Apache Drill - Why, What, How

© MapR Technologies, confidential
© 2014 MapR Technologies
M. C. Srivas, CTO and Founder

MapR is Unbiased Open Source

Linux Is Unbiased
• Linux provides choice
– MySQL
– PostgreSQL
– SQLite
• Linux provides choice
– Apache httpd
– Nginx
– Lighttpd

MapR Is Unbiased
• MapR provides choice
MapR Distribution for Hadoop Distribution C Distribution H
Spark Spark (all of it) and SparkSQL Spark only No
Interactive SQL Impala, Drill, Hive/Tez,
SparkSQL
One option
(Impala)
One option
(Hive/Tez)
Scheduler YARN, Mesos One option
(YARN)
One option
(YARN)
Versions Hive 0.10, 0.11, 0.12, 0.13
Pig 0.11, 012
HBase 0.94, 0.98
One version One version

MapR Distribution for Apache Hadoop
MapR Data Platform
(Random Read/Write)
Enterprise Grade Data Hub Operational
MapR-FS
(POSIX)
MapR-DB
(High-Performance NoSQL)
Security
YARN
Pig
Cascading
Spark
Batch
Spark
Streaming
Storm*
Streaming
HBase
Solr
NoSQL &
Search
Juju
Provisioning
&
Coordination
Savannah*
Mahout
MLLib
ML, Graph
GraphX
MapReduc
e v1 & v2
APACHE HADOOP AND OSS ECOSYSTEM
EXECUTION ENGINES DATA GOVERNANCE AND OPERATIONS
Workflow
& Data
Tez* Governance
Accumulo*
Hive
Impala
Shark
Drill*
SQL
Sqoop Sentry* Oozie ZooKeeper
Flume Knox* Falcon* Whirr
Data
Integration
& Access
HttpFS
Hue
NFS HDFS API HBase API JSON API
MapR Control System
(Management and Monitoring)
* In Roadmap for inclusion/certification
CLI REST API GUI

Hadoop an augmentation for EDW—Why?

But inside, it looks like this …

And this …

Consolidating schemas is very hard.

Consolidating schemas is very hard, causes SILOs

Silos make analysis very difficult
• How do I identify a
unique {customer,
trade} across data sets?
• How can I guarantee
the lack of anomalous
behavior if I can’t see
all data?

© 2014 MapR Technologies 19
Hard to know what’s of value a priori

Hard to know what’s of value a priori

Why Hadoop

Rethink SQL for Big Data
Preserve
•ANSI SQL
• Familiar and ubiquitous
• Performance
• Interactive nature crucial for BI/Analytics
• One technology
• Painful to manage different technologies
• Enterprise ready
• System-of-record, HA, DR, Security, Multi-tenancy,
…

Rethink SQL for Big Data
Preserve
•ANSI SQL
• Familiar and ubiquitous
• Performance
• Interactive nature crucial for BI/Analytics
• One technology
• Painful to manage different technologies
• Enterprise ready
• System-of-record, HA, DR, Security, Multi-tenancy,
…
Invent
• Flexible data-model
• Allow schemas to evolve rapidly
• Support semi-structured data types
• Agility
• Self-service possible when developer and DBA is
same
• Scalability
• In all dimensions: data, speed, schemas, processes,
management

SQL is here to stay

Hadoop is here to stay

YOU CAN’T HANDLE REAL SQL

SQL
select * from A
where exists (
select 1 from B where B.b < 100 );
• Did you know Apache HIVE cannot compute it?
– eg, Hive, Impala, Spark/Shark

Self-described Data
select cf.month, cf.year
from hbase.table1;
• Did you know normal SQL cannot handle the above?
• Nor can HIVE and its variants like Impala, Shark?

Self-described Data
select cf.month, cf.year
from hbase.table1;
• Why?
• Because there’s no meta-store definition available

Self-Describing Data Ubiquitous
Centralized schema
- Static
- Managed by the DBAs
- In a centralized repository
Long, meticulous data preparation process (ETL,
create/alter schema, etc.)
– can take 6-18 months
Self-describing, or schema-less, data
- Dynamic/evolving
- Managed by the applications
- Embedded in the data
Less schema, more suitable for data that has
higher volume, variety and velocity
Apache Drill

A Quick Tour through Apache Drill

Data Source is in the Query
select timestamp, message
from dfs1.logs.`AppServerLogs/2014/Jan/p001.parquet`
where errorLevel > 2

This is a cluster in Apache Drill
- DFS
- HBase
- Hive meta-store

- DFS
- HBase
- Hive meta-store
A work-space
- Typically a sub-directory
- HIVE database

- DFS
- HBase
- Hive meta-store
A work-space
- Typically a sub-directory
- HIVE database
A table
- pathnames
- Hbase table
- Hive table

