I'm pulling data into SAS from a SQL server database with a (SAS DI) extract transformation. The data in the table goes back several years and has a little over 16 million rows; I only need data for the last few years which should amount to roughly 2.6 million rows.
So my extract uses a datetime filter. The autogenerated proc sql looks as follows (after obfuscating the libref, table, and column names):
%let SYSLAST = %nrquote(LIBREF.SRC_TBL);
proc sql;
create table work.WFJVX4PU as
select
col1,
col2,
dt_col,
col3
from &SYSLAST
where dt_col >= &start_dt
;
quit;
The code works and returns the desired rows, but it takes far too long to run when compared with executing a similar query directly on the SQL server. When inspecting the statistics of the execution I discovered that the entire table was brought into SAS before the where clause was applied.
I remember reading that this will occur when attempting to use filters that apply functions to source data (a la datepart(dt_col) >= input("&start_date",date9.)) which is why I tried to pass the datetime used by the filter directly to the SQL engine. I also tried the "datepart" approach and got the same result.
Is there something else I should be doing to apply the filter server side before bringing the data into SAS? This isn't the experience I've had in the past when working with other database tables (e.g. Teradata, MySQL, Oracle, etc).
Additional Details:
The macro variable start_date is defined using the SAS DI prompt creation tool; the auto-generated code for this is
%let start_date_label = 730 days ago (July 02, 2017);
%let start_date_rel = D-730D;
%let start_date = 02Jul2017;
I then create the datetime macro variable start_dt by executing the following in the precode of the job:
%let start_dt = %sysfunc(dhms("&start_date"d,0,0,0));
Below is the libname statement (after obfuscation):
LIBNAME MYLIB SQLSVR CONNECTION=SHARED PRESERVE_TAB_NAMES=YES dbconinit="set ansi_warnings off;use MY_DB_NAME;set nocount on;" Datasrc=MY_DATA_SRC SCHEMA=dbo AUTHDOMAIN="my_sql_authentication" ;
If I modify the where clause to use a literal of the form '2019-06-25' then SAS throws the error
ERROR: Expression using greater than or equal (>=) has components that
are of different data types.
because the dt_col field is of type numeric (format datetime22.3) on the SAS instance of the SQL server table that is registered from the library. If I use a literal as follows:
dt_col >= '25JUN2019'D
then I do get the desired set of rows (despite comparison between a datetime field and a date literal), but the query still takes a long time to execute and the job statistics indicate SAS is still grabbing all 16 million rows to perform this task.
UPDATE
I'm having issues following Tom's advice below. If I execute the following code:
LIBNAME MYLIB SQLSVR CONNECTION=SHARED PRESERVE_TAB_NAMES=YES dbconinit="set ansi_warnings off;use MYDB;set nocount on;" Datasrc=MYSRC SCHEMA=dbo AUTHDOMAIN="my_sql_authentication" ;
/*---- Map the columns ----*/
proc datasets lib = work nolist nowarn memtype = (data view);
delete sql_psthru2;
quit;
proc sql;
create table work.sql_psthru2 as
select
col1,
col2,
dt_col,
col3
from MYLIB.MYTBL
where dt_col>= '25JUN2019'd
;
quit;
then I get data from the database, but if I execute
LIBNAME MYLIB SQLSVR CONNECTION=SHARED PRESERVE_TAB_NAMES=YES dbconinit="set ansi_warnings off;use MYDB;set nocount on;" Datasrc=MYSRC SCHEMA=dbo AUTHDOMAIN="my_sql_authentication" ;
proc sql;
connect using MYLIB;
create table work.sql_psthru as
select * from connection to MYLIB
(select
col1
,col2
,dt_col
from MYTBL
where dt_col >= '2019-06-25'
)
;
quit;
then I receive the error
ERROR: CLI error trying to establish connection: [DataDirect][ODBC
lib] Data source name not found and no default driver specified
immediately after the connect using MYLIB; line.
I've tried many variants of the explicit passthru which I've found all over the internet which I won't post here, but none worked.
