Dremel interactive analysis of web scale datasets

Dremel:
Interactive Analysis of Web-Scale Datasets
Carl Adler
IDSL - Dep. IM - NTUST

Outline
• About Dremel
• Main Features
• Record-Oriented v.s. Column-Oriented
• Data Model
• Nested Columnar Storage
• Query Execution
• Experiments
• Conclusions

About Dremel
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.

Main Features
• Dremel is a large-scale system
• The complement for MapReduce-based interactive query
• The nested data model
• Build on ideas from web search and parallel DBMSs
• Column-striped storage representation

Main Features
Dremel is a large-scale system:
• Reading 1TB of compressed data in 1 sec
-> Needs tens of thousands of disks, concurrently reading.
-> Fault tolerance is critical.

Main Features
The complement for MapReduce-based interactive query:
• Unlike traditional DBs, it is capable of operating on in situ nested data.
• Not a replacement for MR.

Main Features
The nested data model:
• Data used in web are often non-relational.
• Need some flexible data model like json.

Main Features
Build on ideas from web search and parallel DBMSs:
• Serving tree:
Divide a huge and complicated query into several small queries.
• SQL-like interface:
Like Hive and Pig.

Main Features
Column-striped storage representation:
• Read less data from secondary storage and reduce CPU cost due to
compression.
• Column stores have been adopted for analyzing relational data but to
the best of our knowledge have not been extended to nested data
models.

Record-Oriented v.s. Column-Oriented
Record-Oriented Column-Oriented

Record-Oriented v.s. Column-Oriented
• We can just retrieve A.B.C without
reading A.E or A.B.D, etc.
• Challenge: How to scan arbitrary
subset of fields efficiently and process
some analysis in the same
time.

Data Model
• The data model originated in the context of distributed systems (Protocol
Buffers), is used widely at Google, and is available as an open source
implementation.
• The data model is based on strongly-typed nested records.
Its abstract syntax is given by:
𝝉 = dom | < A1 : 𝝉[∗|?], ..., An : 𝝉[∗|?] >

Data Model
𝝉 = dom | < A1 : 𝝉[∗|?], ..., An : 𝝉[∗|?] >
• 𝝉: An atomic type or a record type.
• Atomic type: Integers, floating-point numbers, strings, etc.
• Record: It consist of one or multiple fields.
• Repeated fields (*) may occur multiple times in a record.
• Optional fields (?) may be missing from the record.
• Otherwise, a field is required.

Data Model
This type of data model is language independent and platform-neutral, so a MR
program written in Java can consume records from a data source exposed via a
C++ library.

Nested Columnar Storage
• Values alone do not convey the structure of a record.
• Given two values of a repeated field, we do not know at what ‘level’ the value
repeated (e.g., whether these values are from two different records, or two
repeated values in the same record).
 Repetition Levels
• Given a missing optional field, we do not know which enclosing records were
defined explicitly.
 Definition Levels

Nested Columnar Storage: Repetition Levels
• Repetition Levels:
It tells us at what repeated field in the field’s path the value has repeated.
• The field path Name.Language.Code contains two repeated fields, Name
and Language. Hence, the repetition level of Code ranges between 0 and 2;
level 0 denotes the start of a new record.

Nested Columnar Storage: Repetition Levels

Nested Columnar Storage: Definition Levels
• Definition Levels:
Each value of a field with path p, esp. every NULL, has a definition level
specifying how many fields in p that could be undefined (because they are
optional or repeated) are actually present in the record.

Nested Columnar Storage: Definition Levels

• Splitting Records into Columns:
With this type of data model, write operation is very easy, but we need to
focus on reading. When reading, we don’t need to read the entire records,
and we can just read those columns we need to form a partial data model.

Complete record assembly automaton. Edges are labeled with repetition levels.

Query Execution
• Dremel’s query language is based on SQL and is designed to be efficiently
implementable on columnar nested storage.
• Each SQL statement takes as input one or multiple nested tables and their
schemas and produces a nested table and its output schema.