Combine data sources on the fly
• JSON
• CSV
• ORC (ie, all Hive types)
• Parquet
• HBase tables
• … can combine them
Select USERS.name, USERS.emails.work
from
dfs.logs.`/data/logs` LOGS,
dfs.users.`/profiles.json` USERS,
where
LOGS.uid = USERS.uid and
errorLevel > 5
order by count(*);

Can be an entire directory tree
// On a file
select errorLevel, count(*)
from dfs.logs.`/AppServerLogs/2014/Jan/part0001.parquet`
group by errorLevel;

// On a file
// On the entire data collection: all years, all months
from dfs.logs.`/AppServerLogs`
group by errorLevel

// On a file
group by errorLevel
dirs[1] dirs[2]

// On a file
group by errorLevel
where dirs[1] > 2012
, dirs[2]
dirs[1] dirs[2]

Querying JSON
{ name: classic
fillings: [
{ name: sugar cal: 400 }]}
{ name: choco
fillings: [
{ name: sugar cal: 400 }
{ name: chocolate cal: 300 }]}
{ name: bostoncreme
fillings: [
{ name: cream cal: 1000 }
{ name: jelly cal: 600 }]}
donuts.json

Cursors inside Drill
DrillClient drill = new DrillClient().connect( …);
ResultReader r = drill.runSqlQuery( "select * from `donuts.json`");
while( r.next()) {
String donutName = r.reader( “name").readString();
ListReader fillings = r.reader( "fillings");
while( fillings.next()) {
int calories = fillings.reader( "cal").readInteger();
if (calories > 400)
print( donutName, calories, fillings.reader( "name").readString());
}
}
{ name: classic
fillings: [
{ name: choco
fillings: [
{ name: bostoncreme
fillings: [

Direct queries on nested data
// Flattening maps in JSON, parquet and other
nested records
select name, flatten(fillings) as f
from dfs.users.`/donuts.json`
where f.cal < 300;
// lists the fillings < 300 calories
{ name: classic
fillings: [
{ name: choco
fillings: [
{ name: bostoncreme
fillings: [

Complex Data Using SQL or Fluent API
// SQL
Result r = drill.sql( "select name, flatten(fillings) from
`donuts.json` where fillings.cal < 300`);
// or Fluent API
Result r = drill.table(“donuts.json”)
.lt(“fillings.cal”, 300).all();
while( r.next()) {
String name = r.get( “name").string();
List fillings = r.get( “fillings”).list();
while(fillings.next()) {
print(name, calories, fillings.get(“name”).string());
}
}
{ name: classic
fillings: [
{ name: choco
fillings: [
{ name: plain: 280 }]}
{ name: bostoncreme
fillings: [

Queries on embedded data
// embedded JSON value inside column donut-json inside column-family cf1
of an hbase table donuts
select d.name, count( d.fillings),
from (
select convert_from( cf1.donut-json, json) as d
from hbase.user.`donuts` );

Queries inside JSON records
// Each JSON record itself can be a whole database
// example: get all donuts with at least 1 filling with > 300 calories
max(d.fillings.cal) within record as mincal
from ( select convert_from( cf1.donut-json, json) as d
from hbase.user.`donuts` )
where mincal > 300;

a
• Schema can change over course of query
• Operators are able to reconfigure themselves on schema
change events
– Minimize flexibility overhead
– Support more advanced execution optimization based on actual data
characteristics

De-centralized metadata
// count the number of tweets per customer, where the customers are in Hive, and
their tweets are in HBase. Note that the hbase data has no meta-data information
select c.customerName, hb.tweets.count
from hive.CustomersDB.`Customers` c
join hbase.user.`SocialData` hb
on c.customerId = convert_from( hb.rowkey, UTF-8);

So what does this all mean?

A Drill Database
• What is a database with Drill/MapR?

A Drill Database
• Just a directory, with a bunch of related files

A Drill Database
• Just a directory, with a bunch of related files
• There’s no need for artificial boundaries
– No need to bunch a set of tables together to call it a “database”

A Drill Database
/user/srivas/work/bugs
symptom version date bugid dump-name
impala crash 3.1.1 14/7/14 12345 cust1.tgz
cldb slow 3.1.0 12/7/14 45678 cust2.tgz
BugList Customers
name rep se dump-name
xxxx dkim junhyuk cust1.tgz
yyyy yoshi aki cust2.tgz

Queries are simple
select b.bugid, b.symptom, b.date
from dfs.bugs.’/Customers’ c, dfs.bugs.’/BugList’ b
where c.dump-name = b.dump-name

Queries are simple
select b.bugid, b.symptom, b.date
from dfs.bugs.’/Customers’ c, dfs.bugs.’/BugList’ b
where c.dump-name = b.dump-name
Let’s say I want to cross-reference against your list:
select bugid, symptom
from dfs.bugs.’/Buglist’ b, dfs.yourbugs.’/YourBugFile’ b2
where b.bugid = b2.xxx

What does it mean?