An interesting side note is that I believe the reason the SAS DI statistics are indicating that all 16 million rows are returned is that the extract transformation auto-generates the following macro:
%macro etls_recordCheck;
%let etls_recCheckExist = %eval(%sysfunc(exist(MYLIB.MYTBL, DATA)) or
%sysfunc(exist(MYLIB.MYTBL, VIEW)));
%if (&etls_recCheckExist) %then
%do;
proc sql noprint;
select count(*) into :etls_recnt from MYLIB.MYTBL;
quit;
%end;
%mend etls_recordCheck;
%etls_recordCheck;
So the culprit in the long execution time is not that the full dataset is being returned to SAS (I removed the macro and the code still takes far too long to run).
You could try explicitly writing the code in the remote database. So if you already have a libref named LIBREF defined you can use that in the CONNECT statement in PROC SQL.
proc sql;
connect using libref ;
create table WFJVX4PU as
select * from connection to libef
(select
col1
,col2
,dt_col
,col3
from SRC_TBL
where dt_col >= &start_dt
)
;
quit;
Just make sure everything inside the () is valid syntax for that database system. Included the values of the macro variable START_DT.
I have a front end application which allows others to create elaborate SQL queries based on options selected. This SQL code is then used as a pass through query, and is displayed as part of a form.
The pass through queries created are exceptionally slow however. When ran in Access, they range from 3 minutes to 20 minutes. SQL Server Management Studio 2012 manages to complete identical queries in 45 seconds to 3 minutes. How can this time difference be minimized or eliminated?
tldr:
Time to run query in SSMS: 45 seconds
Time to run query in Access: 3 minutes
Time to open Access form containing query: 3 minutes
Structure behind the question
The queries are structured in the following manner
SELECT ...
FROM
(
SELECT ...
FROM
(
SELECT ... FROM Archive.YearXX
UNION
SELECT ... FROM Archive.YearXX
UNION
SELECT ... FROM Current.Year16
)
LEFT JOIN ReferenceData.Table ON...
)
WHERE ...
GROUP BY ...
ORDER BY ...
where the various options are able to change almost anything in the code, from which tables are pulled in the UNION segment, which fields are returned in the SELECT statement, what is included in the WHERE, GROUP BY, and ORDER BY clauses.
All SQL Servers are in the same local location, and the tables referenced range from 20 fields by 100,000 rows to 30 fields by 13,000,000 rows
Current connection string (actual data redacted):
ODBC;Driver=SQL Server;Server=[server name];DATABASE=[database];UID=[username];PWD=[password];
VBA code which creates QueryDef:
Set qdf = CurrentDb.CreateQueryDef("GeneratedReportData", GenSQL(False))
qdf.Connect = "ODBC;Driver=SQL Server;Server=[server name];DATABASE=[database];UID=[username];PWD=[password];"
qdf.ReturnsRecords = True
qdf.ODBCTimeout = 0
Constraints
I am unable to change the languages or software I am using.
There is little that can be done about the structure and use of subqueries.
We have a DB2 database in AS400. Added a linked server, all went well, however occasionally the table is locked, even when we doing some select queries. Thinking about table hint in SQL Server, does the linked server query (e.g. select * from ...) support table hint?
Doubtful, but I don't know for sure.
Are you using openquery() , or 4 part names?
A query using 4 part names like so:
select * from LNKSVRNAME.IBMINAME.MYSCHEMA.MYTABLE where somecolumn = '00335';
Pulls back all the rows from MYTABLE and does the WHERE filtering on MS SQL Server.
In contrast, using openquery() like so:
select * from openquery(LNKSVRNAME, 'select * from MYSCHEMA.MYTABLE where somecolumn = ''00335''');
Sends the query to the IBM i, and only the matching rows from MYTABLE are pulled back into MS SQL Server.
If the table is being locked exclusively, there's not much you can do. However, if you're running into row locks. You may want to look at the following DB2 for IBM i clauses
FOR READ ONLY
SKIP LOCKED DATA or USE CURRENTLY COMMITTED or WAIT FOR OUTCOME
So something like this:
select * from openquery(LNKSVRNAME, 'select * from MYSCHEMA.MYTABLE where somecolumn = ''00335'' FOR READ ONLY USE CURRENTLY COMMITTED');
Note If you are actually talking to an AS/400, FOR READ ONLY is all you'll have available. But if you're talking to a relatively recent IBM POWER System running a relatively recent version of IBM i, then the concurrent-access-resolution clauses I've shown should be available.
