Clear is an ORM built specifically for PostgreSQL in Crystal.
It's probably the most advanced ORM for PG on Crystal in term of features offered. It features Active Record pattern models, and low-level SQL builder.
You can deal out of the box with jsonb, tsvectors, cursors, CTE, bcrypt password,
array, uuid primary key, foreign constraints... and other things !
It also has a powerful DSL to construct where
and having
clauses.
The philosophy beneath is to please me (and you !) with emphasis made on business code readability and minimal setup.
The project is quite active and well maintened, too !
In few seconds, you want to use Clear if:
- You want an expressive ORM. Put straight your thought to your code !
- You'd like to use advanced Postgres features without hassle
- You are at aware of the pro and cons of Active Records pattern
You don't want to use Clear if:
- You're not willing to use on PostgreSQL
- You look for a minimalist ORM / Data Mapper
- You need something which doesn't evolve, with breaking changes. Clear is still in alpha but starting to mature !
- Active Record pattern based ORM
- Expressiveness as mantra - even with advanced features like jsonb, regexp... -
# Like ...
Product.query.where { ( type == "Book" ) & ( metadata.jsonb("author.full_name") == "Philip K. Dick" ) }
# ^--- will use @> operator, to relay on your gin index. For real.
Product.query.where { ( products.type == "Book" ) & ( products.metadata.jsonb("author.full_name") != "Philip K. Dick" ) }
# ^--- this time will use -> notation, because no optimizations possible :/
# Or...
User.query.where { created_at.in? 5.days.ago .. 1.day.ago }
# Or even...
ORM.query.where { ( description =~ /(^| )awesome($| )/i ) }.first!.name # Clear! :-)
- Proper debug information
- Log and colorize query. Show you the last query if your code crash !
- If failing on compile for a good reason, give proper explaination (or at least try)
- Migration system
- Validation system
- N+1 query avoidance strategy
- Transaction, rollback & savepoint
- Access to CTE, locks, cursors, scope, pagination, join, window, multi-connection and many others features
- Model lifecycle/hooks
- JSONB, UUID, FullTextSearch
In shards.yml
dependencies:
clear:
github: crystal-garage/clear
branch: develop
Then:
require "clear"
Clear offers some mixins, just include them in your classes to clear them:
class User
include Clear::Model
column id : Int64, primary: true
column email : String
column first_name : String?
column last_name : String?
column encrypted_password : Crypto::Bcrypt::Password
def password=(x)
self.encrypted_password = Crypto::Bcrypt::Password.create(password)
end
end
Number
,String
,Time
,Boolean
andJsonb
structures are already mapped.Array
of primitives too. For other type of data, just create your own converter !
class Clear::Model::Converter::MyClassConversion
def self.to_column(x) : MyClass?
case x
when String
MyClass.from_string(x)
when Slice(UInt8)
MyClass.from_slice(x)
else
raise "Cannot convert from #{x.class} to MyClass"
end
end
def self.to_db(x : UUID?)
x.to_s
end
end
Clear::Model::Converter.add_converter("MyClass", Clear::Model::Converter::MyClassConversion)
Most of the ORM for Crystal are mapping column type as Type | Nil
union.
It makes sens so we allow selection of some columns only of a model.
However, this have a caveats: columns are still accessible, and will return nil,
even if the real value of the column is not null !
Moreover, most of the developers will enforce nullity only on their programming language level via validation, but not on the database, leading to inconsistency.
Therefore, we choose to throw exception whenever a column is accessed before it has been initialized and to enforce presence through the union system of Crystal.
Clear offers this through the use of column wrapper.
Wrapper can be of the type of the column as in postgres, or in UNKNOWN
state.
This approach offers more flexibility:
User.query.select("last_name").each do |usr|
puts usr.first_name #Will raise an exception, as first_name hasn't been fetched.
end
u = User.new
u.first_name_column.defined? #Return false
u.first_name_column.value("") # Call the value or empty string if not defined :-)
u.first_name = "bonjour"
u.first_name_column.defined? #Return true now !
Wrapper give also some pretty useful features:
u = User.new
u.email = "[email protected]"
u.email_column.changed? # TRUE
u.email_column.revert
u.email_column.defined? # No more
Clear offers has_many
, has_one
, belongs_to
and has_many through
associations:
class Security::Action
belongs_to role : Role
end
class Security::Role
has_many user : User
end
class User
include Clear::Model
has_one user_info : UserInfo
has_many posts : Post
belongs_to role : Security::Role
# Use of the standard keys (users_id <=> security_role_id)
has_many actions : Security::Action, through: Security::Role
end
Clear offers a collection system for your models. The collection system
takes origin to the lower API Clear::SQL
, used to build requests.
To fetch one model:
# 1. Get the first user
User.query.first #Get the first user, ordered by primary key
# Get a specific user
User.find!(1) #Get the first user, or throw exception if not found.
