Cloud Frontiers: A Deep Dive into Serverless Spatial Data and FME
Shark SQL and Rich Analytics at Scale
1. Shark: SQL and Rich
Analytics at Scale
Reynold Xin, Josh Rosen, Matei Zaharia, Michael Franklin, Scott
Shenker, Ion Stoica
AMPLab, UC Berkeley
June 25 @ SIGMOD 2013
2. Challenges
Data size growing
» Processing has to scale out over large
clusters
» Faults and stragglers complicate DB design
Complexity of analysis increasing
» Massive ETL (web crawling)
» Machine learning, graph processing
» Leads to long running jobs
4. What’s good about
MapReduce?
1. Scales out to thousands of nodes in a fault-
tolerant manner
2. Good for analyzing semi-structured data and
complex analytics
3. Elasticity (cloud computing)
4. Dynamic, multi-tenant resource sharing
5.
6. “parallel relational database systems are
significantly faster than those that rely on the
use of MapReduce for their query engines”
“I totally agree.”
7.
8. This Research
1. Shows MapReduce model can be extended to
support SQL efficiently
» Started from a powerful MR-like engine (Spark)
» Extended the engine in various ways
2. The artifact: Shark, a fast engine on top of MR
» Performant SQL
» Complex analytics in the same engine
» Maintains MR benefits, e.g. fault-tolerance
9. MapReduce Fundamental Properties?
Data-parallel operations
» Apply the same operations on a defined set of data
Fine-grained, deterministic tasks
» Enables fault-tolerance straggler mitigation
10.
11. Why Were Databases Faster?
Data representation
» Schema-aware, column-oriented, etc
» Co-partition co-location of data
Execution strategies
» Scheduling/task launching overhead (~20s in Hadoop)
» Cost-based optimization
» Indexing
Lack of mid-query fault tolerance
» MR’s pull model costly compared to DBMS “push”
See Pavlo 2009, Xin 2013.
12. Why Were Databases Faster?
Data representation
» Schema-aware, column-oriented, etc
» Co-partition co-location of data
Execution strategies
» Scheduling/task launching overhead (~20s in Hadoop)
» Cost-based optimization
» Indexing
Lack of mid-query fault tolerance
» MR’s pull model costly compared to DBMS “push”
See Pavlo 2009, Xin 2013.
Not fundamental to
“MapReduce”
Can be
surprisingly
cheap
13. Introducing Shark
MapReduce-based architecture
» Uses Spark as the underlying execution engine
» Scales out and tolerate worker failures
Performant
» Low-latency, interactive queries
» (Optionally) in-memory query processing
Expressive and flexible
» Supports both SQL and complex analytics
» Hive compatible (storage, UDFs, types, metadata, etc)
14. Spark Engine
Fast MapReduce-like engine
» In-memory storage for fast iterative computations
» General execution graphs
» Designed for low latency (~100ms jobs)
Compatible with Hadoop storage APIs
» Read/write to any Hadoop-supported systems, including
HDFS, Hbase, SequenceFiles, etc
Growing open source platform
» 17 companies contributing code
15. More Powerful MR Engine
General task DAG
Pipelines functions
within a stage
Cache-aware data
locality reuse
Partitioning-aware
to avoid shuffles
join
union
groupBy
map
Stage
3
Stage
1
Stage
2
A:
B:
C:
D:
E:
F:
G:
=
previously
computed
partition
16. Client
CLI
JDBC
Hive Architecture
Meta
store
Hadoop Storage (HDFS, S3, …)
Driver
SQL
Parser
Query
Optimizer
Physical Plan
Execution
MapReduce
17. Client
CLI
JDBC
Shark Architecture
Meta
store
Hadoop Storage (HDFS, S3, …)
Driver
SQL
Parser
Spark
Cache Mgr.
Physical Plan
Execution
Query
Optimizer
18. Extending Spark for SQL
Columnar memory store
Dynamic query optimization
Miscellaneous other optimizations (distributed
top-K, partition statistics pruning a.k.a. coarse-
grained indexes, co-partitioned joins, …)
19. Columnar Memory Store
Simply caching records as JVM objects is inefficient
(huge overhead in MR’s record-oriented model)
Shark employs column-oriented storage, a
partition of columns is one MapReduce “record”.
1
Column
Storage
2
3
john
mike
sally
4.1
3.5
6.4
Row
Storage
1
john
4.1
2
mike
3.5
3
sally
6.4
Benefit: compact representation, CPU efficient
compression, cache locality.
20.
21. How do we optimize:
SELECT * FROM table1 a JOIN table2 b ON a.key=b.key
WHERE my_crazy_udf(b.field1, b.field2) = true;
Hard to estimate cardinality!
22. Partial DAG Execution (PDE)
Lack of statistics for fresh data and the prevalent
use of UDFs necessitate dynamic approaches to
query optimization.
PDE allows dynamic alternation of query plans
based on statistics collected at run-time.
24. PDE Statistics
Gather customizable statistics at per-partition
granularities while materializing map output.
» partition sizes, record counts (skew detection)
» “heavy hitters”
» approximate histograms
Can alter query plan based on such statistics
» map join vs shuffle join
» symmetric vs non-symmetric hash join
» skew handling
25. Complex Analytics Integration
Unified system for SQL,
machine learning
Both share the same set
of workers and caches
def logRegress(points: RDD[Point]): Vector {
var w = Vector(D, _ = 2 * rand.nextDouble - 1)
for (i - 1 to ITERATIONS) {
val gradient = points.map { p =
val denom = 1 + exp(-p.y * (w dot p.x))
(1 / denom - 1) * p.y * p.x
}.reduce(_ + _)
w -= gradient
}
w
}
val users = sql2rdd(SELECT * FROM user u
JOIN comment c ON c.uid=u.uid)
val features = users.mapRows { row =
new Vector(extractFeature1(row.getInt(age)),
extractFeature2(row.getStr(country)),
...)}
val trainedVector = logRegress(features.cache())
33. Conclusion
Leveraging a modern MapReduce engine and
techniques from databases, Shark supports both
SQL and complex analytics efficiently, while
maintaining fault-tolerance.
Growing open source community
» Users observe similar speedups in real use cases
» http://shark.cs.berkeley.edu
» http://www.spark-project.org