Query Execution
Sample query, its result, and output schema.

Query Execution
Architecture:
• Dremel uses a multi-level serving tree to execute queries.
• A root server receives incoming queries, reads metadata from the tables,
and routes the queries to the next level in the serving tree. The leaf servers
communicate with the storage layer or access the data on local disk.

Query Execution
System architecture and execution inside a server node.

Query Execution
• Consider a simple aggregation query below:
SELECT A, COUNT(B) FROM T GROUP BY A
• When the root server receives the above query, it determines all tablets, i.e.,
horizontal partitions of the table, that comprise T and rewrites the query as
follows:
SELECT A, SUM(c) FROM (R1
1 UNION ALL ... R1
n) GROUP BY A
• Tables R1
1 , …, R1
n are the results of queries sent to the nodes 1, …, n at level
1 of the serving tree:

Query Execution
• Tables R1
1 , …, R1
n are the results of queries sent to the nodes 1, …, n at level
1 of the serving tree:
R1
i = SELECT A, COUNT(B) AS c FROM T1
i GROUP BY A
• T1
i is a disjoint partition of tablets in T processed by server i at level 1.
• Here, we can know that the dataset will smaller than the original one, and
each dataset can be processed faster.

Query Execution
• Because Dremel is a multi-user system(usually several queries are executed
simultaneously).
• A query dispatcher schedules queries based on their priorities and balances
the load. Its other important role is to provide fault tolerance when one
server becomes much slower than others or a tablet replica becomes
unreachable.

Query Execution
• A system with 3000 leaf servers
• Each leaf server using 8 threads
• 3000 * 8 = 24000 (slots)
• A table spanning 100,000 tablets
• Assigning about 5 tablets / slot

Experiments
• The basic data access characteristics on a single machine
• How columnar storage benefits MR execution
• Dremel’s performance

Experiments
Table
name
Number of
records
Size (unrepl.,
compressed)
Number
of fields
Data
center
Repl.
factor
T1 85 billion 87 TB 270 A 3×
T2 24 billion 13 TB 530 A 3×
T3 4 billion 70 TB 1200 A 3×
T4 1+ trillion 105 TB 50 B 3×
Datasets used in the experimental study

Experiments – Single Machine
Performance breakdown when reading from a local disk
(300K-record fragment of Table T1)
T1 85 billion 87 TB 270 A 3×

Experiments – MR and Dremel
Q1: SELECT SUM( CountWords (txtField)) / COUNT(*) FROM T1
T1 85 billion 87 TB 270 A 3×

Experiments – Serving Tree Topology
Q2: SELECT country, SUM( item.amount ) FROM T2 GROUP BY country
Q3: SELECT domain, SUM( item.amount ) FROM T2 WHERE domain CONTAINS ’.net’ GROUP BY domain
T2 24 billion 13 TB 530 A 3×

Experiments – Per-tablet Histograms
The area under each histogram corresponds to 100%. As the figure
indicates, 99% of Q2 (or Q3) tablets are processed under one second
(or two seconds).

Experiments – Scalability
In each run, the total expended CPU time is nearly identical, at about
300K seconds, whereas the user-perceived time decreases near-linearly
with the growing size of the system.

Experiments – Stragglers
Q6: SELECT COUNT(DISTINCT a) FROM T5
In contrast to the other datasets, T5 is two-way replicated. Hence, the
likelihood of stragglers slowing the execution is higher since there are
fewer opportunities to reschedule the work.

Conclusions
• Dremel is a custom, scalable data management solution built from simpler
components. It complements the MR paradigm.
• We outlined the key aspects of Dremel, including its storage format, query
language, and execution.
• Multi-level execution trees & Columnar data layout
• In the future, it might be widely adopted in the world.

Reference
• Dremel: Interactive Analysis of Web-Scale Datasets

Dremel interactive analysis of web scale datasets

Recommended

More Related Content

What's hot (20)

Similar to Dremel interactive analysis of web scale datasets (20)

Recently uploaded (20)

Dremel interactive analysis of web scale datasets