# Usage of query provides a `find_by` kind of method:
u : User? = User.query.find { email =~ /yacine/i }
To prepare a collection, juste use Model#query
.
Collections include SQL::Select
object, so all the low level API
(where
, join
, group_by
, lock
...) can be used in this context.
# Get multiple users
User.query.where { (id >= 100) & (id <= 200) }.each do |user|
# Do something with user !
end
#In case you know there's millions of row, use a cursor to avoid memory issues !
User.query.where { (id >= 1) & (id <= 20_000_000) }.each_cursor(batch: 100) do |user|
# Do something with user; only 100 users will be stored in memory
# This method is using pg cursor, so it's 100% transaction-safe
end
Call aggregate functions from the query is possible. For complex aggregation,
I would recommend to use the SQL::View
API (note: Not yet developed),
and keep the model query for fetching models only
# count
user_on_gmail = User.query.where { email.ilike "@gmail.com%" }.count #Note: count return is Int64
# min/max
max_id = User.query.where { email.ilike "@gmail.com%" }.max("id", Int32)
# your own aggregate
weighted_avg = User.query.agg("SUM(performance_weight * performance_score) / SUM(performance_weight)", Float64)
Associations are basically getter which create predefined SQL. To access to an association, just call it !
User.query.each do |user|
puts "User #{user.id} posts:"
user.posts.each do |post| #Works, but will trigger a request for each user.
puts "• #{post.id}"
end
end
For every association, you can tell Clear to encache the results to avoid
N+1 queries, using with_XXX
on the collection:
# Will call two requests only.
User.query.with_posts.each do |user|
puts "User #{user.id} posts:"
user.posts.each do |post|
puts "• #{post.id}"
end
end
Note than Clear doesn't perform a join method, and the SQL produced will use
the operator IN
on the association.
In the case above:
- The first request will be
SELECT * FROM users;
- Thanks to the cache, a second request will be called before fetching the users:
SELECT * FROM posts WHERE user_id IN ( SELECT id FROM users )
I have plan in a late future to offer different query strategies for the cache (e.g. joins, unions...)
When you use the caching system of the association, using filters on association will invalidate the cache, and N+1 query will happens.
For example:
User.query.with_posts.each do |user|
puts "User #{user.id} published posts:"
# Here: The cache system will not work. The cache on association
# is invalidated by the filter `where`.
user.posts.where({published: true}).each do |post|
puts "• #{post.id}"
end
end
The way to fix it is to filter on the association itself:
User.query.with_posts(&.where({published: true})).each do |user|
puts "User #{user.id} published posts:"
# The posts collection of user is already encached with the published filter
user.posts.each do |post|
puts "• #{post.id}"
end
end
Note than, of course in this example user.posts
are not ALL the posts but only the
published
posts
Thanks to this system, we can stack it to encache long distance relations:
# Will cache users<=>posts & posts<=>category
# Total: 3 requests !
User.query.with_posts(&.with_category).each do |user|
#...
end
In case you want columns computed by postgres, or stored in another table, you can use fetch_column
.
By default, for performance reasons, fetch_columns
option is set to false.
users = User.query.select(email: "users.email",
remark: "infos.remark").join("infos"){ infos.user_id == users.id }.to_a(fetch_columns: true)
# Now the column "remark" will be fetched into each user object.
# Access can be made using `[]` operator on the model.
users.each do |u|
puts "email: `#{u.email}`, remark: `#{u["remark"]?}`"
end
inspect
over model offers debugging insights:
p # => #<Post:0x10c5f6720
@attributes={},
@cache=
#<Clear::Model::QueryCache:0x10c6e8100
@cache={},
@cache_activation=Set{}>,
@content_column=
"...",
@errors=[],
@id_column=38,
@persisted=true,
@published_column=true,
@read_only=false,
@title_column="Lorem ipsum torquent inceptos"*,
@user_id_column=5>
In this case, the *
means a column is changed and the object is dirty and diverge from the database.
One thing very important for a good ORM is to offer vision of the SQL called under the hood. Clear is offering SQL logging tools, with SQL syntax colorizing in your terminal.
For activation, simply setup the logger to DEBUG
level !
Log.builder.bind "clear.*", :debug, Log::IOBackend.new(STDOUT)
Object can be persisted, saved, updated:
u = User.new
u.email = "[email protected]"
u.save! #Save or throw if unsavable (validation failed).
Columns can be checked & reverted:
u = User.new
u.email = "[email protected]"
u.email_column.changed? # < Return "true"
u.email_column.revert # Return to #undef.
Presence validator is done using the type of the column:
class User
include Clear::Model
column first_name : String # Must be present
column last_name : String? # Can be null
end
There's a case when a column CAN be null inside Crystal, if not persisted, but CANNOT be null inside Postgres.
