Database Design

Useful tips for designing robust database schema. This guide is more of a decision reference (1) for people wanting to design better database schema for startups (2).

Decision reference: You will probably come across any one of the problems below when designing database schema (e.g. should I use JSONB? How do I design tagging schema? How do I keep historical records?) and be presented with different options and trade-offs. If two standards are equal (should table name be singular or plural), then it's up to you to pick one and make it a standard and keeping it consistent. Rather than giving you a it depends answer, this guide is meant to share the what-ifs, that is the decision I made, and the outcome. Yes, the codes are actually being written and used in different applications I wrote. You don't have to agree with the approaches - I used to take some of the more complex approaches (for the sake of best practice), but over time I realised that it is unnecessary. Simple is best.
Startups: A lot of startups starts by using ORMs or frameworks that provides a lot of convenience when it comes to dealing with database. If you are thinking hey, we need to be agile, that's why we are okay introducing some technical debts let me tell you something, you can be both fast and produce quality work (not compromising database design). Most of the time, when new features are being introduced, or when new situation aroses (e.g. analytics are not accurate, because we did not store timestamp with timezones in database, using integer id instead of uuid, and people start hacking your system, etc), the application just won't cut it anymore. The last thing I want to live with is a poorly designed database schema. You can switch language and frameworks for your frontend/backend servers, but if you design your database wrongly, you have to live with it. Also, I mentioned startups because my knowledge on database design is limited to a user scale of <5m. Also, the approach taken by larger companies might vary, and the technology used might differ as they are focusing a lot on performance, reliability and running databases across the world. For this guide, we are talking specifically about mysql and postgres, and probably just running them in cloud providers like AWS/Google Cloud.

HELP ME IMPROVE THIS GUIDE :)

There are few things you could help me with:

provide feedback (is the topic relevant? did you find another way to approach it? or did you find some edge cases not covered)
give me opportunity to work on something 🙃
help me with writing (I don't have a structured way of writing 😞, and I want to improve on it)

principles

use singular noun
~~use uuid if possible~~
use ordered UUID v1 (only v1 can be ordered sequentially), stored as BINARY(16) when the data is dynamically generated. For reference table, stick to auto-incremented primary keys since the values are static and won't change that much either. If the number of items can be less than [a-z0-9], then use char(1) as primary key, since they can be more verbose than just int (m for male, f for female, o for others etc)
some issues with auto-incremented id is that it needs to be converted to the correct type (int64/uint64) in the application to prevent users from submitting string alphabets. The same complexity lies with UUID, which needs validation too. This is only necessary if we want to avoid the call to the db. Casting the type to int8 means only up to 127 ids are supported!
use soft delete
no null fields, except date (why? When using a strongly typed language as the database client, dealing with null (or nil pointer) is a pain. It is easier to go with sane default values. Also, when you start working with reporting tools, there's some additional logic or edge cases with database NULL that you need to handle)

Notes

inner joins is faster that subquery most of the time
apply logic in the database if you are going to have many different applications
use optimized uuid for faster querying
include the default auto incremented id
PostgreSQL automatically creates indexes on primary keys and unique constraints, but not on the referencing side of foreign key relationships.
put shared logic in template databases - they are like your common folders
(postgres) don't use serial, use generated as always identity for primary keys postgres if non-uuid keys are required
(postgres) use text, for postgres, text and varchar has no performance difference unlink mysql (which don't even has text)
for reference table, use the naming convention entity_type, e.g. notification_type, role_type, and use identity keys
you can use custom function as default keys, this is useful when require insert into a different table as foreign keys (e.g. party relationship)
the equivalent of api is schema views
use schema to split migrations and functionality, e.g. auth schema contains all auth related operations
use extra/other column naming for jsonb
use valid_through for tstzrange
use <entities>_count for counts
use id, created_at, updated_at, deleted_at
be careful when using auto incremented ids. Some application have slug for username, and if there is no checking on the username, and the user used an integer id as a name, then the query will always resolve wrongly
(postgres) adding new column to existing table will always end up in the last position. In Mysql, you can specify to place it before or after an column.

