Imagine a table with the following structure on PostgreSQL 9.0:
create table raw_fact_table (text varchar(1000));
For the sake of simplification I only mention one text column, in reality it has a dozen. This table has 10 billion rows and each column has lots of duplicates. The table is created from a flat file (csv) using COPY FROM.
To increase performance I want to convert to the following star schema structure:
create table dimension_table (id int, text varchar(1000));
The fact table would then be replaced with a fact table like the following:
create table fact_table (dimension_table_id int);
My current method is to essentially run the following query to create the dimension table:
Create table dimension_table (id int, text varchar(1000), primary key(id));
then to create fill the dimension table I use:
insert into dimension_table (select null, text from raw_fact_table group by text);
Afterwards I need to run the following query:
select id into fact_table from dimension inner join raw_fact_table on (dimension.text = raw_fact_table.text);
Just imagine the horrible performance I get by comparing all strings to all other strings several times.
On MySQL I could run a stored procedure during the COPY FROM. This could create a hash of a string and all subsequent string comparison is done on the hash instead of the long raw string. This does not seem to be possible on PostgreSQL, what do I do then?
Sample data would be a CSV file containing something like this (I use quotes also around integers and doubles):
"lots and lots of text";"3";"1";"2.4";"lots of text";"blabla"
"sometext";"30";"10";"1.0";"lots of text";"blabla"
"somemoretext";"30";"10";"1.0";"lots of text";"fooooooo"
Just imagine the horrible performance
I get by comparing all strings to all
other strings several times.
When you've been doing this a while, you stop imagining performance, and you start measuring it. "Premature optimization is the root of all evil."
What does "billion" mean to you? To me, in the USA, it means 1,000,000,000 (or 1e9). If that's also true for you, you're probably looking at between 1 and 7 terabytes of data.
My current method is to essentially
run the following query to create the
dimension table:
Create table dimension_table (id int, text varchar(1000), primary key(id));
How are you gonna fit 10 billion rows into a table that uses an integer for a primary key? Let's even say that half the rows are duplicates. How does that arithmetic work when you do it?
Don't imagine. Read first. Then test.
Read Data Warehousing with PostgreSQL. I suspect these presentation slides will give you some ideas.
Also read Populating a Database, and consider which suggestions to implement.
Test with a million (1e6) rows, following a "divide and conquer" process. That is, don't try to load a million at a time; write a procedure that breaks it up into smaller chunks. Run
EXPLAIN <sql statement>
You've said you estimate at least 99% duplicate rows. Broadly speaking, there are two ways to get rid of the dupes
Inside a database, not necessarily the same platform you use for production.
Outside a database, in the filesystem, not necessarily the same filesystem you use for production.
If you still have the text files that you loaded, I'd consider first trying outside the database. This awk one-liner will output unique lines from each file. It's relatively economical, in that it makes only one pass over the data.
awk '!arr[$0]++' file_with_dupes > file_without_dupes
If you really have 99% dupes, by the end of this process you should have reduced your 1 to 7 terabytes down to about 50 gigs. And, having done that, you can also number each unique line and create a tab-delimited file before copying it into the data warehouse. That's another one-liner:
awk '{printf("%d\t%s\n", NR, $0);}' file_without_dupes > tab_delimited_file
If you have to do this under Windows, I'd use Cygwin.
If you have to do this in a database, I'd try to avoid using your production database or your production server. But maybe I'm being too cautious. Moving several terabytes around is an expensive thing to do.
But I'd test
SELECT DISTINCT ...
before using GROUP BY. I might be able to do some tests on a large data set for you, but probably not this week. (I don't usually work with terabyte-sized files. It's kind of interesting. If you can wait.)
Just to questions:
- it neccessary to convert your data in 1 or 2 steps?
- May we modify the table while converting?
Running more simplier queries may improve your performance (and the server load while doing it)
One approach would be:
generate dimension_table (If i understand it correctly, you don't have performance problems with this) (maybe with an additional temporary boolean field...)
repeat: choose one previously not selected entry from dimension_table, select every rows from raw_fact_table containing it and insert them into fact_table. Mark dimension_table record as done, and next... You can write this as a stored procedure, and it can convert your data in the background, eating minimal resources...
Or another (probably better):
create fact_table as EVERY record from raw_fact_table AND one dimension_id. (so including dimension_text and dimension_id rows)
create dimension_table
create an after insert trigger for fact_table which:
searches for dimension_text in fact_table
if not found, creates a new record in dimension_table
updates dimension_id to this id
in a simle loop, insert every record from raw_fact_table to fact_table
You are omitting some details there at the end, but I don't see that there necessarily is a problem. It is not in evidence that all strings are actually compared to all other strings. If you do a join, PostgreSQL could very well pick a smarter join algorithm, such as a hash join, which might give you the same hashing that you are implementing yourself in your MySQL solution. (Again, your details are hazy on that.)