It's for example the case of the id
column, which take value after saving !
In this case, you can write:
class User
column id : Int64, primary: true, presence: false #id will be set using pg serial !
end
Thus, in all case this will fail:
u = User.new
u.id # raise error
When you save your model, Clear will call first the presence validators, then
call your custom made validators. All you have to do is to reimplement
the validate
method:
class MyModel
#...
def validate
# Your code goes here
end
end
Validation fails if model#errors
is not empty:
class MyModel
#...
def validate
if first_name_column.defined? && first_name != "ABCD" #< See below why `defined?` must be called.
add_error("first_name", "must be ABCD!")
end
end
end
Please use unique
feature of postgres. Unique validator at crystal level is a
non-go and lead to terrible race concurrency issues if your deploy on multiple nodes/pods.
It's an anti-pattern and must be avoided at any cost.
In the case you try validation on a column which has not been initialized, Clear will complain, telling you you cannot access to the column. Let's see an example here:
class MyModel
#...
def validate
add_error("first_name", "should not be empty") if first_name == ""
end
end
MyModel.new.save! #< Raise unexpected exception, not validation failure :(
This validator will raise an exception, because first_name has never been initialized. To avoid this, we have many way:
# 1. Check presence:
def validate
if first_name_column.defined? #Ensure we have a value here.
add_error("first_name", "should not be empty") if first_name == ""
end
end
# 2. Use column object + default value
def validate
add_error("first_name", "should not be empty") if first_name_column.value("") == ""
end
# 3. Use the helper macro `on_presence`
def validate
on_presence(first_name) do
add_error("first_name", "should not be empty") if first_name == ""
end
end
#4. Use the helper macro `ensure_than`
def validate
ensure_than(first_name, "should not be empty", &.!=(""))
end
#5. Use the `ensure_than` helper (but with block notation) !
def validate
ensure_than(first_name, "should not be empty") do |column|
column != ""
end
end
I recommend the 4th method in most of the cases you will faces. Simple to write and easy to read !
Clear offers of course a migration system.
Migration should have an order
column set.
This number can be wrote at the end of the class itself:
class Migration1
include Clear::Migration
def change(dir)
#...
end
end
Another way is to write down all your migrations one file per migration, and
naming the file using the [number]_migration_description.cr
pattern.
In this case, the migration class name doesn't need to have a number at the end of the class name.
# in src/db/migrations/1234_create_table.cr
class CreateTable
include Clear::Migration
def change(dir)
#...
end
end
Migration must implement the method change(dir : Migration::Direction)
Direction is the current direction of the migration (up or down).
It provides few methods: up?
, down?
, up(&block)
, down(&block)
You can create a table:
def change(dir)
create_table(:test) do |t|
t.column :first_name, :string, index: true
t.column :last_name, :string, unique: true
t.index "lower(first_name || ' ' || last_name)", using: :btree
t.timestamps
end
end
I strongly encourage to use the foreign key constraints of postgres for your references:
t.references to: "users", on_delete: "cascade", null: false
There's no plan to offer on Crystal level the on_delete
feature, like
dependent
in ActiveRecord. That's a standard PG feature, just set it
up in migration
Models add a layer of computation. Below is a sample with a very simple model (two integer column), with fetching of 100k rows over 1M rows database, using --release flag:
Method | Total time | Speed | |
---|---|---|---|
Simple load 100k | 12.04 | ( 83.03ms) (± 3.87%) | 2.28× slower |
With cursor | 8.26 | ( 121.0ms) (± 1.25%) | 3.32× slower |
With attributes | 10.30 | ( 97.12ms) (± 4.07%) | 2.67× slower |
With attributes and cursor | 7.55 | (132.52ms) (± 2.39%) | 3.64× slower |
SQL only | 27.46 | ( 36.42ms) (± 5.05%) | fastest |
Simple load 100k
is using an array to fetch the 100k rows.With cursor
is querying 1000 rows at a timeWith attribute
setup a hash to deal with unknown attributes in the model (e.g. aggregates)With attribute and cursor
is doing cursored fetch with hash attributes createdSQL only
build and execute SQL using SQL::Builder
As you can see, it takes around 100ms to fetch 100k rows for this simple model (SQL included). If for more complex model, it would take a bit more of time, I think the performances are quite reasonable, and tenfold or plus faster than Rails's ActiveRecord.
This shard is provided under the MIT license.
All contributions are welcome ! As a specialized ORM for PostgreSQL, be sure a great contribution on a very specific PG feature will be incorporated to this shard. I hope one day we will cover all the features of PG here !
In order to run the test suite, you will need to have the PostgresSQL service locally available via a socket for access with psql. psql will attempt to use the 'postgres' user to create the test database. If you are working with a newly installed database that may not have the postgres user, this can be created with createuser -s postgres
.