BID209: Best Practices for Implementing with Sybase IQ

Mark D. Mumy Principal Systems Consultant [email protected] 24 January 2012

About This Presentation



Setup and Administration

Hardware Configuration Creating the Database IQ Server Settings Connectivity Maintenance Tasks Memory Usage

Development and Design

Data Model Recommendations Database Programming Data Manipulation

Target Audience
 

Database Administrators
All Levels

Query Developers
All Levels

For This Presentation



Lets Keep It Interactive

I will entertain questions on a subject Would be happy to speak to you offline, if desired

 

No Question is too Basic

Chances are others may have the same question

Open Question Forum at the end

Time Permitting

Setup and Administration

     

Hardware Configuration Creating the Database IQ Server Settings Connectivity Maintenance Tasks Memory Usage

Hardware Configuration Drive Arrays



Volume Manager
Not a necessary or recommended component for IQ Can use MPXIO on Sun Platform to provide controller failover, dynamic multi-pathing, and world wide naming convention Additional overhead and software that has no value add to an IQ installation Acceptable for an IQ single node operation where hardware does not have the ability to apply RAID to the disk devices

Hardware Configuration Drive Arrays



RAID Level
Recommend using RAID 5 as a blend between performance, protection, and cost

Recommend using raw devices (a must for multi-node IQ) For details on drive array specifications and configurations see the Sun Reference Architecture Whitepapers and EMC Whitepapers on array configurations with IQ

Hardware Configuration Memory



How Much RAM for IQ

As much as possible! Most systems should have at least 2 GB Rough rule of thumb is 2 GB per CPU Dont forget additional RAM for each load being performed Shared memory is only needed on 32-bit systems and some 64-bit systems to wire memory into RAM HP-UX 64-bit should not use iqsmem

Hardware Configuration Processors and Disk Controllers



Optimal CPU configuration

Queries: 1 CPU per active query more if queries are complex and can be run in parallel Loads: 1 CPU per HG index plus 1 CPU per 2-5 columns being loaded

Disk Controllers
1 fiber controller per 5 write CPUs 1 fiber controller per 10 read CPUs In mixed mode operation use the 5:1 ratio Recommend a minimum of 2 controllers

Creating The Database Creation Options



Creation Options
Block Size and Page Size Minimum should be 128K page size (8K block) Use 256K page for larger databases (>200 GB or so) Larger the page size the more RAM that will be required CASE IGNORE vs. CASE RESPECT Use CASE RESPECT whenever possible Removes case comparison steps and improves performance Java in the Database Install if Java will be needed in the database No impact if installed and not used

10

Creating The Database File Placement

 

Use relative links makes it easier to relocate files Use symbolic links for IQ MAIN and IQ TEMP
/IQM_devs/IQ_MAIN_00 /IQM_devs/IQ_TEMP_00

Place the transaction log and transaction log mirror on significantly large file system
Recommend 5 GB of filesystem space per IQ node for database related storage Transaction logs, IQ MSG, scripts, catalog file, etc

11

Creating The Database Files Locations



Filename rules
Catalog: DATABASE_NAME.db Transaction log: DATABASE_NAME.tlog IQ MSG: DATABASE_NAME.iqmsg IQ Main Store: DATABASE_NAME_iqmain_000 000 would be an incrementing number IQ Temp Store: DATABASE_NAME_iqtmp_000 000 would be an incrementing number IQ Configuration File: params.cfg (default)

12

Creating The Database Top Level Directory



Use for writer and readers (multi-node configuration) to communicate IQ 12.4.3 and earlier
Must be shared across all hosts Generally done via NFS, but some customers use cluster based filesystems (Veritas Cluster Manager)

IQ 12.5
No longer a shared filesystem! Should be local filesystem on each host that is protected accordingly The name and path can be unique for each node No need for NFS or cluster based filesystems