What does it mean?
• No ETL
• Reach out directly to the particular table/file
• As long as the permissions are fine, you can do it
• No need to have the meta-data
– None needed

Another example
• convert_from( xx, json) invokes the json parser inside Drill

Another example
• What if you could plug in any parser?

Another example
• What if you could plug in any parser
– XML?
– Semi-conductor yield-analysis files? Oil-exploration readings?
– Telescope readings of stars?
– RFIDs of various things?

No ETL
• Basically, Drill is querying the raw data directly
• Joining with processed data
• NO ETL
• Folks, this is very, very powerful
• NO ETL

Seamless integration with Apache Hive
• Low latency queries on Hive tables
• Support for 100s of Hive file formats
• Ability to reuse Hive UDFs
• Support for multiple Hive metastores in a single query

Underneath the Covers

Basic Process
Zookeeper
DFS/HBase DFS/HBase DFS/HBase
Drillbit
Distributed Cache
Drillbit
Distributed Cache
Drillbit
Distributed Cache
Query 1. Query comes to any Drillbit (JDBC, ODBC, CLI, protobuf)
2. Drillbit generates execution plan based on query optimization & locality
3. Fragments are farmed to individual nodes
4. Result is returned to driving node
c c c

Stages of Query Planning
Parser
Logical
Planner
Physical
Planner
Query
Foreman
Plan
fragments
sent to drill
bits
SQL
Query
Heuristic and
cost based
Cost based

Query Execution
SQL Parser
Optimizer
Scheduler
Pig Parser
Physical Plan
Mongo
Cassandra
HiveQL Parser
RPC Endpoint
Distributed Cache
Storage Engine Interface
OOppereartaotorsr s
Foreman
Logical Plan
HDFS
HBase
JDBC Endpoint ODBC Endpoint

A Query engine that is…
• Columnar/Vectorized
• Optimistic/pipelined
• Runtime compilation
• Late binding
• Extensible

Columnar representation
A B C D E
A
B
C
D
On disk
E

Columnar Encoding
• Values in a col. stored next to one-another
– Better compression
– Range-map: save min-max, can skip if not
present
• Only retrieve columns participating in
query
• Aggregations can be performed without
decoding
A
B
C
D
On disk
E

Run-length-encoding & Sum
• Dataset encoded as <val> <run-length>:
– 2, 4 (4 2’s)
– 8, 10 (10 8’s)
• Goal: sum all the records
• Normally:
– Decompress: 2, 2, 2, 2, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8
– Add: 2 + 2 + 2 + 2 + 8 + 8 + 8 + 8 + 8 + 8 + 8 + 8 + 8 + 8
• Optimized work: 2 * 4 + 8 * 10
– Less memory, less operations

Bit-packed Dictionary Sort
• Dataset encoded with a dictionary and bit-positions:
– Dictionary: [Rupert, Bill, Larry] {0, 1, 2}
– Values: [1,0,1,2,1,2,1,0]
• Normal work
– Decompress & store: Bill, Rupert, Bill, Larry, Bill, Larry, Bill, Rupert
– Sort: ~24 comparisons of variable width strings
• Optimized work
– Sort dictionary: {Bill: 1, Larry: 2, Rupert: 0}
– Sort bit-packed values
– Work: max 3 string comparisons, ~24 comparisons of fixed-width dictionary bits

Drill 4-value semantics
• SQL’s 3-valued semantics
– True
– False
– Unknown
• Drill adds fourth
– Repeated

Vectorization
• Drill operates on more than one record at a time
– Word-sized manipulations
– SIMD-like instructions
• GCC, LLVM and JVM all do various optimizations automatically
– Manually code algorithms
• Logical Vectorization
– Bitmaps allow lightning fast null-checks
– Avoid branching to speed CPU pipeline

Runtime Compilation is Faster
• JIT is smart, but
more gains with
runtime
compilation
• Janino: Java-based
Java compiler
From http://bit.ly/16Xk32x

Drill compiler
Loaded class
Merge byte-code
of the two
classes
Janino compiles
runtime
byte-code
CodeModel
generates code
Precompiled
byte-code
templates

Optimistic
0
20
40
60
80
100
120
140
160
Speed vs. check-pointing
No need to checkpoint
Apache Drill Checkpoint frequently

Optimistic Execution
• Recovery code trivial
– Running instances discard the failed query’s intermediate state
• Pipelining possible
– Send results as soon as batch is large enough
– Requires barrier-less decomposition of query