Search Path

SET search_path=onetsoc,public;
SHOW search_path;
SET search_path TO default;

Styleguides

Yes, there are styleguide (and perhaps linters) for everything.

https://www.sqlstyle.guide/

Migration file naming convention

Reference the naming convention from active record:

https://edgeguides.rubyonrails.org/active_record_migrations.html#using-the-change-method

Useful Statements

-- Sets a default created date
created_at datetime     NOT NULL DEFAULT CURRENT_TIMESTAMP

-- Sets a default updated date, and updates it whenever a row is updated
updated_at datetime     NOT NULL DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP

-- Adds a new constraint that checks if the first user id is smaller than the second.
CONSTRAINT check_one_way CHECK (user_id1 < user_id2)

-- Adds a constraint that checks if the combination of both columns is unique.
CONSTRAINT uq_user_id_1_user_id_2 UNIQUE (user_id1, user_id2)

-- Sets a foreign key constraint, and updates the foreign key when the primary key changes, or delete the row when it is deleted.
FOREIGN KEY (user_id1) REFERENCES user (id) ON UPDATE CASCADE ON DELETE CASCADE

-- Sets a foreign key constraint, and updates the foreign key when the primary key changes, or set the foreign key to null when it is deleted.
FOREIGN KEY (relationship) REFERENCES ref_relationship (status) ON UPDATE CASCADE ON DELETE SET NULL

-- Create a composite primary key from two columns.
PRIMARY KEY (user_id1, user_id2)

Sorting:

-- Sorting integer string in the correct order
SELECT * FROM <table> ORDER BY CAST(<column> AS unsigned)

Data Type: IPV4 and IPV6

You have two possibilities (for an IPv4 address) :

a varchar(15), if your want to store the IP address as a string. e.g. 192.128.0.15 for instance
an integer (4 bytes), if you convert the IP address to an integer. e.g. 3229614095 for the IP I used before

`ipv4` INT UNSIGNED
INSERT INTO `table` (`ipv4`) VALUES (INET_ATON("127.0.0.1"));
SELECT INET_NTOA(`ipv4`) FROM `table`;

`ipv6` VARBINARY(16)
INSERT INTO `table` (`ipv6`) VALUES (INET6_ATON("127.0.0.1"));
SELECT INET6_NTOA(`ipv6`) FROM `table`;

To use a single column for both IPV4 and IPV6:

CREATE TABLE `sensor` (
  `ip` varbinary(16) NOT NULL DEFAULT '0x'
)
-- Insert IPv6.
insert into sensor (ip) values (INET6_ATON("2001:0db8:85a3:0000:0000:8a2e:0370:7334"));

-- Insert IPv4.
insert into sensor (ip) values (INET6_ATON("255.255.255.0"));

select INET6_NTOA(ip) from sensor;
+------------------------------+
| INET6_NTOA(ip)               |
+------------------------------+
| 2001:db8:85a3::8a2e:370:7334 |
| 255.255.255.0                |
+------------------------------+
2 rows in set (0.00 sec)

Sample query from ipnation:

SELECT c.country 
FROM ip2nationCountries c, ip2nation i 
WHERE i.ip < INET_ATON('your_ip_address') 
AND c.code = i.country 
ORDER BY i.ip DESC 
LIMIT 0,1;

References:

http://www.ip2nation.com/

Data Type: Country

Al Jumahiriyah al Arabiyah al Libiyah ash Shabiyah al Ishtirakiyah al Uzma also known as Libya is the world's longest country name at 74 characters with spaces and 63 characters without.

country varchar(74) NOT NULL DEFAULT ''

Data Type: Address

address_line_1 VARCHAR(255) NOT NULL DEFAULT '',
address_line_2 VARCHAR(255) NOT NULL DEFAULT '',

-- Longest city name: Llanfairpwllgwyngyllgogerychwyrndrobwllllantysiliogogogoch (58 chars.)
city VARCHAR(58) NOT NULL DEFAULT '',

-- Longest state name: The State of Rhode Island and Providence Plantations (52 chars.)
state VARCHAR(56) NOT NULL DEFAULT '',
-- There is no maximum size for a postcode. Currently, the longest postal code is 10 char. Iran has 10 diguts and the US have 4 and 5 seperated by a hyphen. Brazil is 9 and Canada is 7, I think.
postal_code VARCHAR(16) NOT NULL DEFAULT '',
country VARCHAR(74) NOT NULL DEFAULT '',