-- add unique index
CREATE UNIQUE INDEX uidx ON dimension_table USING hash(text);
-- for non case-sensitive hash(upper(text))
try hash(text); and btree(text) to see which one is faster
I an see several ways of solving your problem
There is md5 function in PostgreSql
md5(string) Calculates the MD5 hash of string, returning the result in hexadecimal
insert into dimension_table (select null, md5(text), text from raw_fact_table group by text)
add md5 field into raw_fact_table as well
select id into fact_table from dimension inner join raw_fact_table on (dimension.md5 = raw_fact_table.md5);
Indexes on MD5 filed might help as well
Or you can calculate MD5 on the fly while loading the data.
For example our ETL tool Advanced ETL processor can do it for you.
Plus it can load data into multiple tables same time.
There is a number of on-line tutorials available on our web site
For example this one demonstrates loading slow changing dimension
http://www.dbsoftlab.com/online-tutorials/advanced-etl-processor/advanced-etl-processor-working-with-slow-changing-dimension-part-2.html
Related
I am writing a query with code to select all records from a table where a column value is contained in a CSV. I found a suggestion that the best way to do this was using ARRAY functionality in PostgresQL.
I have a table price_mapping and it has a primary key of id and a column customer_id of type bigint.
I want to return all records that have a customer ID in the array I will generate from csv.
I tried this:
select * from price_mapping
where ARRAY[customer_id] <# ARRAY[5,7,10]::bigint[]
(the 5,7,10 part would actually be a csv inserted by my app)
But I am not sure that is right. In application the array could contain 10's of thousands of IDs so want to make sure I am doing right with best performance method.
Is this the right way in PostgreSQL to retrieve large collection of records by pre-defined column value?
Thanks
Generally this is done with the SQL standard in operator.
select *
from price_mapping
where customer_id in (5,7,10)
I don't see any reason using ARRAY would be faster. It might be slower given it has to build arrays, though it might have been optimized.
In the past this was more optimal:
select *
from price_mapping
where customer_id = ANY(VALUES (5), (7), (10)
But new-ish versions of Postgres should optimize this for you.
Passing in tens of thousands of IDs might run up against a query size limit either in Postgres or your database driver, so you may wish to batch this a few thousand at a time.
As for the best performance, the answer is to not search for tens of thousands of IDs. Find something which relates them together, index that column, and search by that.
If your data is big enough, try this:
Read your CSV using a FDW (foreign data wrapper)
If you need this connection often, you might build a materialized view from it, holding only needed columns. Refresh this when new CSV is created.
Join your table against this foreign table or materialized viev.
I am looking for a way to create a Redshift query that will retrieve data from a table that is generated daily. Tables in our cluster are of the form:
event_table_2016_06_14
event_table_2016_06_13
.. and so on.
I have tried writing a query that appends the current date to the table name, but this does not seem to work correctly (invalid operation):
SELECT * FROM concat('event_table_', to_char(getdate(),'YYYY_MM_DD'))
Any suggestions on how this can be performed are greatly appreciated!
I have tried writing a query that appends the current date to the
table name, but this does not seem to work correctly (invalid
operation):
Redshift does not support that. But you most likely won't need it.
Try the following (expanding on the answer from #ketan):
Create your main table with appropriate (for joins) DIST key, and COMPOUND or simple SORT KEY on timestamp column, and proper compression on columns.
Daily, create a temp table (use CREATE TABLE ... LIKE - this will preserve DIST/SORT keys), load it with daily data, VACUUM SORT.
Copy sorted temp table into main table using ALTER TABLE APPEND - this will copy the data sorted, and will reduce VACUUM on the main table. You may still need VACUUM SORT after that.
After that query your main table normally, probably giving it a range on timestamp. Redshift is optimised for these scenarios, and 99% of times you don't need to optimise table scans yourself - even on tables with billion of rows scans take milliseconds to few seconds. You may need to optimise elsewhere, but that's the second step.
To get insight in the performance of scans, use STL_QUERY system table to find your query ID, and then use STL_SCAN (or SVL_QUERY_SUMMARY) table to see how fast the scan was.
Your example is actually the main use case for ALTER TABLE APPEND.
I am assuming that you are creating a new table everyday.
What you can do is:
Create a view on top of event_table_* tables. Query your data using this view.
Whenever you create or drop a table, update the view.
If you want, you can avoid #2: Instead of creating a new table everyday, create empty tables for next 1-2 years. So, no need to update the view every day. However, do remember that there is an upper limit of 9,900 tables in Redshift.
Edit: If you always need to query today's table (instead of all tables, as I assumed originally), I don't think you can do that without updating your view.
However, you can modify your design to have just one table, with date as sort-key. So, whenever your table is queried with some date, all disk blocks that don't have that date will be skipped. That'll be as efficient as having time-series tables.
Considering a simple database containing a table foo (name text, description text) and a table bar (key text, other_info text) which share a many-to-many relationship (foo_bar (foo_id, bar_id) ).
The fact is, I want to obtain foo rows when a bar.key contains a partial pattern for example 'a' or '1234'. So basically I want a JOIN and a LIKE '%a%'. And it could be the end of my story...
