gt 4420/6422 // spring 2019 // @joy arulraj lecture #4 ...jarulraj/courses/4420-s... · run-length...

DATABASE SYSTEM IMPLEMENTATION

GT 4420/6422 // SPRING 2019 // @JOY_ARULRAJ

LECTURE #4: SYSTEM CATALOGS & DATABASE COMPRESSION

OFFICE HOURS

Prashanth Dintyala→ Office Hours: Mon, 1:30-2:30 PM→ Location: Near KACB 3324→ Email: [email protected]

Sonia Matthew→ Office Hours: Wed, 11-12 AM→ Location: Near KACB 3324→ Email: [email protected]

2

mailto:[email protected]

http://[email protected]

HOMEWORK #1

3

We have posted clarifications on Piazza→ More clarifications will be provided over time

Separately submit the design document on Gradescope→ Homework 1 - Design Doc

TODAY’S AGENDA

System Catalogs

Compression BackgroundNaïve CompressionOLAP Columnar CompressionOLTP Index Compression

4

SYSTEM CATALOGS

Group of tables and views that describe the structure of the database.→ Statistics related to queries

Each system catalog table contains information about specific elements in the database.→ Statistics related to queries→ Statistics related to data distribution→ List of indexes→ List of user tables

5

SYSTEM CATALOGS

Almost every DBMS stores their a database's catalog in itself (i.e., using it storage manager).→ Wrap object abstraction around tuples to avoid invoking

SQL queries to retrieve catalog data→ Specialized code for "bootstrapping" catalog tables.

The entire DBMS should be aware of transactions in order to automatically provide ACID guarantees for DDL (i.e., schema change) commands and concurrent txns.

6

SYSTEM CATALOGS

MySQL→ Special hard-coded scripts to alter the catalog→ Non-transactional catalog changes

PostgreSQL→ Uses SQL commands to alter the catalog

7

SCHEMA CHANGES

ADD COLUMN:→ NSM: Copy tuples into new region in memory.→ DSM: Just create the new column segment DROP COLUMN:→ NSM #1: Copy tuples into new region of memory.→ NSM #2: Mark column as "deprecated", clean up later.→ DSM: Just drop the column and free memory.CHANGE COLUMN:→ Check whether the conversion is allowed to happen.

Depends on default values.

8

SCHEMA CHANGES

Relevant comment on HackerNews:→ Schema changes fail to complete on databases > 2 TB→ Operation requires double the amount of disk storage for

copying. Takes close to a month to perform such an operation on large tables

→ The big reason that DDL is slow is because these systems haven't tried to make it fast. So, blame the DB designers!

→ Wish it was on my day job list of things I can work on.

9

https://news.ycombinator.com/item?id=18917847

INDEXES

CREATE INDEX:→ Scan the entire table and populate the index.→ Must not block all other txns during index construction.→ Have to record changes made by txns that modified the

table while another txn was building the index.→ When the scan completes, lock the table and resolve

changes that were missed after the scan started.DROP INDEX:→ Just drop the index logically from the catalog.→ It only becomes "invisible" when the txn that dropped it

commits. All existing txns will still have to update it.

10

OBSERVATION

I/O is the main bottleneck if the DBMS has to fetch data from disk.→ CPU cost for decompressing data < I/O cost for fetching

un-compressed data. Compression always helps.

In-memory DBMSs are more complicated→ Compressing the database reduces DRAM requirements

and processing.Key trade-off is speed vs. compression ratio→ In-memory DBMSs (always?) choose speed.

11

REAL-WORLD DATA CHARACTERISTICS

Data sets tend to have highly skeweddistributions for attribute values.→ Example: Zipfian distribution of the Brown Corpus→ Words like “the”, “a” occur very frequently in books

Data sets tend to have high correlation between attributes of the same tuple.→ Example: Order Date to Ship Date (few days)→ (June 5, +5) instead of (June 5, June 10)

12

https://en.wikipedia.org/wiki/Brown_Corpus

DATABASE COMPRESSION

Goal #1: Must produce fixed-length values. Allows us to be efficient while accessing tuples.