Alternative is this, based on OpenID AddressClaim

street_address VARCHAR(255) NOT NULL DEFAULT '' COMMENT 'Full street address component, which MAY include house number, street name, Post Office Box, and multi-line extended street address information. This field MAY contain multiple lines, separated by newlines. Newlines can be represented either as a carriage return/line feed pair ("\r\n") or as a single line feed character ("\n").'
locality VARCHAR(58) NOT NULL DEFAULT '' COMMENT 'City or locality component.'
region VARCHAR(56) NOT NULL DEFAULT '' COMMENT 'State, province, prefecture or region component.'
postal_code VARCHAR(16) NOT NULL DEFAULT '' COMMENT 'Zip code or postal code component.'
country VARCHAR(74) NOT NULL DEFAULT '' COMMENT 'Country name component.'
-- latitude (See below)
-- longitude

References:

postal code length https://stackoverflow.com/questions/325041/i-need-to-store-postal-codes-in-a-database-how-big-should-the-column-be

Data Type: URL

url varchar(2083) NOT NULL DEFAULT '';

Checking can be done at the application side. If the length exceeded that of 2083, just warn the client or suggest them to use a url shortener.

References: https://stackoverflow.com/questions/219569/best-database-field-type-for-a-url

Data Type: Email

https://stackoverflow.com/questions/8242567/acceptable-field-type-and-size-for-email-address

email VARCHAR(255) NOT NULL UNIQUE

Data Type: Geolocation

With MySQL <8.0:

CREATE TABLE locations (
  lat DECIMAL(10,8) NOT NULL, 
  lng DECIMAL(11,8) NOT NULL
);

with MySQL >8.0:

CREATE TABLE locations (
    location POINT SRID 4326 NOT NULL,
    SPATIAL INDEX (location)
);

To insert:

INSERT INTO locations (location) VALUES (ST_PointFromText('Point(1 1)', 4326));

To select:

SELECT ST_AsText(location) FROM locations

References:

https://medium.com/maatwebsite/the-best-way-to-locate-in-mysql-8-e47a59892443

Data Type: TZ

Max length of 32, longest is America/Argentina/ComodRivadavia:

zoneinfo VARCHAR(32) COMMENT "String from zoneinfo [zoneinfo] time zone database representing the End-User's time zone. For example, Europe/Paris or America/Los_Angeles"

References:

Data Type: Locale

BCP47/RFC5646 section 4.4.1 recommends a 35 characters tag length:

locale VARCHAR(35) NOT NULL DEFAULT '' COMMENT "End-User's locale, represented as a BCP47 [RFC5646] language tag. This is typically an ISO 639-1 Alpha-2 [ISO639?1] language code in lowercase and an ISO 3166-1 Alpha-2 [ISO3166?1] country code in uppercase, separated by a dash. For example, en-US or fr-CA. As a compatibility note, some implementations have used an underscore as the separator rather than a dash, for example, en_US",

References:

Data Type: Phone number

Phone numbers are usually stored as E.164.

phone_number VARCHAR(32) NOT NULL DEFAULT '',
phone_number_verified BOOLEAN NOT NULL DEFAULT 0,

TL;DR, don't store phone number as bigint, as trailing zeros will break it. References:

Data Type: Name

Longest name (225 characters)

Barnaby Marmaduke Aloysius Benjy Cobweb Dartagnan Egbert Felix Gaspar Humbert Ignatius Jayden Kasper Leroy Maximilian Neddy Obiajulu Pepin Quilliam Rosencrantz Sexton Teddy Upwood Vivatma Wayland Xylon Yardley Zachary Usansky

References:

http://www.worldrecordacademy.com/society/longest_name_Barnaby_Marmaduke_sets_world_record_112063.html

Data Type: Gender

Column naming can be sex, or alternatively gender:

-- Probably the best bet, but needs to be validated. When in doubt, use this.
sex char(1) 
insert into table (gender) values (IF(LEFT(?,1) in ('m', 'f', 'x', 'o'), LOWER(LEFT(?,1)), ''));
-- We can also just take the first character of the string with the left function.
insert into table(gender) values (LEFT('female', 1));

-- With enum. Allows only 'm', 'f', 'M', or 'F'. Don't use enum - it will rebuild the whole database when we update it.
sex enum('m','f') DEFAULT 'm' 

-- With set.
sex set('m', 'f') // Allows '', 'm', 'M', 'f', 'F', or 'm,f'

References:

Note: We could have used check constraint, but it is ignored by MySQL.