I want to test some schemas and indexes, and I was wondering if there is a functionality in PL/SQL Developer that can generate test data (so I won't have to create sequences and loops to insert data in the tables).
Loops and PL/SQL aren't always necessary; this trick might be helpful:
insert into emp(id, name, salary)
select rownum, 'Employee ' || to_char(rownum), dbms_random.value(2, 9) * 1000
from dual
connect by level <= 100;
will generate 100 records, named Employee 1 through Employee 100 with random "round" salaries between 2000 and 9000.
The two main techniques are:
Use of connect by level <= n to generate n rows in a query on dual.
Use of dbms_random package; there's also a very useful function dbms_random.string which can be used -- like its name suggests -- to generate random strings of a certain length containing certain characters.
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 8 years ago.
Improve this question
I'm surprised this hasn't been posted yet. Any interesting tricks that you know about in Postgres? Obscure config options and scaling/perf tricks are particularly welcome.
I'm sure we can beat the 9 comments on the corresponding MySQL thread :)
Since postgres is a lot more sane than MySQL, there are not that many "tricks" to report on ;-)
The manual has some nice performance tips.
A few other performance related things to keep in mind:
Make sure autovacuum is turned on
Make sure you've gone through your postgres.conf (effective cache size, shared buffers, work mem ... lots of options there to tune).
Use pgpool or pgbouncer to keep your "real" database connections to a minimum
Learn how EXPLAIN and EXPLAIN ANALYZE works. Learn to read the output.
CLUSTER sorts data on disk according to an index. Can dramatically improve performance of large (mostly) read-only tables. Clustering is a one-time operation: when the table is subsequently updated, the changes are not clustered.
Here's a few things I've found useful that aren't config or performance related per se.
To see what's currently happening:
select * from pg_stat_activity;
Search misc functions:
select * from pg_proc WHERE proname ~* '^pg_.*'
Find size of database:
select pg_database_size('postgres');
select pg_size_pretty(pg_database_size('postgres'));
Find size of all databases:
select datname, pg_size_pretty(pg_database_size(datname)) as size
from pg_database;
Find size of tables and indexes:
select pg_size_pretty(pg_relation_size('public.customer'));
Or, to list all tables and indexes (probably easier to make a view of this):
select schemaname, relname,
pg_size_pretty(pg_relation_size(schemaname || '.' || relname)) as size
from (select schemaname, relname, 'table' as type
from pg_stat_user_tables
union all
select schemaname, relname, 'index' as type
from pg_stat_user_indexes) x;
Oh, and you can nest transactions, rollback partial transactions++
test=# begin;
BEGIN
test=# select count(*) from customer where name='test';
count
-------
0
(1 row)
test=# insert into customer (name) values ('test');
INSERT 0 1
test=# savepoint foo;
SAVEPOINT
test=# update customer set name='john';
UPDATE 3
test=# rollback to savepoint foo;
ROLLBACK
test=# commit;
COMMIT
test=# select count(*) from customer where name='test';
count
-------
1
(1 row)
The easiest trick to let postgresql perform a lot better (apart from setting and using proper indexes of course) is just to give it more RAM to work with (if you have not done so already). On most default installations the value for shared_buffers is way too low (in my opinion). You can set
shared_buffers
in postgresql.conf. Divide this number by 128 to get an approximation of the amount of memory (in MB) postgres can claim. If you up it enough this will make postgresql fly. Don't forget to restart postgresql.
On Linux systems, when postgresql won't start again you will probably have hit the kernel.shmmax limit. Set it higher with
sysctl -w kernel.shmmax=xxxx
To make this persist between boots, add a kernel.shmmax entry to /etc/sysctl.conf.