13

IQ Server Settings Query Performance



FORCE_NO_SCROLL_CURSORS
Should always be set to ON Very few applications require this to be OFF Can improve query performance

14

IQ Server Settings Query Plans



Query Plan Settings to Provide Optimal Query Information to DBAs and Sybase Engineering
set temporary option query_plan='on'; set temporary option query_plan_after_run='on'; set temporary option query_timing='on'; set temporary option query_detail='on'; set temporary option dml_options10='on'; set temporary option query_plan_as_html='on'; set temporary option Query_Name = Query Name

Should not be set globally as the IQ MSG file will grow rapidly

15

IQ Server Settings Storage



Append_Load
Can be used to improve load performance Will not reuse Row IDs or the space occupied by those Row IDs Great for systems where large, contiguous chunks of data are deleted May not be good if random rows are deleted as it can lead to fragmentation and allocated, but unused space

16

IQ Server Settings Storage Continued



Disk_Striping
If ON, IQ will stripe writes to all available devices If OFF, the first device must be full before the next is used

Disk_Striping_Packed
If ON, it forces better space usage and less fragmentation Fragmentation is indicated when out of space messages are returned but the main dbspace used is less than 100% full Allows Adaptive Server IQ to better utilize small pieces of unused space that remain after compression

17

IQ Server Settings Data Loads



LOAD_MEMORY_MB
Set to 0 (default) on systems with enough RAM Set to <= 500 on systems where RAM is tight or there are simultaneous loads taking place Rough rule of thumb for load memory requirements Run: select tablewidth( tablename ) Divide that value by 2 New number is rough calculation of MB needed to load that table Remember that each load requires separate memory

18

Network Connectivity TCP/IP Ports

     

When configuring a TCP/IP port for IQ to listen on, use something other than the default of 2638 ASA and IQ use port 2638 as a broadcast listener Multiple servers on the same host share the default port for broadcast listening, but not for client network traffic Broadcast listening takes precedence over client network traffic as it is the first port started Easier and safer to assume that 2638 is in use already by ASA/IQ and to use another port for client network traffic Many things must be considered if an IQ server is to use port 2638 for client network traffic on a host with more than one IQ server (server start order, multiplex synchronization, etc) See http://my.sybase.com/detail?id=1012955 for details

19

Client Access and Network Connectivity



Different internal environments and settings for ODBC, JDBC, and Open Client connections JDBC and ODBC are recommended
JConnect is Sybases implementation of JDBC Use the JDBC 2 driver (JConnect 5.2/5.5) com.sybase.jdbc2.jdbc.SybDriver

Open Client
Use with caution Most applications written using Open Client expect an ASE server

20

Client Access and Network Connectivity Open Client vs. ODBC



See Chapter 32 of the ASA Users Manual as well as the ASA Reference Section Transact-SQL and SQL/92 compatibility options for complete list of differences If writing stored procedures or embedded application code, make sure to explicitly make settings for compatibility as these options will get set to different values for Open Client vs. ODBC connections
   

ALLOW_NULLS_BY_DEFAULT QUOTED_IDENTIFIER STRING_RTRUNCATION ANSI_BLANKS

  

ANSINULL CHAINDED FLOAT_AS_DOUBLE

21

Client Access and Network Connectivity Open Client vs. ODBC

 

ODBC and JDBC connectivity support all IQ datatypes Open Client does not support all IQ datatypes
Unsigned integer Depending on version of Open Client, bigint and unsigned bigint should be OK (OC 12.5 and later)

22

Client Access and Network Connectivity



AutoPreCommit Within ODBC

Set registry setting AutoPreCommit to Y Forces applications to issue a COMMIT before each query Go to the registry and update the corresponding Sybase Data Source Name (DSN) created, by adding a new value 'AutoPreCommit with a value of Y' HKEY_LOCAL_MACHINE/SOFTWARE/ODBC/ODBC.INI/{DSN}