However, my table foo contains 10 millions rows and queries take time... Moreover, the table will soon have hundreds millions of rows! In the production environnement, the database is read-only mode.
So I started thinking about materialized view as:
foo_view (name text, description text, bar_key text)
which saves me the JOIN.
And I started using trigram indexes to have faster partial matches, and it works.
But it's not enough...
I started thinking about ARRAY to saves me the JOIN and the DISTINCT part as
foo_view (name text unique, description text, bar_keys text[])
I used the parray_gin extension to have partial searches and indexes with arrays. For now, this is my faster solution.
But it is not enough... (for one million rows it's good but not for more...)
I started thinking about partionnement but with ten millions rows I don't see a significative difference...
My database is not a complex one and I can create as table/view as necessary. My problem is: I will have to query hundreds millions of rows using a partial search and I can't manage to do it fast.
I know that the hardware has a crucial part in performance, but that's not something I can change. Maybe I will later. But for now, my question is strictly software-side.
I need to know if the solutions I tried were good ones or if I missed something.
I have seen performance tweaks for delete on normal tables in t-sql.
But are there performance tweaks on deletes on table variables to be done?
EDIT
Here's an example:
The plot gets thicker, as UserExclusionsEvaluate is actually a CTE, but I'm going to try and optimise it around the table variable first (if possible). The CTE itself runs very quickly. Just the delete that is slow.
DELETE FROM #UsersCriteria
FROM #UsersCriteria UsersCriteria
WHERE UserId IN (SELECT UserID FROM UserExclusionsEvaluate WHERE PushRuleExclusionMet = 1)
In it's current incarnation, #UsersCriteria is:
DECLARE #UsersCriteria TABLE
(
UserId int primary key,
PushRuleUserCriteriaType int
)
I've tried #UsersCriteria as non primary and experimented with clustered non-clustered.
It's probable also the problem is with the IN. I've also tried a JOIN on a subquery.
EDIT:
GOOD NEWS!
After lots of playing with the SQL, including moving the suquery into a chained CTE, attempting table hints etc etc etc.
A simple change from a table variable to a temp table dramatically improved the performance.
Which is really interesting, as deletes ran fine byself, the subquery (on the CTE) ran fine byitself. But mixing the two ran mega slow.
I'm guessing that the optimiser can't kick in when using a CTE in a subquery? Maybe when mixed with the delete.
Not really.
Unless you defined a PK in the DECLARE which may work: there are no statistics for table variables and the table is assumed to have 1 row only
Well there is a limited amount you can do. However, if you have a large data set in the table variable, you should be using a temp table instead if you need better performance.
You could also do the deletes in batches (say 1000 at a time).
Otherwise, show us your delete statment and we'll see if we see anything that can be imporved.
NO.
Table variables are unindexable and transient. They have no statistics.
They are not intended to store very large amounts of data.
If you have a table variable that is big enough to give you performance problems when you delete from it, you're using them in an unintended way. Put that data into a #Temp table or a real table so you have more control.
If you're creating a temporary table within a stored procedure and want to add an index or two on it, to improve the performance of any additional statements made against it, what is the best approach? Sybase says this:
"the table must contain data when the index is created. If you create the temporary table and create the index on an empty table, Adaptive Server does not create column statistics such as histograms and densities. If you insert data rows after creating the index, the optimizer has incomplete statistics."
but recently a colleague mentioned that if I create the temp table and indices in a different stored procedure to the one which actually uses the temporary table, then Adaptive Server optimiser will be able to make use of them.
On the whole, I'm not a big fan of wrapper procedures that add little value, so I've not actually got around to testing this, but I thought I'd put the question out there, to see if anyone had any other approaches or advice?
A few thoughts:
If your temporary table is so big that you have to index it, then is there a better way to solve the problem?
You can force it to use the index (if you are sure that the index is the correct way to access the table) by giving an optimiser hint, of the form:
SELECT *
FROM #table (index idIndex)
WHERE id = #id
If you are interested in performance tips in general, I've answered a couple of other questions about that at some length here:
Favourite performance tuning tricks
How do you optimize tables for specific queries?
What's the problem with adding the indexes after you put data into the temp table?
One thing you need to be mindful of is the visibility of the index to other instances of the procedure that might be running at the same time.
I like to add a guid to these kinds of temp tables (and to the indexes), to make sure there is never a conflict. The other benefit of this approach is that you could simply make the temp table a real table.
Also, make sure that you will need to query the data in these temp tables more than once during the running of the stored procedure, otherwise the cost of index creation will outweigh the benefit to the select.
In Sybase if you create a temp table and then use it in one proc the plan for the select is built using an estimate of 100 rows in the table. (The plan is built when the procedure starts before the tables are populated.) This can result in the temp table being table scanned since it is only "100 rows". Calling a another proc causes Sybase to build the plan for the select with the actual number of rows, this allows the optimizer to pick a better index to use. I have seen significant improvedments using this approach but test on your database as sometimes there is no difference.