Slides for presentation on Cloudera Impala by Scott Leberknight of Near Infinity

1. Scott
Leberknight
Cloudera's
7/9/2013

2. History
lesson...

3. Google Map/Reduce
paper (2004)
Cutting & Cafarella
create Hadoop (2005)

4. Google Dremel paper (2010)
Facebook creates Hive (2007)*

5. Cloudera announces Impala
(October 2012)
HortonWorks' Stinger
(February 2013)
Apache Drill proposal
(August 2012)

6. * Hive => "SQL on Hadoop"
Write SQL queries
Translate into Map/Reduce job(s)
Convenient & easy
High-latency (batch processing)

7. What is Impala?
In-memory, distributed SQL
query engine (no Map/Reduce)
Native code (C++)
Distributed
(on HDFS data nodes)

8. Why Impala?
Interactive data analysis
Low-latency response
(roughly, 4 - 100x Hive)
Deploy on existing Hadoop clusters

9. Why Impala? (cont'd)
Data stored in HDFS avoids...
...duplicate storage
...data transformation
...moving data

10. Why Impala? (cont'd)
SPEED!

11. statestored & Hive metastore
(for database metadata)
Overview
impalad daemon runs on HDFS nodes
Queries run on "relevant" nodes
Supports common HDFS file formats
(for cluster metadata)

12. Overview (cont'd)
Does not use Map/Reduce
Not fault tolerant !
(query fails if any query on any node fails)
Submit queries via Hue/Beeswax
Thrift API, CLI, ODBC, JDBC

13. SQL Support
SELECT
Projection
UNION
INSERT OVERWRITE
INSERT INTO
ORDER BY
(w/ LIMIT)
Aggregation
Subqueries
(uncorrelated)
JOIN (equi-join only,
subject to memory
limitations)
(subset of Hive QL)

14. HBase Queries
Maps HBase tables via Hive
metastore mapping
Row key predicates => start/stop row
Non-row key predicates => SingleColumnValueFilter
HBase scan translations:

15. (Very) Unscientific Benchmarks

16. 9 queries, run in CDH Quickstart VM
Macbook Pro Retina, mid 2012
16GB RAM,
4GB for VM (VMWare 5),
Intel i7 2.6GHz quad-core processor
Hardware
No other load on system during queries
Pseudo-cluster + Impala daemons
CDH 4.2, Impala 1.0

17. Benchmarks (cont'd)
(from simple projection queries to
multiple joins, aggregation, multiple
predicates, and order by)
Impala vs. Hive performance
"TPC-DS" sample dataset
(http://www.tpc.org/tpcds/)

18. Query "A"
select
c.c_first_name,
c.c_last_name
from customer c
limit 50;

19. Query "B"
select
c.c_first_name,
c.c_last_name,
ca.ca_city,
ca.ca_county,
ca.ca_state
from customer c
join customer_address ca
on c.c_current_addr_sk = ca.ca_address_sk
limit 50;

20. Query "C"
select
c.c_first_name,
c.c_last_name,
ca.ca_city,
ca.ca_county,
ca.ca_state
from customer c
join customer_address ca
on c.c_current_addr_sk = ca.ca_address_sk
where lower(c.c_last_name) like 'smi%'
limit 50;

21. Query "D"
select distinct cd_credit_rating
from customer_demographics;

22. Query "E"
select
cd_credit_rating,
count(*)
from customer_demographics
group by cd_credit_rating;

23. Query "F"
select
c.c_first_name,
c.c_last_name,
ca.ca_city,
ca.ca_county,
ca.ca_state,
cd.cd_marital_status,
cd.cd_education_status
from customer c
join customer_address ca
on c.c_current_addr_sk = ca.ca_address_sk
join customer_demographics cd
on c.c_current_cdemo_sk = cd.cd_demo_sk
where
lower(c.c_last_name) like 'smi%' and
cd.cd_credit_rating in ('Unknown', 'High Risk')
limit 50;

24. Query "G"
select
count(c.c_customer_sk)
from customer c
join customer_address ca
on c.c_current_addr_sk = ca.ca_address_sk
join customer_demographics cd
on c.c_current_cdemo_sk = cd.cd_demo_sk
where
ca.ca_zip in ('20191', '20194') and
cd.cd_credit_rating in ('Unknown', 'High Risk');

