I am applying a group of data mining algorithms to a dataset comprised of a set of customers along with a large number of descriptive attributes that summarize various aspects of their past behavior. There are more than 10,000 attributes, each stored as a column in a table with the customer id as the primary key. For several reasons, it is necessary to pre-compute these attributes rather than calculating them on the fly. I generally try to select customer with a specified attribute set. The algorithms can combine any arbitrary number of these attributes together in a single SELECT statement and join the required tables. All the tables have the same number of rows (one per customer).
I am wondering what's the best way to structure these tables of attributes. Is it better to group the attributes into tables of 20-30 columns, requiring more joins on average but fewer columns per SELECT, or have tables with the maximum number of columns to minimize the number of joins, but having potentially all 10K columns joined at once?
I also thought of using one giant 3-column customerID-attribute-value table and storing all the info there, but it would be harder to structure a "select all customers with these attributes-type query that I need."
I'm using MySQL 5.0+, but I assume this is a general SQL-ish question.
From my expirience using tables with 10,000 columns is very-very-very bad idea. What if in future this number will be increased?
If there are a lot of attributes you shouldn't use a horizontal scaled tables (with large number of columns). You should create a new table attributes and place alltributes values into it. Then connect this table with Many-To-One relationship to main entry table
Maybe the second way is to use no-SQL (like MongoDB) systems
As #odiszapc said, you have to use a meta-model structure, like for instance:
CREATE TABLE customer(ID INT NOT NULL PRIMARY KEY, NAME VARCHAR(64));
CREATE TABLE customer_attribute(ID INT NOT NULL, ID_CUSTOMER INT NOT NULL, NAME VARCHAR(64), VALUE VARCHAR(1024));
Return basic informations of given customer:
SELECT * FROM customers WHERE name='John';
Return customer(s) matching certain attributes:
SELECT c.*
FROM customer c
INNER JOIN attribute a1 ON a1.id_customer = c.id
AND a1.name = 'address'
AND a1.value = '1078, c/ los gatos madrileños'
INNER JOIN attribute a2 ON a2.id_customer = c.id
AND a2.name = 'age'
AND a2.value = '27'
Your generator should generate the inner joins on the fly.
Proper indexes on the tables should allow all this engine to go relatively fast (if we assume 10k attributes per customer, and 10k customers, that's actually pretty much a challenge...)
10,000 columns is much. The SELECT statement will be very long and messy if you wouldn't use *. I think you can narrow the attributes down to most useful and meaningful ones, eliminating others
Related
Currently, this is what my SELECT code looks like:
SELECT student.stu_code, user.f_name, user.l_name
FROM user
INNER JOIN student
ON student.stu_code = user.user_id
INNER JOIN course
ON course.stu_code ?????;
Basically, to elaborate the student table inherits from user table, therefore I had user_id = stu_code. What I'm confused about is how to join course table with student table.
Let's say that the course table has a course code (PK), a few other attributes and a stu_code column, however, the student code column has multiple values inside a single column to represent that multiple students are taking the course and stored as VARCHAR.
Example: Student table has stu_code string value of '123' and course table has a stu_code with string value of '123, 246, 369'.
How would I go about joining these two tables together and separating the stu_code in the course table so that it represents 3 separate stu_code values -> i.e. '123', '246', '369'.
Any help is greatly appreciated!
however, the student code column has multiple values inside a single column to represent that multiple students are taking the course and stored as VARCHAR.
Your data model is broken. Put your effort into fixing the data model. You want a junction/association table courseStudents or perhaps enrolled, with columns like:
stu_code (foreign key to students)
course_code (foreign key to students)
enrollment_date
and so on
What is wrong with your data model? Here are a few things:
You are storing numbers as a string.
You are putting multiple values into a string column.
You cannot define foreign key relationships.
SQL has poor string handling capabilities.
SQL has a great way to store lists of things. It is not called "string". It is called "table".
Your data model is ~broken~ hindering you from elegant solutions.
You cannot join your two tables efficiently. While they might both contain strings they do not contain data with the same rules. Thus, you must transform the data in order to join them so you could do this in a few ways but one way is using regular expression function.
You can use it to evaluate a test on whether the stu_code matches the list of codes. Further, you can do this dynamically ... constructing the test string itself based upon values from the left and right
join based on REGEXP
SELECT student.stu_code, user.f_name, user.l_name
FROM user
INNER JOIN student
ON student.stu_code = user.user_id
INNER JOIN course
ON student.stu_code REGEXP CONCAT('[[:<:]]',course.stu_code,'[[:>:]]')
Assuming tables and data:
Student
- - - -
stu_code
123
Course
- - - -
stu_code
'123, 246, 369'
Example:
http://sqlfiddle.com/#!9/672b57f/4
about the regular expression
in mysql the regex syntax can be a little bit different. [[:<:]] is the character class in spencer notation for word boundary.
if you have a new enough version of mysql/mariadb you can use more typical ICU notation of \b.
more about that here : https://dev.mysql.com/doc/refman/8.0/en/regexp.html
about efficiency
in large datasets the performance will be awful. you will have to scan all records and you will have to perform the function on all of them. In a large set you might get some gains by joining on like first (which is faster than regexp). This will be much faster at filtering-out and then the regexp can deal with filtering-in.