Data Type: Currency

compliant with Generally Accepted Accounting Principles (GAAP):

currency DECIMAL(13,4)

For percentage:

-- For your case (0.00% to 100.00%) you'd want decimal(5,4).
gst DECIMAL(5, 4)

-- For the most common case (0% to 100%) you'd want decimal(3,2).
discount DECIMAL(3, 2)

Data Type: Stock Ticker

Tickers on the NYSE range from one to five characters long, with those of length five typically being used for mutual funds and ETFs (VFIAX is the symbol for Vanguard 500 index).

symbol VARCHAR(5)

Marital Status

Code	Description	Definition
1	Single	This refers to a person who has never been married.
2	Married	This refers to a person who is recognised as married under the marriage laws in Singapore. It includes a person who has remarried.
3	Widowed	This refers to a person whose spouse(s) is/are deceased and who has not remarried.
4	Separated	This refers to a person who has been legally separated or estranged from his/her spouse(s) and who has not remarried.
5	Divorced	This refers to a person whose marriage(s) has/have been legally dissolved and who has not remarried.
x	Not Reported	This includes instances where the marital status is unknown, not reported or where there is no/insufficient information available on the marital status

marital_status ENUM('single', 'married', 'widowed', 'separated', 'divorced', 'not reported');

References:

https://www.singstat.gov.sg/-/media/files/standards_and_classifications/scms.pdf

References

Gender X: https://www.lifesitenews.com/news/generation-x-germany-to-allow-third-blank-gender-for-birth-certificates

One-to-One Relationship

Example of 1-to-1 relationship between user and preference table:

CREATE TABLE IF NOT EXISTS user (
  name VARCHAR(255),
  id INT UNSIGNED AUTO_INCREMENT,
  PRIMARY KEY (id)
);

CREATE TABLE IF NOT EXISTS preference (
  user_id INT UNSIGNED AUTO_INCREMENT,
  interest TEXT,
  -- ...other fields
  PRIMARY KEY (id),
  FOREIGN KEY (id) REFERENCES user(id)
);

Country Table

https://www.ip2location.com/free/country-multilingual

Issues

There was a scenario where User A is logged in User B account (bug), the reason is very simple. User A previous user id is 2, and the JWT token is not expired. When the db was cleared, and User B recreated two new users, which has user id 1 and 2. So now, the JWT token only store the user id 2, when User A calls the API, since the previous id is 2, User A can view User B profile. Pitfalls of integer id.

Thoughts

Should I create a differen table for user profile and password? No. https://stackoverflow.com/questions/17683571/should-i-create-2-tables-first-for-usernames-and-passwords-and-other-for-user https://www.quora.com/Should-we-keep-the-user-name-and-password-in-the-same-table-where-the-other-personal-information-is https://dba.stackexchange.com/questions/148909/is-it-a-good-practice-to-isolate-login-information-username-password-in-a-sep

Useful features for postgres

lateral join
sum conditional
filter option
RANK, DENSE_RANK, ROW_NUMBER
partial index

select id, name, 
	review_count_rank,
	recent_review_count_rank,
	rating_rank,
	word_count_rank,
	review_count_rank + recent_review_count_rank + rating_rank + word_count_rank as total
from (
	select 
		product_items.id as id,
		product_items.name,
		
	--	product_items.cached_reviews_count,
		DENSE_RANK() OVER (order by product_items.cached_reviews_count desc) review_count_rank,
		
	--	COALESCE(tmp.review_count, 0) as recent_review_count,
		DENSE_RANK() OVER (order by COALESCE(tmp.review_count, 0) desc) recent_review_count_rank,
		
	--	product_items.cached_rating,
		DENSE_RANK() OVER (order by product_items.cached_rating desc) rating_rank,
		
	--	COALESCE(tmp.word_count, 0) as word_count,
		DENSE_RANK() OVER (order by COALESCE(tmp.word_count,0) desc) word_count_rank
	from product_items 
	left join (
		select 
			pir.item_id as item_id, 
			count(*) as review_count,
			sum(array_length(regexp_split_to_array(pir.text, '\s'),1)) as word_count
		from product_items pi
		left join product_item_reviews pir
			on (pi.id = pir.item_id)
		where pir.deleted_at is null
			and pi.deleted_at is null
			and pir.created_at >= current_timestamp - interval '30 day'
		group by pir.item_id
	) tmp on (tmp.item_id = product_items.id)
	where category_id = 6
		and product_items.deleted_at is null
	order by review_count_rank, 
		rating_rank,
		recent_review_count_rank,
		word_count_rank
	) tmp
order by total;

Null

Advantages of null fields (or when to use null):

we can use null field with unique values, so that empty strings will not be counted (they are considered unique)
It depends on the domain you are working on. NULL means absence of value (i.e. there is no value), while empty string means there is a string value of zero length.