A whole bunch of Postgresql tricks can be found here:
http://www.postgres.cz/index.php/PostgreSQL_SQL_Tricks
Postgres has a very powerful datetime handling facility thanks to its INTERVAL support.
For example:
select NOW(), NOW() + '1 hour';
now | ?column?
-------------------------------+-------------------------------
2009-04-18 01:37:49.116614+00 | 2009-04-18 02:37:49.116614+00
(1 row)
select current_date ,(current_date + interval '1 year')::date;
date | date
---------------------+----------------
2014-10-17 | 2015-10-17
(1 row)
You can cast many strings to an INTERVAL type.
COPY
I'll start. Whenever I switch to Postgres from SQLite, I usually have some really big datasets. The key is to load your tables with COPY FROM rather than doing INSERTS. See documentation:
http://www.postgresql.org/docs/8.1/static/sql-copy.html
The following example copies a table to the client using the vertical bar (|) as the field delimiter:
COPY country TO STDOUT WITH DELIMITER '|';
To copy data from a file into the country table:
COPY country FROM '/usr1/proj/bray/sql/country_data';
See also here:
Faster bulk inserts in sqlite3?
My by far favorite is generate_series: at last a clean way to generate dummy rowsets.
Ability to use a correlated value in a LIMIT clause of a subquery:
SELECT (
SELECT exp_word
FROM mytable
OFFSET id
LIMIT 1
)
FROM othertable
Abitlity to use multiple parameters in custom aggregates (not covered by the documentation): see the article in my blog for an example.
One of the things I really like about Postgres is some of the data types supported in columns. For example, there are column types made for storing Network Addresses and Arrays. The corresponding functions (Network Addresses / Arrays) for these column types let you do a lot of complex operations inside queries that you'd have to do by processing results through code in MySQL or other database engines.
Arrays are really cool once you get to know 'em.
Lets say you would like to store some hyper links between pages. You might start by thinking about creating a Table kinda like this:
CREATE TABLE hyper.links (
tail INT4,
head INT4
);
If you needed to index the tail column, and you had, say 200,000,000 links-rows (like wikipedia would give you), you would find yourself with a huge Table and a huge Index.
However, with PostgreSQL, you could use this Table format instead:
CREATE TABLE hyper.links (
tail INT4,
head INT4[],
PRIMARY KEY(tail)
);
To get all heads for a link you could send a command like this (unnest() is standard since 8.4):
SELECT unnest(head) FROM hyper.links WHERE tail = $1;
This query is surprisingly fast when it is compared with the first option (unnest() is fast and the Index is way way smaller). Furthermore, your Table and Index will take up much less RAM-memory and HD-space, especially when your Arrays are so long that they are compressed to a Toast Table. Arrays are really powerful.
Note: while unnest() will generate rows out of an Array, array_agg() will aggregate rows into an Array.
Materialized Views are pretty easy to setup:
CREATE VIEW my_view AS SELECT id, AVG(my_col) FROM my_table GROUP BY id;
CREATE TABLE my_matview AS SELECT * FROM my_view;
That creates a new table, my_matview, with the columns and values of my_view. Triggers or a cron script can then be setup to keep the data up to date, or if you're lazy:
TRUNCATE my_matview;
INSERT INTO my_matview SELECT * FROM my_view;
Inheritance..infact Multiple Inheritance (as in parent-child "inheritance" not 1-to-1 relation inheritance which many web frameworks implement when working with postgres).
PostGIS (spatial extension), a wonderful add-on that offers comprehensive set of geometry functions and coordinates storage out of the box. Widely used in many open-source geo libs (e.g. OpenLayers,MapServer,Mapnik etc) and definitely way better than MySQL's spatial extensions.
Writing procedures in different languages e.g. C, Python,Perl etc (makes your life easir to code if you're a developer and not a db-admin).
Also all procedures can be stored externally (as modules) and can be called or imported at runtime by specified arguments. That way you can source control the code and debug the code easily.
A huge and comprehensive catalogue on all objects implemented in your database (i.e. tables,constraints,indexes,etc).
I always find it immensely helpful to run few queries and get all meta info e.g. ,constraint names and fields on which they have been implemented on, index names etc.