Perhaps your model was based upon an assumption of having a courses table with very few rows?
It ironic because you have made your course table unnecessarily large. You would actually be better off with an intermediary table that represents the many-to-many nature (the fact that students can take many courses and courses can have many students) with 1 row per unique relationship. While this table would be an order of magnitude "longer" it would be leaner and it could be indexed and query performance would be faster.
The courses table does not need to have any awareness of the student list and thus you can alter courses by removing courses.stu_code once you change the model (aside: It might be useful if courses cached a hint of the expected student count for that course)
possible link table
would be a new table like this (note how it only ever needs these 2 columns)
stu_course_lnk
- - - - - - - -
stu_code course_id
123 ABC
124 ABC
...
123 XYZ
...
124 LMN
then you add joins of
...
student.stu_code = stu_course_lnk.stu_code
and
stu_course_lnk.course_id = course.id
...
I have two tables - books and images. The books table has many columns - including id (primary key), name (which is not unique), releasedate, etc. The images table have two columns - id (which is not unique, i.e one book id may have multiple images associated with it, and we need all those images. This column has a non-unique index), and poster (which is unique primary key, all images lie in the same bucket, hence cannot have duplicate names). My requirement is given a book name, find all images associated with it (along with the year of release and the bucketname for each image, the bucketname being just a number in this case).
I am running this query:
select books.id,poster,bucketname,year(releasedate) from books
inner join images where images.bookId = books.id and books.name = "<name>";
A sample result set may look like this:
As you can see there are two results matching - one with id 2 and year 1989, having 5 images, other one with id 261009, year 2013 and one image.
The problem is, the query is extremely slow. It takes around .14 seconds from MySQL console itself, under zero load (in production there may be several concurrent requests and they may be queued, leading to further delay), which is unacceptable for autocomplete. Can anyone tell me how to optimize the query by adding correct indices/keys to the tables? If it is not possible from MySQL, suggestions regarding a proper Redis schema would be useful as well.
Edit: Approx no. of rows in images - 480k, in books - 285k. In future, autocomplete will show result for book authors as well as book names, hence the query will need to expand to take into account a separate table authors where each author will have an id and name, just like a book.
For optimal performance, you want suitable covering indexes available. For example:
... on `books` (`name`,`id`,`releasedate`)
... on `images` (`bookid`,`poster`,`bucketname`)
We want name as the leading column in the index, because of the equality predicate in the WHERE clause. We want id and releasedate also included in the index to make it a "covering index", so the query can be satisfied from the index, without a need to visit pages of the underlying table to retrieve values.
We want bookid as the leading column because of the reference in the ON clause. Again, having poster and bucketname available right in the index make it a "covering" index.
Use EXPLAIN to see the query execution plan.
Also, note that the inner join operation won't return a row from books if a matching row in images is not found. If we want to return a row from books even when no image is available, we could use an outer join.
I'd write the query like this:
SELECT b.id
, i.poster
, i.bucketname
, YEAR(b.releasedate)
FROM books b
LEFT
JOIN images i
ON i.bookid = b.id
WHERE b.name = ?
I'm running into all kinds of thought problems while planning my database:
Outline:
The database is a patient database with a large number of patients.
Each patient has tons of data, eg: bloodpressure values on different dates.
Questions:
Would it be easier to create tables per patient e.g.
"bob_builder_BPvalues" or to create one table for the BP values eg. "BP_values" and then have all the patients values in there linked via foreign keys?
As I have so much data per patient, it does not seem to make sense to mix blood pressure value of each patient into one single table as this would look very messy to a human. Which approach would be faster in terms of processing and sorting through the data?
Let's say you have 10 patients:
With your first approach, you'd end up with 10 different tables always containing the same type of data.
For each query on a single patient, you would have to build a dynamic query joining to the right table:
SELECT ...
FROM patients
INNER JOIN bobby_measures ON ... -- this has to be crafted dynamically each time
WHERE patients.name = 'bobby'
And what if you want to make some stats on some kind of data for a range of dates for all patients ? Querying this becomes a nightmare, even with 10 patients. So guess what happens when you have 1000...
On the other hand, your second choice makes (arguably) human reading of the database more difficult. But being read by a human is not one of the objectives of databases.
With a single patientData table (or as many tables you want, one per datatype if needed, bloodPressure and stuff), everything becomes simpler. You can query any patient using the same query, changing only the patient id, you can make all the queries you want for a range of dates, filtering on some datatype, or whatever.
SELECT ...
FROM patients
INNER JOIN patientData ON ...
WHERE patients.name in ('bobby', 'joe'...)
AND patientData.type = 'blood pressure'
AND patientData.date BETWEEN ... AND ...
