The SQL OFFSET keyword is evil. It basically behaves like SLEEP in other programming langauges: the bigger the number, the slower the execution. Fetching results in a page-by-page fashion in SQL doesn't require OFFSET at all but an even simpler SQL clause. Besides being faster, you don't have to cope with drifting results if new data is inserted between two page fetches.

1. © 2013 by Markus Winand
iStockPhoto/Mitshu
Pagination done
the Right Way
@MarkusWinand
@SQLPerfTips

2. © 2013 by Markus Winand
Takeaways
‣OFFSET kills performance
‣Paging can be done without OFFSET
‣It’s faster and has fewer side effects

3. © 2013 by Markus Winand
Note
In this presentation
index means B-tree index.

4. A Trivial Example
A query to fetch the 10 most recent news:
se!e"# *
$%om news
whe%e #op&" ' 1234
o!de! b" da#e des$, %d des$
&%m%# 10;
$!ea#e %ndex .. on news(#op%$);
Using o%de% b( to get the most recent first and
!&m&# to fetch only the first 10.
Alternative SQL-2008 syntax (since PostgreSQL 8.4)
$e#"h $&%s# 10 %ows on!(

5. Worst Case: No Index for o%de% b(
L&m&# (a$#ua& !ows'10)
-> So%# (a$#ua& !ows'10)
So%# Me#hod: #op-N heapso!# Memo!": 18(B
-> B&#map Heap S"an (!ows'10000)
Re"he") *ond: (#op&" ' 1234)
-> B&#map Index S"an (!ows'10000)
Index *ond: (#op&" ' 1234)

6. Worst Case: No Index for order by
The limiting factor is the number of rows that match
the whe%e clause
(“Base-Set Size”).
The database might use
an index to satisfy the
whe%e clause, but must
still fetch all matching
rows to “sort” them.

7. Another Benchmark: Fetch Next Page
Fetching the next page is easy using the o$$se#
keyword:
se!e"# *
$%om news
whe%e #op&" ' 1234
o%de% b( da#e des", &d des"
o))se# 10
!&m&# 10;

8. Worst Case: No Index for order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'20)
So%# Me#hod: #op-N heapso!# Memo!": 19(B
-> B&#map Heap S"an (a"#ua! %ows'10000)
Re"he") *ond: (#op&" ' 1234)
-> B&#map Index S"an (a"#ua! %ows'10000)
Index *ond: (#op&" ' 1234)

9. Worst Case: No Index for order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'30)
So%# Me#hod: #op-N heapso!# Memo!": 20(B
-> B&#map Heap S"an (a"#ua! %ows'10000)
Re"he") *ond: (#op&" ' 1234)
-> B&#map Index S"an (a"#ua! %ows'10000)
Index *ond: (#op&" ' 1234)

10. Worst Case: No Index for order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'40)
So%# Me#hod: #op-N heapso!# Memo!": 22(B
-> B&#map Heap S"an (a"#ua! %ows'10000)
Re"he") *ond: (#op&" ' 1234)
-> B&#map Index S"an (a"#ua! %ows'10000)
Index *ond: (#op&" ' 1234)

11. Worst Case: No Index for order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'10000)
So%# Me#hod: ex#e!na& me!*e D%s(: 1200(B
-> B&#map Heap S"an (a"#ua! %ows'10000)
Re"he") *ond: (#op&" ' 1234)
-> B&#map Index S"an (a"#ua! %ows'10000)
Index *ond: (#op&" ' 1234)

12. Worst Case: No Index for order by
Sorting might become the limiting factor when
browsing farther back.
Fetching the last page
can take considerably
longer than fetching
the first page.

13. Improvement 1: Indexed order by
se!e"# *
$%om news
whe%e #op&" ' 1234
o%de% b( da#e des", &d des"
o$$se# 10
!&m&# 10;
$!ea#e %ndex .. on news (#op%$, da#e, %d);
A single index to support the whe%e and o%de% b(
clauses.