Goal #2: Allow the DBMS to postpone decompression as long as possible during query execution. Operate directly on compressed data.

Goal #3: Must be a lossless scheme. No data should be lost during this transformation.

13

LOSSLESS VS. LOSSY COMPRESSION

When a DBMS uses compression, it is always lossless because people don’t like losing data.

Any kind of lossy compression is has to be performed at the application level.→ Example: Sensor data. Readings are taken every second,

but we may only store average across every minute.

Some new DBMSs support approximate queries→ Example: BlinkDB, SnappyData, XDB, Oracle (2017)

14

http://blinkdb.org/

http://www.snappydata.io/

https://initialdlab.github.io/XDB/

https://oracle-base.com/articles/12c/approximate-query-processing-12cr2

ZONE MAPS

Pre-computed aggregates for blocks of data.DBMS can check the zone map first to decide whether it wants to access the block.

15

Original Data

val100200300400400

ZONE MAPS


16

Zone Map

val1004002801400

typeMINMAXAVGSUM

5COUNT

Original Data

val100200300400400

ZONE MAPS


17

Zone Map

val1004002801400

typeMINMAXAVGSUM

5COUNT

Original Data

val100200300400400

SELECT * FROM tableWHERE val > 600

COMPRESSION GRANULARITY

Choice #1: Block-level→ Compress a block of tuples for the same table.Choice #2: Tuple-level→ Compress the contents of the entire tuple (NSM-only).Choice #3: Attribute-level→ Compress a single attribute value within one tuple.→ Can target multiple attributes for the same tuple.Choice #4: Column-level→ Compress multiple values for one or more attributes

stored for multiple tuples (DSM-only).

18

NAÏVE COMPRESSION

Compress data using a general purpose algorithm. Scope of compression is only based on the data provided as input.→ LZO (1996), LZ4 (2011), Snappy (2011), Zstd (2015)

Considerations→ Computational overhead (gzip is super slow)→ Compress vs. decompress speed.

19

https://en.wikipedia.org/wiki/Lempel%E2%80%93Ziv%E2%80%93Oberhumer

https://en.wikipedia.org/wiki/LZ4_(compression_algorithm)

https://en.wikipedia.org/wiki/Snappy_(software)

https://en.wikipedia.org/wiki/Zstandard

NAÏVE COMPRESSION

Choice #1: Entropy Encoding→ More common sequences use less bits to encode, less

common sequences use more bits to encode.

Choice #2: Dictionary Encoding→ Build a lookup table that maps logical identifiers to data

chunks. Example: 1 ~ “the”→ Replace those values in the original data with logical

identifiers which can be later uncompressed using the lookup table. Example: <1, 5, 3> ~ <“the lookup table”>

20

MYSQL INNODB COMPRESSION

21

Source: MySQL 5.7 Documentation

Buffer Pool Disk Pages

Compressed page0modification log



https://dev.mysql.com/doc/refman/5.7/en/innodb-compression-internals.html


22

[1,2,4,8] KB








23

[1,2,4,8] KB









24

[1,2,4,8] KB







Updates



25

[1,2,4,8] KB



Uncompressedpage0





Updates



26

16 KB

[1,2,4,8] KB



Uncompressedpage0





Updates


NAÏVE COMPRESSION

The data has to be decompressed first before it can be read and (potentially) modified.→ This limits the “scope” of the compression scheme.