Packet Sizes
Larger packet sizes will help with large data retrieval Use p option in IQ configuration file to increase size Use CommBufferSize parameter in ODBC connection string

23

Client Access and Network Connectivity



Network Speed
Data retrieval will be as fast as the network 100 MB of data will take 80 seconds on a 10 Mbit LAN 100 MB of data will take 8 seconds on a 100 Mbit LAN 100 MB of data will take 0.8 seconds on a Gigabit LAN LAN speed may be the performance bottleneck for queries that return large amounts of data The faster the network cards and LAN the better off concurrency will be: more available bandwidth per user Adding multiple network cards to the IQ node will help with network

Network Interface Cards

24

Client Access and Network Connectivity



ASE Component Integration Services (CIS)

Can have ASE CIS reference IQ tables (proxy tables) Use ASE 12.5 CIS class of ASIQ Prior to ASE 12.5 ASAnywhere or ASEnterprise classes had to be used Not a viable option if joining non-IQ tables with an IQ proxy table Can map multiple ASE logins to a single IQ login Data modifications should not be performed on proxy tables Can map ASE proxy tables to stored procedure calls in IQ

25

Client Access and Network Connectivity



IQ Component Integration Services (CIS)

Can only be used on the Solaris and WinNT versions of IQ (12.4.3) All platforms support CIS with IQ 12.5 Can map to tables in Oracle (ODBC), ASE (ODBC & JDBC), ASA (ODBC & JDBC), DB2 (ODBC), MS SQL Server (ODBC), and any ODBC source Not a viable option if joining IQ tables with non-IQ proxy tables Data modifications may be performed on proxy tables

26

Maintenance Tasks Database Consistency Checks



IQ 12.4.3
The routine to check the database for potential corruptions is SP_IQCHECKDB There are no run-time options for this command, however the DBCC options control the behavior of the SP_IQCHECKDB command Recommendations Run Level 1 (DBCC_OPTION=1) DBCCs every 1 to 4 weeks (will process 1-2 GB per second) Run Level 0 (DBCC_OPTION=0) DBCCs every 1 to 3 months (will process 50-100 MB per second) Use DBCC_TEAM_SIZE option to control parallel execution

27

Maintenance Tasks Database Consistency Checks



IQ 12.5
New syntax for IQ 12.5 See the Reference Manual Recommendations Perform CHECK every 1 to 4 weeks Perform VERIFY every 1 to 3 months Perform REPAIR if errors are reported from VERIFY Use RESOURCE_PERCENT parameter to control CPU usage during DBCC.

28

Maintenance Tasks Parallel Create index



Create Multiple Indexes in a batch

Syntax: BEGIN PARALLEL IQ Create HG Index ; Create LF Index ; END PARALLEL IQ ;

Recommendation
Create no more indexes in a batch than you have CPU's Note: Two indexes on same column will be serial

29

Maintenance Tasks Setting the IQ Caches



Start by allocating memory to caches

40% IQ Main Cache 60% IQ Temporary Cache

 

Adjust allocations from here based on monitoring Number of HG indexes in a table may alter this plan for loading data Make sure not to starve other operations that need RAM (OS, LOAD_MEMORY_MB, other applications, etc)

30

Maintenance Tasks IQ Monitoring

 

IQ Monitor is a diagnostic tool for DBAs It collects and reports internal counters from the IQ Buffer Caches
Main Cache interaction with the IQ Store Temporary Cache interaction with the IQ Temp Store

IQ Monitor offers a series of views of the counters to showing differing aspects of the server and buffer cache workload

31

Maintenance Tasks IQ Monitoring



Provides different views of buffer activity

summary report of both caches detailed report of one cache i/o activity of a cache a debug report of all buffer cache activity

You must specify an option when you start the monitor to specify what view to monitor One Monitor may be running for a cache
May have one for each (Main & Temp)

32

Managing Memory How Much Can IQ Address?

