
= andalso (y, z) elem parents);
- from a suchThat isAncestor (a, “arwen”);
val it = [“earendil”, “elrond”] : string list
- from d suchThat isAncestor (“earendil”, d);
val it = [“elrond”, “arwen”, “elladan”, “elrohir”] : string list
Recursively-defined predicate relation

Morel
Functional query language
Rich type system, concise as SQL, Turing complete
Combines (relational) query optimization with
(lambda) FP optimization techniques
Execute in Java interpreter and/or data-parallel
engines from e in emps,
d in depts
where e.deptno = d.deptno
yield {d.deptno, e.name};
from line in lines,
word in split line
group word compute c = count;
from a suchThat isAncestor (a, “arwen”);

Thank you!
Questions?
Julian Hyde, Looker/Google • @julianhyde
https://github.com/hydromatic/morel • @morel_lang
https://calcite.apache.org • @ApacheCalcite

Other operators
from in
where
yield
order … desc
group … compute
count max min sum
intersect union except
exists notExists elem notElem only

Compared to other languages
Haskell – Haskell comprehensions are more general (monads vs lists). Morel is
focused on relational algebra and probably benefits from a simpler type system.
Builder APIs (e.g. LINQ, FlumeJava, Cascading, Apache Spark) – Builder APIs
are two languages. E.g. Apache Spark programs are Scala that builds an
algebraic expression. Scala code (especially lambdas) is sprinkled throughout the
algebra. Scala compilation is not integrated with algebra optimization.
SQL – SQL’s type system does not have parameterized types, so higher order
functions are awkward. Tables and columns have separate namespaces, which
complicates handling of nested collections. Functions and temporary variables
cannot be defined inside queries, so queries are not Turing-complete (unless you
use recursive query gymnastics).

Standard ML: types
Type Example
Primitive types bool
char
int
real
string
unit
true: bool
#"a": char
~1: int
3.14: real
"foo": string
(): unit
Function types string -> int
int * int -> int
String.size
fn (x, y) => x + y * y
Tuple types int * string (10, "Fred")
Record types {empno:int, name:string} {empno=10, name="Fred"}
Collection types int list
(bool * (int -> int)) list
[1, 2, 3]
[(true, fn i => i + 1)]
Type variables 'a List.length: 'a list -> int

Functional programming ↔ relational programming
Functional
programming
in-the-small
- fun squareList [] = []
= | squareList (x :: xs) = x * x :: squareList xs;
val squareList = fn : int list -> int list
- squareList [1, 2, 3];
val it = [1,4,9] : int list
Functional
programming
in-the-large
- fun squareList xs = List.map (fn x => x * x) xs;
val squareList = fn : int list -> int list
Relational
programming
- fun squareList xs =
= from x in xs
= yield x * x;

wordCount again
wordCount
in-the-small
- fun wordCount list = ...;
val wordCount = fn : string list -> {count:int, word:string} list
wordCount
in-the-large using
mapReduce
- fun mapReduce mapper reducer list = ...;
val mapReduce = fn : ('a -> ('b * 'c) list) ->
('b * 'c list -> 'd) -> 'a list -> ('b * 'd) list
- fun wc_mapper line =
= List.map (fn w => (w, 1)) (split line);
val wc_mapper = fn : string -> (string * int) list
- fun wc_reducer (key, values) =
= List.foldl (fn (x, y) => x + y) 0 values;
val wc_reducer = fn : 'a * int list -> int
- fun wordCount list = mapReduce wc_mapper wc_reducer list;
val wordCount = fn : string list -> {count:int, word:string} list
Relational
implementation of
mapReduce
- fun mapReduce mapper reducer list =
= from e in list,
= (k, v) in mapper e
= group k compute c = (fn vs => reducer (k, vs)) of v;

group …. compute
- fun median reals = ...;
val median = fn : real list -> real
- from e in emps
= group x = e.deptno mod 3,
= e.job
= compute c = count,
= sum of e.sal,
= m = median of e.sal + e.comm;
val it = {c:int, job:string, m:real, sum:real, x.int} list

https://github.com/hydromatic/morel/discussions/106
Join the discussion!

LucidDB
C++
Calcite evolution - origins as an SMP DB
JDBC server
JDBC client
Physical
operators
Rewrite rules
Catalog
Storage & data
structures
SQL parser &
validator
Query
planner
Relational
algebra
Java

Optiq
Calcite evolution - pluggable components
JDBC server
JDBC client
Physical
operators
Rewrite rules
SQL parser &
validator
Query
planner
Relational
algebra

Optiq
Calcite evolution - pluggable components
JDBC server
JDBC client
SQL parser &
validator
Query
planner
Adapter
Pluggable
rewrite rules
Pluggable
stats / cost
Pluggable
catalog
Physical
operators
Storage
Relational
algebra

Apache Calcite
Calcite evolution - separate JDBC stack
Avatica
JDBC server
JDBC client
Pluggable
rewrite rules
Pluggable
stats / cost
Pluggable
catalog
ODBC client
Adapter
Physical
operators
Storage
SQL parser &
validator
Query
planner
Relational
algebra

Apache Calcite
Calcite evolution - federation via adapters
Pluggable
rewrite rules
Pluggable
stats / cost
Pluggable
catalog
Adapter
Physical
operators
Storage
SQL parser &
validator
Query
planner
Relational
algebra
SQL

Apache Calcite
JDBC adapter
Pluggable
rewrite rules
Pluggable
stats / cost
Enumerable
adapter
MongoDB
adapter
File adapter
(CSV, JSON, Http)
Apache Kafka
adapter
Apache Spark
adapter
Pluggable
catalog
SQL
SQL parser &
validator
Query
planner
Relational
algebra

Apache Calcite
Pluggable
rewrite rules
Pluggable
stats / cost
Enumerable
adapter
Pluggable
catalog
SQL
SQL parser &
validator
Query
planner
Relational
algebra

Apache Calcite
JDBC adapter
Pluggable
rewrite rules
Pluggable
stats / cost
Pluggable
catalog
SQL
SQL parser &
validator
Query
planner
Relational
algebra

Apache Calcite
Calcite evolution - SQL dialects
Pluggable
rewrite rules
Pluggable parser, lexical,
conformance, operators
Pluggable
SQL dialect
SQL
SQL
SQL parser &
validator
Query
planner
Relational
algebra
JDBC adapter

Apache Calcite
Calcite evolution - other front-end languages
SQL
Adapter
Physical
operators
Storage
SQL parser &
validator
Query
planner
Relational
algebra

Calcite evolution - other front-end languages
Pig
RelBuilder
Adapter
Physical
operators
Morel
Storage
Query
planner
Relational
algebra
Datalog
SQL parser &
validator
SQL

Is there a perfect data-parallel programming language? (Experiments with Morel and Apache Calcite)

More Related Content

What's hot

Similar to Is there a perfect data-parallel programming language? (Experiments with Morel and Apache Calcite)

More from Julian Hyde

Recently uploaded

Is there a perfect data-parallel programming language? (Experiments with Morel and Apache Calcite)