Ways to sort array alphabetically in postgres.

This is one interesting problem that I faced when designing a friendship table - I need to create two rows with both the user id (user_id, friend_id) pair. However, querying becomes complex, as now I querying for the pair requires a union (and indices on both side). One way to solve it is to create another column that is the hash of both ids, sorted. The idea is to create a trigger that will sort both ids, hash them as md5, and store it in another column.

select (select array(select unnest (ARRAY[user_id, friend_id]) as x ORDER BY x)  as j) from relationship;
select md5(array_to_string(array_agg(id), '')) 
from (
	select * 
	from (values ('6769d922-ac68-11ea-8c70-9b8806d7aa41'), ('6769d922-ac68-11ea-8c70-9b8806d7aa41')) 
	as f(id) 
	order by f
) tmp;

Alternative way:

select 
	MD5(row(
		case 
			when user_id < friend_id 
			then (user_id, friend_id)
			else (friend_id, user_id)
		end
	)::text)
from (
	values 
	('7d7849d0-b94f-11ea-92be-43016fd48059', '8175c79c-b94f-11ea-92be-ab6d21fe7fb3'),
	('8175c79c-b94f-11ea-92be-ab6d21fe7fb3', '7d7849d0-b94f-11ea-92be-43016fd48059')
) as f(user_id, friend_id);

Finding missing index on foreign keys:

https://stackoverflow.com/questions/970562/postgres-and-indexes-on-foreign-keys-and-primary-keys

Using Identity Column (Postgres)

^ All postgres related topics should be tagged.

Identity column is the recommended approach over serial.

CREATE TABLE IF NOT EXISTS world (
-	id serial PRIMARY KEY,
+	id integer GENERATED ALWAYS AS IDENTITY PRIMARY KEY,
	name text
);

One advantage is we can't directly override the id:

INSERT INTO world (name) VALUES('will produce id 1');
INSERT INTO world (id, name) OVERRIDING SYSTEM VALUE VALUES(10, 'will produce id 10');
INSERT INTO world (name) VALUES('will produce id 2');

Using custom function as default key (Postgres)

We can actually use custom functions to generate the default key in Postgres. The example below shows an example of party and organization table.

we always have to create a party first before creating a person or organization, and the reference the id
this can be simplified by using a function

CREATE TABLE IF NOT EXISTS party(
	id uuid PRIMARY KEY DEFAULT gen_random_uuid(),
	type text not null check (type in ('person', 'organization'))
)

CREATE OR REPLACE FUNCTION gen_party_id(_type text) 
RETURNS uuid AS $$
	INSERT INTO party (type) VALUES (_type)
	RETURNING id;
$$ LANGUAGE SQL VOLATILE;

CREATE TABLE IF NOT EXISTS organization (
	id uuid PRIMARY KEY NOT NULL DEFAULT gen_party_id('organization'),
	type text NOT NULL DEFAULT 'organization' CHECK (type = 'organization'),
	name text,
	foreign key (id, type) references party(id, type)
);

volland / database-design Goto Github PK

database-design's Introduction

Database Design

HELP ME IMPROVE THIS GUIDE :)

principles

Notes

Search Path

Styleguides

Migration file naming convention

Useful Statements

Data Type: IPV4 and IPV6

Data Type: Country

Data Type: Address

Data Type: URL

Data Type: Email

Data Type: Geolocation

Data Type: TZ

Data Type: Locale

Data Type: Phone number

Data Type: Name

Data Type: Gender

Data Type: Currency

Data Type: Stock Ticker

Marital Status

References

One-to-One Relationship

Country Table

Issues

Thoughts

Useful features for postgres

Null

Ways to sort array alphabetically in postgres.

Finding missing index on foreign keys:

Using Identity Column (Postgres)

Using custom function as default key (Postgres)

database-design's People

Contributors

Watchers

Recommend Projects

Recommend Topics

Recommend Org