Lazy beats Smart and Fast

Lazy beats Smart & Fast
Julian Hyde | DataEngConf SF
2018/04/17

@julianhyde
SQL
Query planning
Query federation
OLAP
Streaming
Hadoop
ASF member
Original author of Apache Calcite
PMC Apache Arrow, Calcite, Drill, Eagle, Kylin
Architect at Looker

A “simple” query
Data
● 2010 U.S. census
● 100 million records
● 1KB per record
● 100 GB total
System
● 4x SATA 3 disks
● Total read throughput 1 GB/s
Query
Goal
● Compute the answer to the query in
under 5 seconds
SELECT SUM(householdSize)
FROM CensusHouseholds;

Solutions
Sequential scan Query takes 100 s (100 GB at 1 GB/s)
Parallelize Spread the data over 40 disks in 10 machines
Query takes 10 s
Cache Keep the data in memory
2nd query: 10 ms
3rd query: 10 s
Materialize Summarize the data on disk
All queries: 100 ms
Materialize +
cache + adapt
As above, building summaries on demand

Lazy > Smart + Fast
(Lazy + adaptive is even better)

Overview
How do you tune a data system? How can (or should) a data system tune itself?
What problems have we solved to bring these things to Apache Calcite?
Part 1: Strategies for organizing data
Part 2: How to make systems self-organizing?

SELECT d.name, COUNT(*) AS c
FROM Emps AS e
JOIN Depts AS d USING (deptno)
WHERE e.age < 40
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
Requires: declarative language (SQL),
query planner
Original goal: data independence
Enables: query optimization, new
algorithms and data structures
Scan [Emps] Scan [Depts]
Join [e.deptno = d.deptno]
Filter [e.age < 30]
Aggregate [deptno, COUNT(*) AS c]
Filter [c > 5]
Project [name, c]
Sort [c DESC]

Apache Calcite
Apache top-level project
Query planning framework used in many
projects and products
Also works standalone: embedded federated
query engine with SQL / JDBC front end
Apache community development model
http://calcite.apache.org
http://github.com/apache/calcite

Ways of organizing data
Format (CSV, JSON, binary)
Layout: row- vs. column-oriented (e.g. Parquet, ORC), cache friendly (e.g. Arrow)
Storage medium (disk, flash, RAM, NVRAM, ...)
Non-lossy copy: sorted / partitioned
Lossy copies of data: project, filter, aggregate, join
Combinations of the above
Logical optimizations >> physical optimizations

Index
A sorted, projected materialized
view
Accelerates queries that use
ranges, correlated lookups, sorting,
aggregate, distinct
CREATE TABLE Emp (empno INT,
name VARCHAR(20), deptno INT);
CREATE INDEX I_Emp_Deptno
ON Emp (deptno, name);
SELECT DISTINCT deptno FROM Emp
WHERE deptno BETWEEN 20 AND 40
ORDER BY deptno;
empno name deptno
100 Fred 20
110 Barney 10
120 Wilma 30
130 Dino 10
deptno name rowid
10 Barney af5634.0001
10 Dino af5634.0003
20 Fred af5634.0000
30 Wilma af5634.0002

Add the remaining columns
No longer need “rowid”
Lossless
During planning, treat indexes
as tables, and index lookups
as joins
Covering index
empno name deptno
100 Fred 20
110 Barney 10
120 Wilma 30
130 Dino 10
deptno name empno
10 Barney 100
10 Dino 130
20 Fred 20
30 Wilma 30
CREATE INDEX I_Emp_Deptno2 (
deptno INTEGER,
name VARCHAR(20))
COVER (empno);

Materialized view
CREATE MATERIALIZED
VIEW EmpsByDeptno AS
SELECT deptno, name, deptno
FROM Emp
ORDER BY deptno, name;
Scan [Emps]
Scan [EmpsByDeptno]
Sort [deptno, name]
empno name deptno
100 Fred 20
110 Barney 10
120 Wilma 30
130 Dino 10
deptno name empno
10 Barney 100
10 Dino 130
20 Fred 20
30 Wilma 30
As a materialized view, an
index is now just another
table
Several tables contain the
information necessary to
answer the query - just pick
the best