   

Solaris and AIX (64-bit) limited to RAM + swap space with no kernel changes necessary AIX (32-bit) limited to 2.5 GB using iqsmem to achieve this HP-UX limited to kernel value of maxdsiz_64bit
maxdsiz_64bit should never be set to exceed RAM + swap space

Windows NT/2000/XP Must use a 4GT supported version

2 GB for main and temp buffer caches 3 GB total for the process Load memory counts against 3 GB total, not 2 GB limit on buffer caches

33

Managing Memory How Much Can IQ Address?

 

Keep in mind that more memory may mean tuning the O/S On Solaris, a system with 288 GB spent a tremendous amount of time managing memory with standard configuration
Changed following kernel parameters to help (/etc/system): set autoup=900 set tune_t_fsflushr=1 set lotsfree=0x10000 set bufhwm=8000 set maxphys=1048576 set maxusers=2048 set rlim_fd_max=2048 Keep in mind that above options were for a limited test and were the result of site specific analysis performed by Sun and should be used as reference only

Refer to hardware and O/S specific tuning guides for help

34

Managing Memory Inventory Your System Memory



Assess whats running and how much memory it uses (exclusive of IQ) Typically these include
Operating System OLAP Servers Middleware Other applications

35

Managing Memory
Deduct that total from RAM Allow 20% for File System Cache
If using file systems   

 

Deduct another 10% (just in case) The rest is for IQ and IQ Caches A picture tells this story

36

Memory The Big Picture



Start with the O/S and other applications Determine what the IQ server needs (normal load) IQ Overhead (Heap for loading) Allocate IQ Caches
40% IQ Main Cache 60% IQ Temp Cache

Operating System + All Other Apps IQ Server IQ Overhead IQ Main Cache IQ Temp cache

37

Development and Design

Data Model Recommendations Database Programming Data Manipulation

  

38

Data Model Recommendations Proper Datatype Sizing

 

Use the smallest datatypes possible for data Be aware of all datatypes in IQ there may be more than you know If hour, minute and second information is not necessary, use DATE instead of DATETIME If the data will fit within a TINYINT or SMALLINT datatype use that rather than INTEGER or BIGINT Allows the engine to store data in smaller units (1-byte TINYINT or 2-byte SMALLINT versus 4-byte INTEGER or 8-byte BIGINT Dont over allocate storage when defining NUMERIC() or DECIMAL() as it can be costly for data that doesnt need all that space

39

Data Model Recommendations IQ UNIQUE

  

Use whenever possible to help minimize storage usage IQ 12.5

Used solely for FP index type (storage)

IQ 12.4.3 and earlier

Imperative to use on all columns even if cardinality exceeds 64K Optimizer may make use of IQ UNIQUE value at query runtime Does not need to be exact, but should be close to cardinality

40

Data Model Recommendations IQ UNIQUE and the FP index



 

Consider using Minimize_Storage option in Sybase IQ 12.5. This will place an IQ UNIQUE(255) on every column for every table created and removes the need to use IQ UNIQUE If the value is <= 255 then IQ will place a 1-byte FP index on the column 1 byte of storage per row If the value is > 255 but <= 65536 then IQ will place a 2-byte FP index on the column 2 bytes of storage per row May slightly hinder data loads, but improve query speeds May incur onetime slight load slowdown while 2-byte FP is converted to flat FP, but this usually happens during the first load

41

Data Model Recommendations NULL Values



Always specify NULL or NOT NULL

Open Client and ODBC connections have different default behavior when table is created

 

Allows the optimizer a better guess at join criteria NULL data does not save space on the database page as it would in ASE Will be compressed out when stored on disk

42

Data Model Recommendations Unsigned Datatypes

   

Use unsigned datatypes where possible Use for surrogate keys and join columns Unsigned data comparisons are quicker The caveat to this is that Open Client may misinterpret the value if it is too large as it does not understand large unsigned data
Can convert to signed integer, numeric, or decimal if returning data to an Open Client application This caveat applies to moving data between IQ servers with INSERT FROM LOCATION