These schemes also do not consider the high-level meaning or semantics of the data. → Example: Relationship between order and shipping dates

27

OBSERVATION

We can perform exact-match comparisons and natural joins on compressed data if predicates and data are compressed the same way.→ Range predicates are more tricky…

28

OBSERVATION


29

SELECT * FROM usersWHERE name = 'Andy'

NAME SALARYAndy 99999

Prashanth 88888

OBSERVATION


30



Prashanth 88888

NAME SALARYXX AAYY BB

OBSERVATION


31


SELECT * FROM usersWHERE name = XX


Prashanth 88888

NAME SALARYXX AAYY BB

COLUMNAR COMPRESSION

Compression Schemes→ Run-length Encoding→ Bitmap Encoding→ Delta Encoding→ Incremental Encoding→ Mostly Encoding→ Dictionary Encoding

32

RUN-LENGTH ENCODING

Compress runs of the same value in a single column into triplets:→ Example: [‘Atlanta’, ‘Atlanta’, ‘Atlanta’] ~ [‘Atlanta’, 3]→ The value of the attribute.→ The # of elements in the run.→ The start position in the column segment.

Requires the columns to be sorted intelligently to maximize compression opportunities.

33

DATABASE COMPRESSIONSIGMOD RECORD 1993

RUN-LENGTH ENCODING

34

Original Data

id

21

43

76

98

sex

MM

FM

FM

MM

RUN-LENGTH ENCODING

35

Original Data

id

21

43

76

98

sex

MM

FM

FM

MM

RUN-LENGTH ENCODING

36

Compressed Data

id

21

43

76

98

sex

(F,3,1)(M,0,3)

(F,5,1)(M,4,1)

(M,6,2)

Original Data

id

21

43

76

98

sex

MM

FM

FM

MM

RLE Triplet- Value- Offset- Length

RUN-LENGTH ENCODING

37

Compressed Data

id

21

43

76

98

sex

(F,3,1)(M,0,3)

(F,5,1)(M,4,1)

(M,6,2)

Original Data

id

21

43

76

98

sex

MM

FM

FM

MM


RUN-LENGTH ENCODING

38

Compressed DataSorted Data

id

21

63

98

74

sex

MM

MM

MM

FF


RUN-LENGTH ENCODING

39

Compressed DataSorted Data

id

21

63

98

74

sex

MM

MM

MM

FF

id

21

63

97

74

sex

(F,7,2)(M,0,6)


BITMAP ENCODING

Store a separate Bitmap for each unique value for a particular attribute where an offset in the vector corresponds to a tuple.→ The ith position in the Bitmap corresponds to the ith tuple

in the table.→ Typically segmented into chunks to avoid allocating large

blocks of contiguous memory.

Only practical if the value cardinality is low.

40

MODEL 204 ARCHITECTURE AND PERFORMANCEHigh Performance Transaction Systems 1987

BITMAP ENCODING

41

Original Data

id

21

43

76

98

sex

MM

FM

FM

MM

BITMAP ENCODING

42

Original Data

id

21

43

76

98

sex

MM

FM

FM

MM

BITMAP ENCODING

43

Compressed DataOriginal Data

id

21

43

76

98

sex

MM

FM

FM

MM

id

21

43

76

98

M

11

01

01

11

F

00

10

10

00

sex

BITMAP ENCODING

44

Compressed DataOriginal Data

id

21

43

76

98

sex

MM

FM

FM

MM

id

21

43

76

98

M

11

01

01

11

F

00

10

10

00

sex

BITMAP ENCODING: EXAMPLE

45

CREATE TABLE customer_dim (id INT PRIMARY KEY,name VARCHAR(32),email VARCHAR(64),address VARCHAR(64),zip_code INT

);


46


);


Assume we have 10 million tuples.43,000 zip codes in the US.→ 10000000 × 32-bits = 40 MB→ 10000000 × 43000 = 53.75 GB

Every time a txn inserts a new tuple, we have to extend 43,000 different bitmaps.

47


);




48


);




49


);




50


);

DELTA ENCODING

Recording the difference between values that follow each other in the same column.→ The base value can be stored in-line or in a separate look-

up table. → Can be combined with RLE to get even better

compression ratios.