Spatial query
Find all restaurants within 1.5 distance units of
where I am:
restaurant x y
Zachary’s pizza 3 1
King Yen 7 7
Filippo’s 7 4
Station burger 5 6
SELECT *
FROM Restaurants AS r
WHERE ST_Distance(
ST_MakePoint(r.x, r.y),
ST_MakePoint(6, 7)) < 1.5
•
•
•
•
Zachary’s
pizza
Filippo’s
King
Yen
Station
burger

Hilbert space-filling curve
● A space-filling curve invented by mathematician David Hilbert
● Every (x, y) point has a unique position on the curve
● Points near to each other typically have Hilbert indexes close together

•
•
•
•
Add restriction based on h, a restaurant’s distance
along the Hilbert curve
Must keep original restriction due to false positives
Using Hilbert index
restaurant x y h
Zachary’s pizza 3 1 5
King Yen 7 7 41
Filippo’s 7 4 52
Station burger 5 6 36
Zachary’s
pizza
Filippo’s
SELECT *
FROM Restaurants AS r
WHERE (r.h BETWEEN 35 AND 42
OR r.h BETWEEN 46 AND 46)
AND ST_Distance(
ST_MakePoint(r.x, r.y),
ST_MakePoint(6, 7)) < 1.5
King
Yen
Station
burger

Telling the optimizer
1. Declare h as a generated column
2. Sort table by h
Planner can now convert spatial range
queries into a range scan
Does not require specialized spatial
index such as r-tree
Very efficient on a sorted table such as
HBase
CREATE TABLE Restaurants (
restaurant VARCHAR(20),
x DOUBLE,
y DOUBLE,
h DOUBLE GENERATED ALWAYS AS
ST_Hilbert(x, y) STORED)
SORT KEY (h);
restaurant x y h
Zachary’s pizza 3 1 5
Station burger 5 6 36
King Yen 7 7 41
Filippo’s 7 4 52

Much valuable data is “data in flight”
Use SQL to query streams (or streams + tables)
Streaming
Data center
SELECT AVG(unitPrice)
FROM Orders
WHERE units > 1000
AND orderDate
BETWEEN ‘2014-06-01’
AND ‘2015-12-31’
SELECT STREAM *
FROM Orders
WHERE units > 1000
Streaming query
Historic query

Hybrid query combines a stream with its
own history
● Orders is used as both as stream
and as “stream history” virtual table
● “Average order size over last year”
should be maintained by the system,
i.e. a materialized view
SELECT STREAM *
FROM Orders AS o
WHERE units > (
SELECT AVG(units)
FROM Orders AS h
WHERE h.productId = o.productId
AND h.rowtime
> o.rowtime - INTERVAL ‘1’ YEAR)
“Orders” used
as a stream
“Orders” used as
a “stream history”
virtual table

Summary - data optimization via
materialized views
Many forms of data optimization can be modeled as materialized views:
● Blocks in cache
● B-tree indexes
● Summary tables
● Spatial indexes
● History of streams
Allows the optimizer to “understand” the optimization and use it (if beneficial)
But who designs the optimizations?

How do data systems learn?
queries
DML
statistics
adaptations
recommender
Goals ● Improve response time, throughput, storage cost
● Predictable, adaptive (short and long term), allow human
intervention
How? ● Humans
● Adaptive systems
● Smart algorithms
Example
adaptations
● Cache disk blocks in memory
● Cached query results
● Data organization, e.g. partition on a different key
● Secondary structures, e.g. b-tree and r-tree indexes

Tiled, in-memory materialized views
A vision for an adaptive data system (we’re not there yet)
tables on
disk
in-memory
materializations
SELECT x, SUM(n) FROM t GROUP BY x