43

Data Model Recommendations Long Varchar and Long Varbinary

    

Can be used to store moderate amounts of text or binary data VARCHAR() or VARBINARY() datatypes
Maximum width is 32K (64K ascii hex for VARBINARY())

LONG BINARY added in 12.5

Maximum width is 64K binary (128K ascii hex)

The WORD index is the only index allowed on VARCHAR() data wider than 255 bytes Storage will be allocated in 256 byte chunks
A 257 byte string will require 512 bytes of storage Much less than the 2K requirement in ASE TEXT and IMAGE columns

44

Data Model Recommendations Large Object Storage

 

  

Can be used to store binary or text based objects Extends the long binary datatype from a maximum size of 64K to an unlimited size!!! FP index is the only viable index Cannot be searched, only stored and retrieved Special function to return the size of an object (byte_length64) Special function to return portions of the object, not the entire contents (byte_substr64)

45

Data Model Recommendations Large Object Storage Continued



Create table example:

create table lob_test ( seq_no int, my_lob long binary )

No change to load table statement, so how is data loaded in the LOB? Through a load table source data file:
1|/tmp/lob_1.txt| 2|/tmp/lob_2.txt| 3|/tmp/lob_3.txt|

Load engine will place the contents of the file specified in the source data file (/tmp/lob_1.txt) into the long binary column for that row Can also extract contents of a binary object to individual files with the BFILE() function
46

Data Model Recommendations Varchar vs. Char

  

Use CHAR() whenever possible All storage in IQ is fixed width VARCHAR() types only add overhead
A VARCHAR(100) columns will require 101 bytes of storage 100 bytes for data 1 byte for the size of data

CHAR() data is blank padded, VARCHAR() is not

47

Data Model Recommendations When and Where to Use Indexes



Always use indexes on:

Join columns (HG index regardless of cardinality) Searchable columns (HG or LF index) Aggregation columns (HNG) Only if a single column is used in an aggregation SUM( A * B ) will not use an HNG, but rather an FP Date, Time, and Datetime columns (DATE, TIME, DTTM)

 

If uncertain, place an LF or HG index on the column Use Primary Key, Unique Constraint, or UNIQUE HG indexes where appropriate

48

Data Model Recommendations When and Where to Use Indexes



A column with an HNG, CMP, or WD index should have a corresponding LF or HG index (very rare circumstances will negate this) Indexes are not needed on columns whose data is ONLY returned to the client (projected) Remove HNG indexes on date/time/datetime columns. Replace them with DATE, TIME, or DTTM indexes Investigate the use of HNG indexes. In certain circumstances having HNG indexes on columns can hurt performance (mostly in aggregations).

49

Data Model Recommendations Simple Index Selection Criteria



Ask yourself these simple questions about each column to determine which indexes belong
Is the cardinality greater than 1500-2000? Yes: place an HG index on this column No: place an LF index on this column Does this column contain date, time, datetime, or timestamp data? Yes: place a DATE, TIME, or DTTM index on this column (these should be used in lieu of the HNG index recommended in IQM 12.4.3 and earlier) Will this column be used in range searches or aggregations? Yes: place an HNG index on this column (if the aggregation contains more than just the column, an HNG may not be appropriate) Will this column be used for word searching? Yes: place a WD index on this column Will this column be compared to another column in the same table? Yes: place a CMP index on the two columns being compared

50

Data Model Recommendations Multi-column Indexes



 

Currently, only HG, UNIQUE HG, UNIQUE CONSTRAINT, and PRIMARY KEY indexes support multiple columns HG inserts are the most expensive in IQ Try to guarantee that inserts will happen at the end of the index
Place generally incrementing data at the beginning of the index list (sequential data) For instance, a transaction date or batch number Something that will try to guarantee a sequential key