14. Improvement 1: Indexed order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'10)
Index *ond: (#op&" ' 0)

15. Improvement 1: Indexed order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'20)
Index *ond: (#op&" ' 0)

16. Improvement 1: Indexed order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'30)
Index *ond: (#op&" ' 0)

17. Improvement 1: Indexed order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'40)
Index *ond: (#op&" ' 0)

18. Improvement 1: Indexed order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'10000)
So%# Me#hod: ex#e!na& me!*e D%s(: 1200(B
-> B&#map Heap S"an (a"#ua! %ows'10000)
Re"he") *ond: (#op&" ' 1234)
-> B&#map Index S"an (a"#ua! %ows'10000)
Index *ond: (#op&" ' 1234)

19. Improvement 1: Indexed order by
Fetching the first page is not affected by the
Base-Set size!
Fetching the next page
is also faster.
However, PostgreSQL
might take a Bitmap
Index Scan when
browsing to the end.

20. © 2013 by Markus Winand
We can do better!

21. © 2013 by Markus Winand
Don’t touch what you don’t need

22. Improvement 2: The Seek Method
Instead of o$$se#, use a whe%e filter to remove the
rows from previous pages.
se!e"# *
$%om news
whe%e #op&" ' 1234
and (da#e, %d) < (p!e+_da#e, p!e+_%d)
o%de% b( da#e des", &d des"
!&m&# 10;
Only select the rows “before” (=earlier date) the
last!row from the previous page.
A definite sort order is really required!

23. © 2013 by Markus Winand
Side Note: Row Values
Besides scalar values, SQL also defines
“row values” or “composite values.”
‣ In the SQL standard since ages (SQL-92)
‣ All comparison operators are well defined
‣ E.g.: (x, y) > (a, b) is true iff
(x > a or (x=a and y>b))
‣ In other words, when (x,y) sorts after (a,b)
‣ Great PostgreSQL support since 8.0!

24. © 2013 by Markus Winand
Row Values: How it works

25. © 2013 by Markus Winand
Row Values: How it works

26. © 2013 by Markus Winand
Row Values: How it works

27. © 2013 by Markus Winand
Row Values: Emulating them
‣ There is only a single database that has sufficient
Row-Values support for this trick (PostgreSQL).
‣ With all other database, you must emulate it
according to this scheme:
!"#$#%&'()*+',)%-.%/
%%%%012%3
%%%%%%%%%%%3&'()*+',)%-%/4
%%%%%%%%%5$
%%%%%%%%%%%3&'()*+',)%.%/%012%&'()*6+%-%/4
%%%%%%%%4
See: http://use-the-index-luke.com/sql/partial-results/fetch-next-page#sb-equivalent-logic

28. Seek-Method without Optimal Index
se!e"# *
$%om news
whe%e #op&" ' 1234
and (da#e, &d) < (p%e+_da#e, p%e+_&d)
o%de% b( da#e des", &d des"
!&m&# 10;
$!ea#e %ndex .. on news (#op%$);

29. Seek Method w/o Index for order by
L&m&# (a$#ua& !ows'10)
-> So%# (a$#ua& !ows'10)
So%# Me#hod: #op-N heapso!# Memo!": 18(B
-> B&#map Heap S"an (!ows'10000)
Re"he") *ond: (#op&" ' 1234)
-> B&#map Index S"an (!ows'10000)
Index *ond: (#op&" ' 1234)

30. Seek Method w/o Index for order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'10)
So%# Me#hod: #op-N heapso!# Memo!": 18(B
-> B&#map Heap S"an (a$#ua& !ows'9990)
Rows Remo+ed b" F%&#e!: 10 (new &n 9.2)
-> B&#map Index S"an (a$#ua& !ows'10000)
Index *ond: (#op&" ' 1234)