-- and so on
Using the right indices on the patientData table(s) and an appropriate presentation layer, all this data becomes totally readable by an average user.
Have a single table for all patients. This can then link to a BloodPressure table using a foreign key. The relationship between ...
Patient 1----* BloodPressureResults
So a single patient can have many blood pressure results.
You would then be able to view the blood pressure results for a specific patient by using a simple query...
SELECT * FROM BloodPressureResults
WHERE Patient_Id = '1'
This would then return you all of the blood pressure results for the patient with an Id of 1.
You would then also be able to add other tables like WeightResults or BloodTestResults in the same way as the BloodPressureResults table
BACKGROUND:
I am developing a search feature that enables users to search three entities at once: classes, student organizations, and events by name.
MySQL 5.6 and Innodb will be utilized on a relatively small server.
TABLES:
Search
entity_id (tiny_int)
entity_type (tiny_int)
full_name (varchar(255))
Index (Primary) -> entity_id, entity_type
Index (FULL TEXT INDEX) -> full_name
Class (entity_type = 1)
class_id
ALL OTHER COLUMNS...
Events (entity_type = 2)
event_id
ALL OTHER COLUMNS
Orgs (entity_type = 3)
org_id
ALL OTHER COLUMNS
QUESTION:
Is it appropriate to index the name of 3 seemingly similar but different data sets (class, event, organization) into one search table through the use of entity_id, entity_type Primary Key?
How would I go about doing a join on the search table with the Class, Events, Orgs tables? Is there a conditional I can do using the entity_type?
Any help or guidance is greatly appreciated.
This seems like a reasonable approach to linking these three table types. It's v similar to what you often see ppl do with 'account_type: user, admin, other' and links to separate settings for each.
To do a join between the main table and all three subs will depend on how similar the columns are. If they are v similar you could do three joins in the same query and put the entity_type as part of the on clause. If they are v different then you may be better served by splitting up the query into 3 distinct joins. You could union these together but it may or may not be worth it.
I'm wondering if this its even posible.
I want to join 2 tables based on the data of table 1.
Example table 1 has column food with its data beeing "hotdog".
And I have a table called hotdog.
IS it possible to do a JOIN like.
SELECT * FROM table1 t join t.food on id = foodid
I know it doesnt work but, its even posible, is there a work arround?.
Thanks in advance.
No, you can't join to a different table per row in table1, not even with dynamic SQL as #Cade Roux suggests.
You could join to the hotdog table for rows where food is 'hotdog' and join to other tables for other specific values of food.
SELECT * FROM table1 JOIN hotdog ON id = foodid WHERE food = 'hotdog'
UNION
SELECT * FROM table1 JOIN apples ON id = foodid WHERE food = 'apples'
UNION
SELECT * FROM table1 JOIN soups ON id = foodid WHERE food = 'soup'
UNION
...
This requires that you know all the distinct values of food, and that all the respective food tables have compatible columns so you can UNION them together.
What you're doing is called polymorphic associations. That is, the foreign key in table1 references rows in multiple "parent" tables, depending on the value in another column of table1. This is a common design mistake of relational database programmers.
For alternative solutions, see my answers to:
Possible to do a MySQL foreign key to one of two possible tables?
Why can you not have a foreign key in a polymorphic association?
I also cover solutions for polymorphic associations in my presentation Practical Object Oriented Models In SQL, and in my book SQL Antipatterns Volume 1: Avoiding the Pitfalls of Database Programming.
Only with dynamic SQL. It is also possible to left join many different tables and use CASE based on type, but the tables would be all have to be known in advance.
It would be easier to recommend an appropriate design if we knew more about what you are trying to achieve, what your design currently looks like and why you've chosen that particular table design in the first place.
-- Say you have a table of foods:
id INT
foodtype VARCHAR(50) (right now it just contains 'hotdog' or 'hamburger')
name VARCHAR(50)
-- Then hotdogs:
id INT
length INT
width INT
-- Then hamburgers:
id INT
radius INT
thickness INT
Normally I would recommend some system for constraining only one auxiliary table to exist, but for simplicity, I'm leaving that out.
SELECT f.*, hd.length, hd.width, hb.radius, hb.thickness
FROM foods f
LEFT JOIN hotdogs hd
ON hd.id = f.id
AND f.foodtype = 'hotdog'
LEFT JOIN hamburgers hb
ON hb.id = f.id
AND f.foodtype = 'hamburger'
Now you will see that such a thing can be code generated (or even for a very slow prototype dynamic SQL on the fly) from SELECT DISTINCT foodtype FROM foods given certain assumptions about table names and access to the table metadata.
The problem is that ultimately whoever consumes the result of this query will have to be aware of new columns showing up whenever a new table is added.
So the question moves back to your client/consumer of the data - how is it going to handle the different types? And what does it mean for different types to be in the same set? And if it needs to be aware of the different types, what's the drawback of just writing different queries for each type or changing a manual query when new types are added given the relative impact of such a change anyway?