51

Data Model Recommendations Join Column

 

Prefer joining integer datatypes (unsigned if possible) Integer comparisons are quicker than character comparisons Keep the datatypes as narrow as possible to improve join performance by reducing disk I/O and memory requirements Prefer using HG index on join columns rather than cardinality appropriate index (LF or HG)

52

Data Model Recommendations Primary Keys



Multi-column primary keys should have an additional LF or HG index placed on each individual column
Must be done manually

UNIQUE CONSTRAINT, UNIQUE HG, and Primary Key are identical structures: HG index with no G-Array Use when possible
Helps optimizer make more informed query path decisions even if index is not used in joins or searches You get an HG Index created automatically This HG index has no G-Array (uses less space)

53

Data Model Recommendations Foreign Keys



Should be used with IQ 12.5 and later to aid in query join performance Use when possible
Helps optimizer make more informed query path decisions even if index is not used in joins or searches You get an HG Index created automatically on the foreign key column in the current table Requires that a primary key exist on referenced table

54

Data Model Recommendations Temporary Tables



3 Types of Temporary Tables

# tables create table #temp table ( col1 int ) Local Temporary Tables declare local temporary table temp table ( col1 int ) Behave just like # tables Global Temporary Tables create global temporary table temp table ( col1 int ) Tables structure is static across connections and reboots

55

Data Model Recommendations Temporary Tables



On Commit Preserve Rows

Use this option so that rows in temporary tables remain after the transaction has been committed

Temporary tables are available at the current level (parent) and all of its children A parent cannot see a child's temporary table(s)

56

Data Model Recommendations Cursors



Avoid using cursors

Generally means row based processing IQ was designed for set based processing Sometimes they cannot be avoided If used, make sure to use NO SCROLL cursors

Open With Hold

Allows the cursor to remain open across transactions If not used, the cursor may be closed when a commit is issued (depends on connectivity type)

57

Database Programming Language Watcom SQL vs. T-SQL

 

IQ (ASA) is not 100% T-SQL Compatible, but very close Recommend using Watcom SQL
All system procedures written with it Many more code examples and more IQ people versed in it Dynamic SQL Better Loop control Full cursor movement rather than just read next Cannot mix T-SQL with Watcom SQL

Watcom SQL has some extensions that T-SQL does not:

Batches and procedures must be written in the same dialect

58

Database Programming Language Watcom SQL vs. T-SQL


      

Behavior differences include:

DECLARE CURSOR GOTO IF PRINT RAISERROR SET WHILE (T-SQL) vs. LOOP
59

    

Global variables Variable Names CALL FOR ASA requires variables to be declared immediately after a BEGIN

Database Programming Language Commit and Rollback



Use transaction control around logical units of work, even read only queries Should commit before a read/write batch is started to ensure latest version of data is available Should issue commit and rollback after batch completion to release all query resources Rollback will free memory resources in use by previous operations For systems with high number of connected users, freeing memory resources can aid in query performance
60

Database Programming Language Custom Functions

 

 

Custom functions can be written in either SQL or Java Great way to encapsulate business logic for transforming data Can have a significant performance impact on queries Functions are executed in the catalog portion of the engine
All result rows may need to be moved to ASA Can be time consuming for large result sets

Turn on query plans to see what impact the functions have on effective query plans

61

Database Programming Language Outer Joins

    

T-SQL Outer Join: *= or =* ANSI Standard/Watcom SQL Outer Join: [left|right|full] outer join Be careful of outer join syntax The T-SQL syntax can be very ambiguous or non-deterministic for IQ to translate
All T-SQL outer joins must be converted to ANSI outer joins and then processed

Use the ANSI standard instead as they are not ambiguous and are always clear in their meaning Visit http://my.sybase.com/detail?id=1017447 for more details