51

Original Datatime

12:0112:00

12:0312:02

12:04

temp

99.499.5

99.699.5

99.4

DELTA ENCODING



compression ratios.

52

Original Datatime

12:0112:00

12:0312:02

12:04

temp

99.499.5

99.699.5

99.4

DELTA ENCODING



compression ratios.

53

Original Datatime

12:0112:00

12:0312:02

12:04

temp

99.499.5

99.699.5

99.4

Compressed Datatime

+112:00

+1+1

+1

temp

-0.199.5

+0.1+0.1

-0.2

DELTA ENCODING



compression ratios.

54

Original Datatime

12:0112:00

12:0312:02

12:04

temp

99.499.5

99.699.5

99.4

Compressed Datatime

+112:00

+1+1

+1

temp

-0.199.5

+0.1+0.1

-0.2

DELTA ENCODING



compression ratios.

55

Original Datatime

12:0112:00

12:0312:02

12:04

temp

99.499.5

99.699.5

99.4

Compressed Datatime

(+1,4)12:00

temp

-0.199.5

+0.1+0.1

-0.2

Compressed Datatime

+112:00

+1+1

+1

temp

-0.199.5

+0.1+0.1

-0.2

INCREMENTAL ENCODING

Type of delta encoding whereby common prefixes or suffixes and their lengths are recorded so that they need not be duplicated.This works best with sorted data.

56

Original Data

robrobbedrobbingrobot



57

Original Data


Common Prefix



58

Original Data


Common Prefix



59

Original Data


Common Prefix

-



60

Original Data


Common Prefix

-



61

Original Data


Common Prefix

-



62

Original Data


Common Prefix

-rob



63

Original Data


Common Prefix

-robrobbrob



64

Original Data


Common Prefix

-robrobbrob

Compressed Data

robbedingot

0343



65

Original Data


Common Prefix

-robrobbrob

Compressed Data

robbedingot

0343

Prefix Length

Suffix

MOSTLY ENCODING

When the values for an attribute are “mostly” less than the largest size, you can store them as a smaller data type.→ The remaining values that cannot be compressed are

stored in their raw form.

66

Source: Redshift Documentation

Original Dataint64

42

699999999

8

http://docs.aws.amazon.com/redshift/latest/dg/c_MostlyN_encoding.html

MOSTLY ENCODING

When the values for an attribute are “mostly” less than the largest size, you can store them as a smaller data type.→ The remaining values that cannot be compressed are

stored in their raw form.

67

Source: Redshift Documentation

Original Dataint64

42

699999999

8

Compressed Datamostly8

42

6XXX

8

offset3

value99999999

http://docs.aws.amazon.com/redshift/latest/dg/c_MostlyN_encoding.html

DICTIONARY COMPRESSION

Most pervasive compression scheme in DBMSs.Replace frequent patterns with smaller codes.

Need to support fast encoding and decoding.Need to also support range queries.→ Example: SALARY > 100

68

DICTIONARY-BASED ORDER-PRESERVING STRING COMPRESSION FOR MAIN MEMORY COLUMN STORESSIGMOD 2009

DICTIONARY COMPRESSION

When to construct the dictionary?What is the scope of the dictionary?How do we allow for range queries?How do we enable fast encoding/decoding?

69

DICTIONARY CONSTRUCTION

Choice #1: All At Once→ Compute the dictionary for all the tuples at a given point

of time.→ New tuples must use a separate dictionary or the all

tuples must be recomputed.

Choice #2: Incremental→ Merge new tuples in with an existing dictionary.→ Likely requires re-encoding of existing tuples.

70

DICTIONARY SCOPE

Choice #1: Block-level→ Only include a subset of tuples within a single table.→ Potentially lower compression ratio, but can add new

tuples more easily.→ Impact of dictionary data corruption is localizedChoice #2: Table-level→ Construct a dictionary for the entire table.→ Better compression ratio, but expensive to update.Choice #3: Multi-Table→ Can be either subset or entire tables.→ Sometimes helps with joins and set operations.