Batches of Values
• Value vectors
– List of values, with same schema
– With the 4-value semantics for each value
• Shipped around in batches
– max 256k bytes in a batch
– max 64K rows in a batch
• RPC designed for multiple replies to a request

Pipelining
• Record batches are pipelined
between nodes
– ~256kB usually
• Unit of work for Drill
– Operators works on a batch
• Operator reconfiguration happens
at batch boundaries
DrillBit
DrillBit DrillBit

Pipelining Record Batches
SQL Parser
Optimizer
Scheduler
Pig Parser
Physical Plan
Mongo
Cassandra
HiveQL Parser
RPC Endpoint
Distributed Cache
Storage Engine Interface
OOppereartaotorsr s
Foreman
Logical Plan
HDFS
HBase
JDBC Endpoint ODBC Endpoint

DISK
Pipelining
• Random access: sort without copy or
restructuring
• Avoids serialization/deserialization
• Off-heap (no GC woes when lots of memory)
• Full specification + off-heap + batch
– Enables C/C++ operators (fast!)
• Read/write to disk
– when data larger than memory
Drill Bit
Memory
overflow
uses disk

Cost-based Optimization
• Using Optiq, an extensible framework
• Pluggable rules, and cost model
• Rules for distributed plan generation
• Insert Exchange operator into physical plan
• Optiq enhanced to explore parallel query plans
• Pluggable cost model
– CPU, IO, memory, network cost (data locality)
– Storage engine features (HDFS vs HIVE vs HBase)
Query
Optimizer
Pluggable
rules
Pluggable
cost model

Distributed Plan Cost
• Operators have distribution property
• Hash, Broadcast, Singleton, …
• Exchange operator to enforce distributions
• Hash: HashToRandomExchange
• Broadcast: BroadcastExchange
• Singleton: UnionExchange, SingleMergeExchange
• Enumerate all, use cost to pick best
• Merge Join vs Hash Join
• Partition-based join vs Broadcast-based join
• Streaming Aggregation vs Hash Aggregation
• Aggregation in one phase or two phases
• partial local aggregation followed by final aggregation
HashToRandomExchange
Sort
Streaming-Aggregation
Data Data Data

Drill 1.0 Hive 0.13 w/ Tez Impala 1.x
Latency Low Medium Low
Files Yes (all Hive file formats,
plus JSON, Text, …)
Yes (all Hive file formats) Yes (Parquet, Sequence,
…)
HBase/MapR-DB Yes Yes, perf issues Yes, with issues
Schema Hive or schema-less Hive Hive
SQL support ANSI SQL HiveQL HiveQL (subset)
Client support ODBC/JDBC ODBC/JDBC ODBC/JDBC
Hive compat High High Low
Large datasets Yes Yes Limited
Nested data Yes Limited No
Concurrency High Limited Medium
Interactive SQL-on-Hadoop options

Apache Drill Roadmap
•Low-latency SQL
•Schema-less execution
•Files & HBase/M7 support
•Hive integration
•BI and SQL tool support via
ODBC/JDBC
Data exploration/ad-hoc queries
1.0
•HBase query speedup
•Nested data functions
•Advanced SQL
functionality
Advanced analytics and
operational data
1.1
•Ultra low latency queries
•Single row
insert/update/delete
•Workload management
Operational SQL
2.0

Apache Drill Resources
• Drill 0.5 released last week
• Getting started with Drill is easy
– just download tarball and start running SQL queries on local files
• Mailing lists
– drill-user@incubator.apache.org
– drill-dev@incubator.apache.org
• Docs: https://cwiki.apache.org/confluence/display/DRILL/Apache+Drill+Wiki
• Fork us on GitHub: http://github.com/apache/incubator-drill/
• Create a JIRA: https://issues.apache.org/jira/browse/DRILL

Active Drill Community
• Large community, growing rapidly
– 35-40 contributors, 16 committers
– Microsoft, Linked-in, Oracle, Facebook, Visa, Lucidworks,
Concurrent, many universities
• In 2014
– over 20 meet-ups, many more coming soon
– 3 hackathons, with 40+ participants
• Encourage you to join, learn, contribute and have fun …

Drill at MapR
• World-class SQL team, ~20 people
• 150+ years combined experience building commercial
databases
• Oracle, DB2, ParAccel, Teradata, SQLServer, Vertica
• Team works on Drill, Hive, Impala
• Fixed some of the toughest problems in Apache Hive

Thank you!
M. C. Srivas
srivas@mapr.com
Did I mention we are hiring…

Apache Drill - Why, What, How

More Related Content

What's hot (20)

Similar to Apache Drill - Why, What, How (20)

Recently uploaded (20)

Apache Drill - Why, What, How