Transaction Time Indexing with Version Compression

1
Transaction Time Indexing with Version Compression

• D. Lomet, M. Hong, R. Nehme and R. Zhang
• The 34th International Conference on Very Large
  Data Bases

2
Overview

• A DBMS called Immortal DB with several
  improvements over existing systems.
• The paper focuses on the internal manipulation of
  data, rather than the more abstract issues like
  query-language design.

3
Outline

• Dealing with compatibility.
• Dealing with full pages.
• Comparing splits.
• Compressing the data.
• Storage utilization.
• Experimental results.

4
A good starting point

• To build a temporal database, a minimal
  requirement is the compatibility with traditional
  databases.
• The authors come up with a really clever idea for
  compatibility issues version chaining!

5
Version chaining

• The traditional DB is obtained simply by
  disabling all pointers!

Oldest version
Tom, 25K, Clothing (2000)
Previous version
Tom, 40K, Clothing (2005)
Most recent version
Dynamic slot array
6
A closer look

• In implementation, versions chains have to be
  stored in memory pages and disk sectors.

John, 30K, Toys (1999)
Tom, 25K, Clothing (2000)
John, 35K, Shoes (2004)
Tom, 40K, Clothing (2005)
USED
John, 40K, Shoes (2007)
FREE
Dynamic slot array
7
Adding information

• The tuples and the slots grow in the opposite
  directions, much like heap and stack.

John, 30K, Toys (1999)
Tom, 25K, Clothing (2000)
John, 35K, Shoes (2004)
Tom, 40K, Clothing (2005)
John, 40K, Shoes (2007)
Tom, 45K, Clothing (2009)
Tom, 45K, Hardware (2009)
Dynamic slot array
8
Deletion

• The clever delete stub trick.

John, 30K, Toys (1999)
Tom, 25K, Clothing (2000)
John, 35K, Shoes (2004)
Tom, 40K, Clothing (2005)
John, 40K, Shoes (2007)
I am a dummy node
Dynamic slot array
9
The explosion

• A page finally becomes full or near full.

John, 30K, Toys (1999)
Tom, 25K, Clothing (2000)
John, 35K, Shoes (2004)
Tom, 40K, Clothing (2005)
John, 40K, Shoes (2007)
Tom, 40K, Clothing (2009)
Tom, 40K, Hardware (2009)
xxx
xxx
xxx
xxx
Dynamic slot array
10
Dealing with full pages

• In a traditional database, a full page is split
  according to the key and then accessed via a
  B-tree.
• Say a page is split into one with keys lt50 and
  the other with keys gt50.
• This is called a key-split in this paper.

11
Key-splits work, but

• In a temporal database, the current data are
  accessed most often.
• Itd be nice to time-split. Then put the
  current page in fast memory and the history page
  in the secondary or tertiary storage!
• So most accesses hit in the fast memory!
• Much like caches in computer architecture!

12
An implementation issue

• Immortal DB should not be built on top of an
  existing DB!
• Otherwise, full pages are key-split anyway.

13
Performing a time-split

• Say the clock is at t now.

John, 30K, Toys (1999)
John, 30K, Toys (1999)
John, 35K, Shoes (2004)
John, 35K, Shoes (2004)
John, 35K, Shoes (2004t-1)
John, 40K, Shoes (NOW)
Dynamic slot array
Dynamic slot array
14
The problem of redundancy

• A query of John back in 2007 should logically go
  to the history page!

John, 30K, Toys (1999)
John, 30K, Toys (1999)
John, 35K, Shoes (2004)
John, 35K, Shoes (2004)
John, 40K, Shoes (2007)
John, 40K, Shoes (2007)
John, 40K, Shoes (2007t-1)
Dynamic slot array
Dynamic slot array
15
A tradeoff

• Neither splitting method is uniformly superior.
• So how often should we time-split? How about
  key-splits?

16
Index pages

• We have so far been looking at data pages.
• Recall that in the OS course, the main memory is
  paged, with some pages containing page-tables.
• The situation is much alike in Immortal DB the
  index pages serve as those containing
  page-tables!

17
Pages must be rectangles

• Because we do time- and key-splits!

Each tuple has a key and a timestamp.
Each page contains several tuples.
A key-split separates a page by adding a
horizontal line.
A time-split separates a page by adding a
vertical line (ignore redundancy for now).
Key
Time
18
An index page
Key
C
F
B
A
D
E
Time
D
B
A
E
F
C
Dynamic slot array
19
Change-log compression

• Adjacent versions of data are usually similar.
• So its nice to record only their differences.

All except the most recent data are compressed.
Dynamic slot array
20
Compression for performance

• The better compression ratio, the less we worry
  about redundancy caused by time-splits.
• Hence time-splits become free lunch.
• By distributing the larger fraction of splits for
  time-splits rather than key-splits, the benefit
  of time-splits is amplified!

21
Storage Utilization

• The paper considers two measurements.
• Single Version Current Utilization(SVCU) aside
  from history pages, what fraction of a page is
  devoted to current data?
• MultiVersion Total Utilization (MVTU) counting
  duplicated data only once, what fraction of a
  page is used?

22
The higher, the better SVCU

• A typical query goes only to current data.
• Higher SVCU means that a single page contains
  richer current data (except for history pages).
• So a higher SVCU increases the chance for a
  typical query to be resolved within one page!

23
The higher, the better MVTU

• A higher MVTU means less waste of space!

24
Effects of splits

• Performing time-splits too often decreases MVTU
  because of redundant data.
• Performing key-splits too often decreases SVCU
  because current data get split into two pages.
• Again, we see a tradeoff between time- and
  key-splits!

25
Effect of insertion

• If most accesses are insertions (rather than
  updates), then a page will be filled with current
  data, so SVCU increases!

Jimmy, 30K, hardware (2009)
Tony, 40K, Accounting (2009)
Claire, 50K, Toy (2009)
Joe, 35K, Toy (2009)
Mary, 42K, advertise (2009)
Dynamic slot array
26
Effect of update (1/2)

• If most accesses are updates, then most data will
  be compressed, so MVTU increases.

Dynamic slot array
27
Effect of update (2/2)

• But updates do not help increase SVCU. Instead,
  updates may cause key-splits, which decrease SVCU!

Dynamic slot array
28
Effect of compression

• Better compression certainly increases MVTU.
• It also increases SVCU!
• It decreases the need for splits, in particular
  key-splits.
• So current data are less likely to become
  scattered.

29
Confirming with experiments (1/2)
30
Confirming with experiments (2/2)
31
Thank you!