For me it all becomes extremely handy when I have to load new data or do massive updates in big tables (I would automatically disable triggers and drop indexes) and then recreate them again easily after processing has finished. Someone did an excellent job of writing handful of these queries.
http://www.alberton.info/postgresql_meta_info.html
Multiple schemas under one database, you can use it if your database has large number of tables, you can think of schemas as categories. All tables (regardless of it's schema) have access to all other tables and functions present in parent db.
You don't need to learn how to decipher "explain analyze" output, there is a tool: http://explain.depesz.com
select pg_size_pretty(200 * 1024)
pgcrypto: more cryptographic functions than many programming languages' crypto modules provide, all accessible direct from the database. It makes cryptographic stuff incredibly easy to Just Get Right.
A database can be copied with:
createdb -T old_db new_db
The documentation says:
this is not (yet) intended as a general-purpose "COPY DATABASE" facility
but it works well for me and is much faster than
createdb new_db
pg_dump old_db | psql new_db
Memory storage for throw-away data/global variables
You can create a tablespace that lives in the RAM, and tables (possibly unlogged, in 9.1) in that tablespace to store throw-away data/global variables that you'd like to share across sessions.
http://magazine.redhat.com/2007/12/12/tip-from-an-rhce-memory-storage-on-postgresql/
Advisory locks
These are documented in an obscure area of the manual:
http://www.postgresql.org/docs/9.0/interactive/functions-admin.html
It's occasionally faster than acquiring multitudes of row-level locks, and they can be used to work around cases where FOR UPDATE isn't implemented (such as recursive CTE queries).
This is my favorites list of lesser know features.
Transactional DDL
Nearly every SQL statement is transactional in Postgres. If you turn off autocommit the following is possible:
drop table customer_orders;
rollback;
select *
from customer_orders;
Range types and exclusion constraint
To my knowledge Postgres is the only RDBMS that lets you create a constraint that checks if two ranges overlap. An example is a table that contains product prices with a "valid from" and "valid until" date:
create table product_price
(
price_id serial not null primary key,
product_id integer not null references products,
price numeric(16,4) not null,
valid_during daterange not null
);
NoSQL features
The hstore extension offers a flexible and very fast key/value store that can be used when parts of the database need to be "schema-less". JSON is another option to store data in a schema-less fashion and
insert into product_price
(product_id, price, valid_during)
values
(1, 100.0, '[2013-01-01,2014-01-01)'),
(1, 90.0, '[2014-01-01,)');
-- querying is simply and can use an index on the valid_during column
select price
from product_price
where product_id = 42
and valid_during #> date '2014-10-17';
The execution plan for the above on a table with 700.000 rows:
Index Scan using check_price_range on public.product_price (cost=0.29..3.29 rows=1 width=6) (actual time=0.605..0.728 rows=1 loops=1)
Output: price
Index Cond: ((product_price.valid_during #> '2014-10-17'::date) AND (product_price.product_id = 42))
Buffers: shared hit=17
Total runtime: 0.772 ms
To avoid inserting rows with overlapping validity ranges a simple (and efficient) unique constraint can be defined:
alter table product_price
add constraint check_price_range
exclude using gist (product_id with =, valid_during with &&)
Infinity
Instead of requiring a "real" date far in the future Postgres can compare dates to infinity. E.g. when not using a date range you can do the following
insert into product_price
(product_id, price, valid_from, valid_until)
values
(1, 90.0, date '2014-01-01', date 'infinity');
Writeable common table expressions
You can delete, insert and select in a single statement:
with old_orders as (
delete from orders
where order_date < current_date - interval '10' year
returning *
), archived_rows as (
insert into archived_orders
select *
from old_orders
returning *
)
select *
from archived_rows;
The above will delete all orders older than 10 years, move them to the archived_orders table and then display the rows that were moved.