31. Seek Method w/o Index for order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'10)
So%# Me#hod: #op-N heapso!# Memo!": 18(B
-> B&#map Heap S"an (a$#ua& !ows'9980)
Rows Remo+ed b" F%&#e!: 20 (new &n 9.2)
-> B&#map Index S"an (a$#ua& !ows'10000)
Index *ond: (#op&" ' 1234)

32. Seek Method w/o Index for order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'10)
So%# Me#hod: #op-N heapso!# Memo!": 18(B
-> B&#map Heap S"an (a$#ua& !ows'9970)
Rows Remo+ed b" F%&#e!: 30 (new &n 9.2)
-> B&#map Index S"an (a$#ua& !ows'10000)
Index *ond: (#op&" ' 1234)

33. Seek Method w/o Index for order by
L&m&# (a"#ua! %ows'10)
-> So%# (a$#ua& !ows'10)
So%# Me#hod: #op-N heapso!# Memo!": 18(B
-> B&#map Heap S"an (a$#ua& !ows'10)
Rows Remo+ed b" F%&#e!: 9990 (new &n 9.2)
-> B&#map Index S"an (a$#ua& !ows'10000)
Index *ond: (#op&" ' 1234)

34. Seek Method w/o Index for order by
Always needs to retrieve the full base set, but the
top-n sort buffer needs
to hold only 10 rows.
The response time remains
the same even when
browsing to the last page.
And the memory footprint
is very low!

35. Seek-Method with Optimal Index
se!e"# *
$%om news
whe%e #op&" ' 1234
and (da#e, &d) < (p%e+_da#e, p%e+_&d)
o%de% b( da#e des", &d des"
!&m&# 10;
$!ea#e %ndex .. on news (#op%$, da#e, %d);

36. Seek Method with index for order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'10)
Index *ond: (#op&" ' 1234)

37. Seek Method with index for order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'10)
Index *ond: ((#op&" ' 1234)
AND (ROW(d#, &d) < ROW(‘...’, 23456)))

38. Seek Method with index for order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'10)
Index *ond: ((#op&" ' 1234)
AND (ROW(d#, &d) < ROW(‘...’, 34567)))

39. Seek Method with index for order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'10)
Index *ond: ((#op&" ' 1234)
AND (ROW(d#, &d) < ROW(‘...’, 45678)))

40. Seek Method with index for order by
L&m&# (a$#ua& !ows'10)
-> Index S"an Ba")wa%d (a$#ua& !ows'10)
Index *ond: ((#op&" ' 1234)
AND (ROW(d#, &d) < ROW(‘...’, 56789)))

41. Seek Method with index for order by
Successively browsing back doesn’t slow down.
Neither the size of the
base set(*) nor the fetched
page number affects the
response time.
(*) the index tree depth still affects the response time.

42. ComparisonOffsetSeek
W/O index for o%de% b( With index for o%de% b(

43. © 2013 by Markus Winand
Too good to be true?
The Seek Method has serious limitations
‣You cannot directly navigate to arbitrary pages
‣because you need the values from the previous page
‣Bi-directional navigation is possible but tedious
‣you need to revers the o%de% b( direction and RV comparison
‣Works best with full row values support
‣Workaround is possible, but ugly and less performant
‣Framework support?
‣jOOQ 3.3 introduced native support for the Seek-Method
‣order_query for Ruby by @glebm: http://github.com/glebm/order_query

44. © 2013 by Markus Winand
A Perfect Match for Infinite Scrolling
The “Infinite Scrolling” UI doesn’t need to ...
‣navigate to arbitrary pages
‣there are no buttons
‣Browse backwards
‣previous pages are
still in the browser
‣Stable results
‣New rows don’t affect the result
No need for de-duplication in JS

45. Also a Perfect Match for PostgreSQL
o%de% b(
support matrix
row values
support matrix
Popular
Popular
Advanced
Advanced

46. © 2013 by Markus Winand
About Markus Winand
Tuning developers for
high SQL performance
Training & co (one-man show):
winand.at
Geeky blog:
use-the-index-luke.com
Author of:
SQL Performance Explained