25. Query "H"
select
c.c_first_name,
c.c_last_name,
ca.ca_city,
ca.ca_county,
ca.ca_state,
cd.cd_marital_status,
cd.cd_education_status
from customer c
join customer_address ca
on c.c_current_addr_sk = ca.ca_address_sk
join customer_demographics cd
on c.c_current_cdemo_sk = cd.cd_demo_sk
where
ca.ca_zip in ('20191', '20194') and
cd.cd_credit_rating in ('Unknown', 'High Risk')
limit 100;

26. select
i_item_id,
s_state,
avg(ss_quantity) agg1,
avg(ss_list_price) agg2,
avg(ss_coupon_amt) agg3,
avg(ss_sales_price) agg4
from store_sales
join date_dim
on (store_sales.ss_sold_date_sk = date_dim.d_date_sk)
join item
on (store_sales.ss_item_sk = item.i_item_sk)
join customer_demographics
on (store_sales.ss_cdemo_sk = customer_demographics.cd_demo_sk)
join store
on (store_sales.ss_store_sk = store.s_store_sk)
where
cd_gender = 'M' and
cd_marital_status = 'S' and
cd_education_status = 'College' and
d_year = 2002 and
s_state in ('TN','SD', 'SD', 'SD', 'SD', 'SD')
group by
i_item_id,
s_state
order by
i_item_id,
s_state
limit 100;
Query "TPC-DS"

27. Query Hive (sec) # M/R jobs Impala (sec) x Hive perf.
A 13.8 1 0.25 54
B 30.0 1 0.41 73
C 33.3 1 0.42 79
D 23.2 1 0.64 36
E 21.6 1 0.62 35
F 59.1 2 1.96 30
G 78.5 3 1.56 50
H 59.6 2 1.89 32
TPC-DS 204.5 6 3.23 63
(remember, unscientific...)

29. A
rchitecture

30. Two daemons
impalad
statestored
impalad on each HDFS data node
statestored - cluster metadata
Thrift APIs, ODBC, JDBC

31. impalad
Query execution
Query coordination
Query planning

32. impalad
Query Coordinator
Query Planner
Query Executor
HDFS DataNode
HBase RegionServer

33. Queries performed in-memory
Intermediate data never hits disk!
Data streamed to clients
C++
runtime code generation
intrinsics for optimization
Execution engine:

34. statestored
Cluster membership
Acts as a cluster
monitor
Not a SPOF
(single point of failure)

35. Metadata
Impala uses Hive metastore
Daemons cache metadata
REFRESH when table
definition/data change
Create tables in Hive or Impala

36. Next up - how queries work...

37. impalad
Query Coordinator
Query Planner
Query Executor
HDFS DataNode
HBase RegionServer
Client Statestore Hive Metastore
table/
database
metadata
SQL
query
impalad
Query Coordinator
Query Planner
Query Executor
HDFS DataNode
HBase RegionServer
impalad
Query Coordinator
Query Planner
Query Executor
HDFS DataNode
HBase RegionServer
cluster
monitoring

38. Read directly from disk
Short-circuit reads
Bypass HDFS DataNode
(avoids overhead of HDFS API)

39. impalad
Query Coordinator
Query Planner
Query Executor
HBase
Region
Server
HDFS
DataNode
Local Filesystem
Read
directly
from disk

41. Current Limitations
(as of version 1.0.1)
No join order optimization
No custom file formats, SerDes or UDFs
Limit required when using ORDER BY
Joins limited by aggregate memory of cluster
("put larger table on left")

42. Current Limitations
(as of version 1.0.1)
No advanced data structures
(arrays, maps, json, etc.)
Only basic DDL (otherwise do in Hive)
Limited file formats and compression
(though probably fine for most people)

43. Future...
Structure types (structs,
arrays, maps, json, etc.)
DDL support
Additional file formats &
compression support
"Performance"
Join optimization
(e.g. cost-based)
UDFs (???)
YARN integration
Fault-tolerance (???)