1.) When you need append notice to query, you can use nested comment
SELECT /* my comments, that I would to see in PostgreSQL log */
a, b, c
FROM mytab;
2.) Remove Trailing spaces from all the text and varchar field in a database.
do $$
declare
selectrow record;
begin
for selectrow in
select
'UPDATE '||c.table_name||' SET '||c.COLUMN_NAME||'=TRIM('||c.COLUMN_NAME||') WHERE '||c.COLUMN_NAME||' ILIKE ''% '' ' as script
from (
select
table_name,COLUMN_NAME
from
INFORMATION_SCHEMA.COLUMNS
where
table_name LIKE 'tbl%' and (data_type='text' or data_type='character varying' )
) c
loop
execute selectrow.script;
end loop;
end;
$$;
3.) We can use a window function for very effective removing of duplicate rows:
DELETE FROM tab
WHERE id IN (SELECT id
FROM (SELECT row_number() OVER (PARTITION BY column_with_duplicate_values), id
FROM tab) x
WHERE x.row_number > 1);
Some PostgreSQL's optimized version (with ctid):
DELETE FROM tab
WHERE ctid = ANY(ARRAY(SELECT ctid
FROM (SELECT row_number() OVER (PARTITION BY column_with_duplicate_values), ctid
FROM tab) x
WHERE x.row_number > 1));
4.) When we need to identify server's state, then we can use a function:
SELECT pg_is_in_recovery();
5.) Get functions's DDL command.
select pg_get_functiondef((select oid from pg_proc where proname = 'f1'));
6.) Safely changing column data type in PostgreSQL
create table test(id varchar );
insert into test values('1');
insert into test values('11');
insert into test values('12');
select * from test
--Result--
id
character varying
--------------------------
1
11
12
You can see from the above table that I have used the data type – ‘character varying’ for ‘id’
column. But it was a mistake, because I am always giving integers as id. So using varchar here is a
bad practice. So let’s try to change the column type to integer.
ALTER TABLE test ALTER COLUMN id TYPE integer;
But it returns:
ERROR: column “id” cannot be cast automatically to type integer SQL
state: 42804 Hint: Specify a USING expression to perform the
conversion
That means we can’t simply change the data type because data is already there in the column. Since the data is of type ‘character varying’ postgres cant expect it as integer though we entered integers only. So now, as postgres suggested we can use the ‘USING’ expression to cast our data into integers.
ALTER TABLE test ALTER COLUMN id TYPE integer USING (id ::integer);
It Works.
7.) Know who is connected to the Database
This is more or less a monitoring command. To know which user connected to which database
including their IP and Port use the following SQL:
SELECT datname,usename,client_addr,client_port FROM pg_stat_activity ;
8.) Reloading PostgreSQL Configuration files without Restarting Server
PostgreSQL configuration parameters are located in special files like postgresql.conf and pg_hba.conf. Often, you may need to change these parameters. But for some parameters to take effect we often need to reload the configuration file. Of course, restarting server will do it. But in a production environment it is not preferred to restarting the database, which is being used by thousands, just for setting some parameters. In such situations, we can reload the configuration files without restarting the server by using the following function:
select pg_reload_conf();
Remember, this wont work for all the parameters, some parameter
changes need a full restart of the server to be take in effect.
9.) Getting the data directory path of the current Database cluster
It is possible that in a system, multiple instances(cluster) of PostgreSQL is set up, generally, in different ports or so. In such cases, finding which directory(physical storage directory) is used by which instance is a hectic task. In such cases, we can use the following command in any database in the cluster of our interest to get the directory path:
SHOW data_directory;
The same function can be used to change the data directory of the cluster, but it requires a server restarts:
SET data_directory to new_directory_path;
10.) Find a CHAR is DATE or not
create or replace function is_date(s varchar) returns boolean as $$
begin
perform s::date;
return true;
exception when others then
return false;
end;
$$ language plpgsql;
Usage: the following will return True
select is_date('12-12-2014')
select is_date('12/12/2014')
select is_date('20141212')
select is_date('2014.12.12')
select is_date('2014,12,12')
11.) Change the owner in PostgreSQL
REASSIGN OWNED BY sa TO postgres;
12.) PGADMIN PLPGSQL DEBUGGER
Well explained here
It's convenient to rename an old database rather than mysql can do. Just using the following command:
ALTER DATABASE name RENAME TO new_name