71

MULTI-ATTRIBUTE ENCODING

Instead of storing a single value per dictionary entry, store entries that span attributes.→ I’m not sure any DBMS actually implements this.

72

Original Data

202

val2

101202

101202

101

val1

BA

CA

BA

101C101B

MULTI-ATTRIBUTE ENCODING

Instead of storing a single value per dictionary entry, store entries that span attributes.→ I’m not sure any DBMS actually implements this.

73

Original Data Compressed Data

202

val2

101202

101202

101

val1

BA

CA

BA

101C101B

val2

101202

101

val1

BA

C

code

YYXX

ZZ

val1+val2

YYXX

ZZXX

YY

ZZXX

YY

ENCODING / DECODING

A dictionary needs to support two operations:→ Encode: For a given uncompressed value, convert it into

its compressed form.→ Decode: For a given compressed value, convert it back

into its original form.

No magic hash function will do this for us.We need two hash tables to support operations in both directions.

74

ORDER-PRESERVING COMPRESSION

The encoded values need to support sorting in the same order as original values.

75

Original Data

nameAndrea

PrashanthAndyDana



76

Original Data

nameAndrea

PrashanthAndyDana

Compressed Data

code10203040

valueAndreaAndyDana

Prashanth

name10402030



77

SELECT * FROM usersWHERE name LIKE 'And%'

Original Data

nameAndrea

PrashanthAndyDana

Compressed Data

code10203040

valueAndreaAndyDana

Prashanth

name10402030



78

SELECT * FROM usersWHERE name LIKE 'And%'

Original Data

nameAndrea

PrashanthAndyDana

Compressed Data

code10203040

valueAndreaAndyDana

Prashanth

name10402030

SELECT * FROM usersWHERE name BETWEEN 10 AND 20


79

Original Data

nameAndrea

PrashanthAndyDana

Compressed Data

code10203040

valueAndreaAndyDana

Prashanth

name10402030


80

SELECT name FROM usersWHERE name LIKE 'And%' ???

Original Data

nameAndrea

PrashanthAndyDana

Compressed Data

code10203040

valueAndreaAndyDana

Prashanth

name10402030


81

SELECT name FROM usersWHERE name LIKE 'And%' Have to perform seq scan

Original Data

nameAndrea

PrashanthAndyDana

Compressed Data

code10203040

valueAndreaAndyDana

Prashanth

name10402030


82

SELECT name FROM usersWHERE name LIKE 'And%'

SELECT DISTINCT nameFROM usersWHERE name LIKE 'And%'

Have to perform seq scan

???

Original Data

nameAndrea

PrashanthAndyDana

Compressed Data

code10203040

valueAndreaAndyDana

Prashanth

name10402030


83

SELECT name FROM usersWHERE name LIKE 'And%'

SELECT DISTINCT nameFROM usersWHERE name LIKE 'And%'

Have to perform seq scan

Only need to access dictionary

Original Data

nameAndrea

PrashanthAndyDana

Compressed Data

code10203040

valueAndreaAndyDana

Prashanth

name10402030

DICTIONARY IMPLEMENTATIONS

Hash Table:→ Fast and compact.→ Unable to support range and prefix queries.

B+Tree:→ Slower than a hash table and takes more memory.→ Can support range and prefix queries.

84

PARTING THOUGHTS

Dictionary encoding is probably the most useful compression scheme because it does not require pre-sorting.

The DBMS can combine different approaches for even better compression.

It is important to wait as long as possible during query execution to decompress data.

85

NEXT CLASS

Physical vs. Logical Logging

86

gt 4420/6422 // spring 2019 // @joy arulraj lecture #4 ...jarulraj/courses/4420-s... · run-length...

Documents