Building materialized views
Challenges:
● Design Which materializations to create?
● Populate Load them with data
● Maintain Incrementally populate when data changes
● Rewrite Transparently rewrite queries to use materializations
● Adapt Design and populate new materializations, drop unused ones
● Express Need a rich algebra, to model how data is derived
Initial focus: summary tables (materialized views over star schemas)

CREATE LATTICE Sales AS
SELECT t.*, c.*, COUNT(*), SUM(s.units)
FROM Sales AS s
JOIN Time AS t USING (timeId)
JOIN Customers AS c USING (customerId)
JOIN Products AS p USING (productId);
Designing summary tables via lattices
CREATE MATERIALIZED VIEW SalesYearZipcode AS
SELECT t.year, c.state, c.zipcode,
COUNT(*), SUM(units)
FROM Sales AS s
GROUP BY 1, 2, 3;
product
product
class
sales
customers
time

Many possible
summary
tables
Key
z zipcode (43k)
s state (50)
g gender (2)
y year (5)
m month (12)
() 1
(z, s, g, y,
m) 912k
(s, g, y,
m) 6k
(z) 43k (s) 50 (g) 2 (y) 5 (m) 12
raw 1m
(y, m) 60(g, y) 10
(z, s)
43.4k
(g, y, m)
120
Fewer than you would
expect, because 5m
combinations cannot
occur in 1m row table
Fewer than you
would expect,
because state
depends on zipcode

Algorithm: Design summary tables
Given a database with 30 columns, 10M rows. Find X summary tables with under
Y rows that improve query response time the most.
AdaptiveMonteCarlo algorithm [1]:
● Based on research [2]
● Greedy algorithm that takes a combination of summary tables and tries to
find the table that yields the greatest cost/benefit improvement
● Models “benefit” of the table as query time saved over simulated query load
● The “cost” of a table is its size
[1] org.pentaho.aggdes.algorithm.impl.AdaptiveMonteCarloAlgorithm
[2] Harinarayan, Rajaraman, Ullman (1996). “Implementing data cubes efficiently”

Lattice (optimized) () 1
(z, s, g, y,
m) 912k
(s, g, y,
m) 6k
(z) 43k (s) 50 (g) 2 (y) 5 (m) 12
(z, g, y,
m) 909k
(z, s, y,
m) 831k
raw 1m
(z, s, g,
m) 644k
(z, s, g,
y) 392k
(y, m) 60
(z, s)
43.4k
(z, s, g)
83.6k
(g, y) 10
(g, y, m)
120
(g, m)
24
Key
z zipcode (43k)
s state (50)
g gender (2)
y year (5)
m month (12)

Data profiling
Algorithm needs count(distinct a, b, ...) for each combination of attributes:
● Previous example had 25 = 32 possible tables
● Schema with 30 attributes has 230 (about 109) possible tables
● Algorithm considers a significant fraction of these
● Approximations are OK
Attempts to solve the profiling problem:
1. Compute each combination: scan, sort, unique, count; repeat 230 times!
2. Sketches (HyperLogLog)
3. Sketches + parallelism + information theory [CALCITE-1616]

Sketches
HyperLogLog is an algorithm that computes
approximate distinct count. It can estimate
cardinalities of 109 with a typical error rate of
2%, using 1.5 kB of memory. [3][4]
With 16 MB memory per machine we can
compute 10,000 combinations of attributes
each pass.
So, we’re down from 109 to 105 passes.
[3] Flajolet, Fusy, Gandouet, Meunier (2007). "Hyperloglog: The analysis of a near-optimal cardinality estimation algorithm"
[4] https://github.com/mrjgreen/HyperLogLog

Given Expected cardinality Actual cardinality Surprise
(gender): 2 (state): 50 (gender, state): 100.0 100 0.000
(month): 12 (zipcode): 43,000 (month, zipcode): 441,699.3 442,700 0.001
(state): 50 (zipcode): 43,000 (state, zipcode): 799,666.7 43,400 0.897
(state, zipcode): 43,400
(gender, state): 100
(gender, zipcode): 85,995
(gender, state, zipcode): 86,799
= min(86,799, 892,234, 892,228)
83,567 0.019
● Surprise = abs(actual - expected) / (actual + expected)
● E(card (x, y)) = n . (1 - ((n - 1) / n) ^ p) n = card (x) * card (y), p = row count
Combining probability & information theory