62

Data Manipulation Load Table



Parallel Load Table

Make sure to put column delimiter after last column Must use ROW DELIMITED BY and DELIMITED BY options in load table command Column and row delimiters must be a single character 1 to 4 byte delimiters allowed for serial loads Parallel load can only have a single character delimiter

63

Data Manipulation Load Table

  

If possible, perform load table from binary datafiles

Can be 3 to 10 times faster than ASCII loads

Can use FILLER() clause with a delimiter or byte count Better performance achieved by casting the date or datetime formats rather than letting IQ guess If possible, issue a single load table with multiple files rather than 1 load table per file to be loaded into a table
Only important for tables with HG, Unique HG, or Primary Key indexes

64

Data Manipulation Insert From Location

 

Great way to move data from any Open Client source Syntax:
insert into TABLE() location SERVERNAME.DBNAME { select statement };

  

IQ username and login must match on remote system Interfaces entry must match the SERVERNAME Can also be used to move data quickly from an ASA table to an IQ in the same server See section on Open Client datatype limitations

65

Data Manipulation Single Row Operations



Avoid at all costs for large data manipulation operations Different from single statement operations that modify many rows Individual INSERT VALUES() will be slower than bulk load operations
Expect no more than 5,000 to 20,000 operations per hour

66

Data Manipulation Single Row Operations



If multiple single row inserts are needed there is a method to improve performance Consider this code (1 insert per row):
insert into #my_table values ( 1, 1 ) insert into #my_table values ( 2, 1 ) insert into #my_table values ( 3, 1 )

Can be converted to this (always 2 inserts):

create table #iq_dummy ( a1 int ) insert into #iq_dummy values ( 1 ) insert into #my_table select 1, 1 from #iq_dummy union all select 2, 1 from #iq_dummy union all select 3, 1 from #iq_dummy
67

Data Manipulation Named Pipes and Flat Files

  

Named pipes can be faster no disk I/O How to make a named pipe on most flavors of UNIX:
mknod PIPE_NAME p

 

Can use BCP, GZIP, UNCOMPRESS, or applications to push IQ formatted data into named pipe LOAD TABLE command can read from named pipe Can also fast extract data to named pipes so that they can be read by another application or even compressed and stored

68

Data Manipulation Partial Width Inserts

   

Can induce fragmentation and overly large space consumption if not watched Row IDs from deleted data are not reused during a partial width insert operation Space from delete data is not reused (because Row IDs are not reused) Partial width inserts are analogous to APPEND_LOAD=on in terms of Row ID and space behavior Only becomes a problem if partial width inserts are a way of life for a table

69

Data Loading The Deep Fact Table



Many databases have a deep table

A growing table with tens of millions of rows Rows typically rolled out and replaced over time

 

HG indexes can be slow to load/delete

Loading time may increase as data added, but it depends

Solution (with a caveat)

Partition the table (Example time: day, week, month) Build a view that is a Union of all the tables

70

Data Loading Partitioned Fact Table

Big Fact Table Partitioned t1 t2 t3 Accessed by View Create View bigtable as Select * from t1 Union All Select * from t2 Union All Select * from t3

71

Data Loading Why Partitioning?

  

Loads are faster and predictable

x million rows will load consistently

IQ may process the Union All in parallel

As long as CPU resources are available

To roll data out, truncate one table

Truncate table is much faster than Delete No changes to DDL required Load new data into the empty (truncated) table

72

Data Extraction
For fastest data unloads use the TEMP_EXTRACT options Can unload data in ASCII or BINARY format
Recommend BINARY format as it is faster to reload into IQ   

Great way to archive portions of the database Can unload to one or more files, serially Avoid using ISQL or DBISQL redirection to a file
Much slower than fast unloads

73

Best Practices for Implementing with Adaptive Server IQ



Th

Questions?

74

BID209: Best Practices for Implementing with Sybase IQ

Mark D. Mumy Principal Systems Consultant [email protected] 24 January 2012