45. Dremel is a scalable, interactive ad-hoc
query system for analysis of read-only
nested data. By combining multi-level
execution trees and columnar data layout, it
is capable of running aggregation queries
over trillion-row tables in seconds. The
system scales to thousands of CPUs and
petabytes of data, and has thousands of
users at Google.
Comparing Impala to Dremel
- http://research.google.com/pubs/pub36632.html

46. Comparing Impala to Dremel
Impala = Dremel features circa 2010 + join
support, assuming columnar data format
(but, Google doesn't stand still...)
Dremel is production, mature
Basis for Google's BigQuery

47. Comparing Impala to Hive
Hive uses Map/Reduce -> high latency
Impala is in-memory, low-
latency query engine
Impala sacrifices fault tolerance
for performance

48. Comparing Impala to Drill
Apache Drill
Based on Dremel
In early stages...

49. "Apache Drill is an open-source software framework that supports
data-intensive distributed applications for interactive analysis of large-
scale datasets. Drill is the open source version of Google's Dremel
system which is available as an IaaS service called Google BigQuery. One
explicitly stated design goal is that Drill is able to scale to 10,000 servers
or more and to be able to process petabyes of data and trillions of
records in seconds. Currently, Drill is incubating at Apache."
- http://incubator.apache.org/drill/drill_overview.html
Comparing Impala to Drill

50. "The Stinger Initiative is a collection of
development threads in the Hive community
that will deliver 100X performance
improvements as well as SQL compatibility."
Comparing Impala to Stinger
- http://hortonworks.com/stinger/

51. Comparing Impala to Stinger
Stinger
Improve Hive performance (e.g. optimize execution plan)
Support for analytics (e.g. OVER clause, window functions)
TEZ framework to optimize execution
Columnar file format
http://hortonworks.com/stinger/

52. Stinger Phase 1 performance...
(Stinger phase 1 is really just Hive 0.11)

53. remember, these numbers are
non-scientific micro-benchmarks!

54. Same 9 queries (as w/ Impala), run
in HortonWorks Sandbox VM
Macbook Pro Retina, mid 2012
16GB RAM,
4GB for VM (VMWare 5),
Intel i7 2.6GHz quad-core processor
Hardware (same as w/ Impala)
No other load on system during queries
HortonWorks Data Platform (HDP) 1.3
Running pseudo-cluster

55. Query Hive (sec)
# M/R
jobs
Stinger
Phase 1 (sec)
# M/R
jobs
x Hive
perf.
A 13.8 1 10.0 1 1.4
B 30.0 1 15.8 1 1.9
C 33.3 1 14.1 1 2.4
D 23.2 1 18.7 1 1.2
E 21.6 1 19.7 1 1.1
F 59.1 2 34.3 1 1.7
G 78.5 3 35.2 1 2.2
H 59.6 2 31.5 1 1.9
TPC-DS 204.5 6 37.2 1 5.5
(remember, unscientific...)

56. Query
Stinger Phase 1
(sec)
Impala (sec) x Stinger perf.
A 10.0 0.25 39
B 15.8 0.41 38
C 14.1 0.42 33
D 18.7 0.64 29
E 19.7 0.62 32
F 34.3 1.96 18
G 35.2 1.56 23
H 31.5 1.89 17
TPC-DS 37.2 3.23 12
(remember, unscientific...)

57. Impala Review
In-memory, distributed
SQL query engine
Integrates into
existing HDFS
Not Map/Reduce
Focus on
performance
(native code)
Competition...
Interactive data
analysis

58. References
Google Dremel - http://research.google.com/pubs/pub36632.html
Apache Drill - http://incubator.apache.org/drill/
TPC-DS dataset - http://www.tpc.org/tpcds/
Stinger Initiative - http://hortonworks.com/blog/100x-faster-hive/
http://hortonworks.com/stinger/
Cloudera Impala resources
http://www.cloudera.com/content/support/en/documentation/cloudera-impala/cloudera-
impala-documentation-v1-latest.html
Cloudera Impala: Real-Time Queries in Apache Hadoop, For Real
http://blog.cloudera.com/blog/2012/10/cloudera-impala-real-time-queries-in-apache-
hadoop-for-real/

59. Photo Attributions
Impala - http://www.flickr.com/photos/gerardstolk/5897570970/
Measuring tape - http://www.morguefile.com/archive/display/24850
Bridge frame - http://www.morguefile.com/archive/display/9699
Balance - http://www.morguefile.com/archive/display/93433
* All others are iStockPhoto (I paid for them...)

60. My Info
twitter.com/sleberknight www.sleberknight.com/blog
scott dot leberknight at gmail dot com