Algorithm
Three ways “surprise” can help:
● If a cardinality is not
surprising, we don’t need to
store it -- we can derive it
● If a combination’s cardinality
is not surprising, it is unlikely
to have surprising children
● If we’re not seeing surprising
results, it’s time to stop
surprise_threshold := 1
queue := {singleton combinations} // (a), (b), ...
while queue is not empty {
batch := remove first 10,000 entries in queue
compute cardinality of each combination in batch
for each actual (computed) cardinality a {
e := expected cardinality of combination
s := surprise(a, e)
if s > surprise_threshold {
store combination and its cardinality
add child combinations to queue // (x, a), (x, b),
...
}
increase surprise_threshold
}
}

Algorithm progress and “surprise” threshold
Progress of algorithm
Rejected as not
sufficiently
surprising
Surprise
threshold rises
as algorithm
progresses
Singleton
combinations are
have surprise = 1
Surprise
threshold rises
after we have
completed the
first batch

Data profiling - summary
The algorithm defeats a combinatorial search space using sketches +
information theory + parallelism
Recommending data structures is an optimization problem; profiling provides
the cost & benefit function
As a by-product, the algorithm discovers unique keys, “almost” keys, and foreign
keys
But which tables are actually joined together in practice?

FROM Sales AS s
FROM Sales AS s
GROUP BY 1, 2, 3;
product
product
class
sales
customers
time
The lattice generates the
summary tables. But who
writes the lattice?
Designing summary tables via lattices (2)

FROM Sales AS s
FROM Sales AS s
GROUP BY 1, 2, 3;
ALTER SCHEMA Sales
INFER LATTICES;
product
product
class
sales
customers
time
Designing summary tables via lattices (3)

Lattice after Query 1 + 2
Query 2
Query 1
Growing and evolving
lattices based on queries
sales
customers
product
product
class
sales
product
product
class
sales
customers
See: [CALCITE-1870] “Lattice suggester”

Summary
Learning systems = manual tuning + adaptive + smart algorithms
Query history + data profiling→ lattices → summary tables
We have discussed summary tables (materialized views based on
join/aggregate in a star schema) but the approach can be applied to other kinds
of materialized views
Relational algebra, incorporating materialized views, is a powerful language that
allows us to combine many forms of data optimization

Thank you! Questions?
@julianhyde | @ApacheCalcite | http://apache.calcite.org
Resources
[CALCITE-1616] Data profiler
[CALCITE-1870] Lattice suggester
[CALCITE-1861] Spatial indexes
[CALCITE-1968] OpenGIS
[CALCITE-1991] Generated columns
Talk: “Data profiling with Apache Calcite” (Hadoop Summit, 2017)
Talk: “SQL on everything, in memory” (Strata, 2014)
Zhang, Qi, Stradling, Huang (2014). “Towards a Painless Index for Spatial Objects”
Harinarayan, Rajaraman, Ullman (1996). “Implementing data cubes efficiently”
Image credit
https://www.flickr.com/photos/defenceimages/6938469933/

Architecture
Conventional database Calcite

Planning queries
MySQL
Splunk
join
Key: productId
group
Key: productName
Agg: count
filter
Condition:
action = 'purchase'
sort
Key: c desc
scan
scan
Table: products
select p.productName, count(*) as c
from splunk.splunk as s
join mysql.products as p
on s.productId = p.productId
where s.action = 'purchase'
group by p.productName
order by c desc
Table: splunk

Optimized query
MySQL
Splunk
join
Key: productId
group
Key: productName
Agg: count
filter
Condition:
action = 'purchase'
sort
Key: c desc
scan
scan
Table: splunk
Table: products
select p.productName, count(*) as c
from splunk.splunk as s
join mysql.products as p
on s.productId = p.productId
where s.action = 'purchase'
group by p.productName
order by c desc

Calcite framework
Cost, statistics
RelOptCost
RelOptCostFactory
RelMetadataProvider
• RelMdColumnUniquensss
• RelMdDistinctRowCount
• RelMdSelectivity
SQL parser
SqlNode
SqlParser
SqlValidator
Transformation rules
RelOptRule
• FilterMergeRule
• AggregateUnionTransposeRule
• 100+ more
Global transformations
• Unification (materialized view)
• Column trimming
• De-correlation
Relational algebra
RelNode (operator)
• TableScan
• Filter
• Project
• Union
• Aggregate
• …
RelDataType (type)
RexNode (expression)
RelTrait (physical property)
• RelConvention (calling-convention)
• RelCollation (sortedness)
• RelDistribution (partitioning)
RelBuilder
JDBC driver
Metadata
Schema
Table
Function
• TableFunction
• TableMacro
Lattice

Materialized views, lattices, tiles
Materialized view - A table whose contents are
guaranteed to be the same as executing a given query.
Lattice - Recommends, builds, and recognizes summary
materialized views (tiles) based on a star schema.
A query defines the tables and many:1 relationships in
the star schema.
Tile - A summary materialized view that belongs to a
lattice. A tile may or may not be materialized. Might be:
● Declared in lattice, or
● Generated via recommender algorithm, or
● Created in response to query.
CREATE MATERIALIZED VIEW t AS
SELECT * FROM emps
WHERE deptno = 10;
CREATE LATTICE star AS
SELECT *
FROM sales_fact_1997 AS s
JOIN product AS p ON …
JOIN product_class AS pc ON …
JOIN customer AS c ON …
JOIN time_by_day AS t ON …;
CREATE MATERIALIZED VIEW zg IN star
SELECT gender, zipcode, COUNT(*),
SUM(unit_sales) FROM star
GROUP BY gender, zipcode;

Combining past and future
select stream *
from Orders as o
where units > (
select avg(units)
from Orders as h
where h.productId = o.productId
and h.rowtime > o.rowtime - interval ‘1’ year)
➢ Orders is used as both stream and table
➢ System determines where to find the records
➢ Query is invalid if records are not available

Controlling when data is emitted
Early emission is the defining
characteristic of a streaming query.
The emit clause is a SQL extension
inspired by Apache Beam’s “trigger”
notion. (Still experimental… and
evolving.)
A relational (non-streaming) query is
just a query with the most conservative
possible emission strategy.
select stream productId,
count(*) as c
from Orders
group by productId,
floor(rowtime to hour)
emit at watermark,
early interval ‘2’ minute,
late limit 1;
select *
from Orders
emit when complete;

Other applications of data profiling
Query optimization:
● Planners are poor at estimating selectivity of conditions after N-way join
(especially on real data)
● New join-order benchmark: “Movies made by French directors tend to have
French actors”
● Predict number of reducers in MapReduce & Spark
“Grokking” a data set
Identifying problems in normalization, partitioning, quality

Further improvements to data profiling
● Build sketches in parallel
● Run algorithm in a distributed framework (Spark or MapReduce)
● Compute histograms
○ For example, Median age for male/female customers
● Seek out functional dependencies
○ Once you know FDs, a lot of cardinalities are no longer “surprising”
○ FDs occur in denormalized tables, e.g. star schemas
● Smarter criteria for stopping algorithm
● Skew/heavy hitters. Are some values much more frequent than others?
● Conditional cardinalities and functional dependencies
○ Does one partition of the data behave differently from others? (e.g. year=2005, state=LA)

Lazy beats Smart and Fast

More Related Content

What's hot (20)

Similar to Lazy beats Smart and Fast (20)

More from Julian Hyde (17)

Recently uploaded (20)

Lazy beats Smart and Fast

Editor's Notes