A type-safe query builder and Active Record ORM for Crystal
SQLite · PostgreSQL · MySQL · CockroachDB

Documentation · Tutorials · How-to guides · Reference · API · Examples · Changelog

Tip

The full guide is at azutopia.gitbook.io/cql: tutorials, task guides, the reference and the design notes. Its source is in docs/, but it reads best on the site.

Why CQL

The v1 release candidate continues PR #65. See the release audit and validation record for the tested configurations and remaining release gates. The automated release workflow owns the version bump; this branch is not a published v1.0.0.

For the native engine, connect_failure_threshold=3&connect_retry_delay=1 enables a connection circuit breaker: failed connects open it, one probe tests recovery after the delay, and healthy existing connections remain usable. It never retries statements. Use sslmode=verify-full to verify the PostgreSQL server, and size the pool and query timeouts for the application. Pool configuration explains the defaults and failure behavior.

• Mistakes stop the build. Schemas, models and queries are Crystal code, so a misspelled column or a value of the wrong type is a compiler error instead of a runtime one.
• Values are always bound. Every value reaches the database as a parameter, and names that can't be bound, such as index and alias names, are validated.
• PostgreSQL needs no extra shard. CQL ships its own wire-protocol driver, with a fiber-safe pool, streaming, pipelining and COPY.

Take a users table with a text name and an integer age. Neither of these lines compiles:

User.where { nmae == "Ada" }  # misspelled column
User.where { age == "forty" } # a String compared with an Int32 column

Error: undefined local variable or method 'nmae' for User::Columns (with ... yield) and top-level (current scope)
Error: CQL: cannot compare a Int32 column with String using `==`; use a Int32 value or another Int32 column

Features

Area	What you get	Guide
Schema and migrations	Tables, indexes and foreign keys in CQL::Schema.define. Versioned CQL::Migration(N) classes that apply, roll back and redo. CQL::SchemaDump reads an existing database back into a schema.	Create a migration
Models	CQL::ActiveRecord::Model(Int64), or Int32, UUID and ULID keys, with validations, callbacks, scopes, timestamps, soft deletes, optimistic locking and opt-in dirty tracking.	Define a model
Associations	belongs_to, has_one, has_many and many_to_many, with preload for the first three. A foreign key whose type doesn't match the target's key fails to compile.	Has many
Typed queries	where { (age >= 18) & (status == "active") }, checked at compile time. Joins, CTEs, window functions, set operations, upsert, row locking, JSON and geospatial predicates, raw fragments with bound values, explain, and to_sql(for: adapter).	Typed where DSL
Bulk work	insert_all, upsert_all and insert_<association>_all send multi-row statements. find_each, find_in_batches and pluck keep large reads small.	Optimize queries
Transactions	Nested savepoints on every adapter.	Transactions
Guardrails	Validated identifiers, and CQL_VALIDATE_SCHEMA_MAPPINGS=1 crystal spec to check model attribute types against the schema.	Safety guardrails
Caching	Memory and Redis stores, a fragment cache with tag invalidation, and a per-request query cache with middleware for Kemal, Azu, Lucky, Amber, SpiderGazelle and HTTP::Server.	Query cache
Observability	CQL::Telemetry.subscribe sees every statement. CQL::Performance adds a profiler, N+1 detection and reports. Both stay off until you enable them or run in development.	Telemetry
PostgreSQL driver	Protocol 3.2 with fallback to 3.0, TLS, an AsyncEngine pool, streaming, pipelining, COPY and LISTEN/NOTIFY.	Native driver

Installation

Add CQL and your database's driver to shard.yml, then run shards install:

dependencies:
  cql:
    github: azutoolkit/cql
    version: ~> 0.0.861
  sqlite3: # for SQLite
    github: crystal-lang/crystal-sqlite3
  mysql: # for MySQL; used by the optional cql/mysql integration
    github: crystal-lang/crystal-mysql
    version: ~> 0.16.0

Note

PostgreSQL and CockroachDB need nothing more, because CQL ships the driver. CQL needs Crystal 1.19.1 or later; CI tests Crystal 1.19.1 and 1.22.1.

For MySQL 8.4, use require "cql/mysql" to enable caching_sha2_password and verified TLS. Set ssl-ca=/path/to/ca.pem in the database URL for a private CA; see MySQL 8.4 setup guide.

Quick Start

Save this as src/app.cr:

require "cql"
require "sqlite3"

AppDB = CQL::Schema.define(:app_db, adapter: CQL::Adapter::SQLite, uri: "sqlite3://./app.db") do
  table :users do
    primary :id, Int64, auto_increment: true
    text :name
  end
  table :posts do
    primary :id, Int64, auto_increment: true
    text :title
    bigint :user_id
  end
end

class User
  include CQL::ActiveRecord::Model(Int64)
  db_context AppDB, :users
  property name : String
  has_many :posts, Post, foreign_key: :user_id
  validate :name, presence: true

  def initialize(@name : String)
  end
end

class Post
  include CQL::ActiveRecord::Model(Int64)
  db_context AppDB, :posts
  property title : String
  property user_id : Int64?
  belongs_to :user, User, foreign_key: :user_id

  def initialize(@title : String, @user_id : Int64? = nil)
  end
end

AppDB.users.create! # CREATE TABLE IF NOT EXISTS users (id INTEGER PRIMARY KEY AUTOINCREMENT, name TEXT NOT NULL)
AppDB.posts.create!

ada = User.create!(name: "Ada")                           # INSERT INTO users (name) VALUES (?)
ada.posts.create(title: "Notes on the Analytical Engine") # INSERT INTO posts (title, user_id) VALUES (?, ?)

# Two queries: SELECT ... FROM posts, then SELECT ... FROM users WHERE users.id IN (?)
Post.preload(:user).all.each { |post| puts "#{post.title} by #{post.user.try(&.name)}" }

puts User.where { name == "Ada" }.count # SELECT COUNT(*) FROM users WHERE users.name = ?
puts User.new("").valid?                # the presence validation fails on a blank name

Run it:

crystal run src/app.cr

Notes on the Analytical Engine by Ada
1
false

Next, Your First CQL App adds migrations, updates, deletes and transactions.

How It Works

flowchart LR
    models["Schema and models<br/>CQL::Schema.define"] --> build["Build<br/>QueryBuilder(T) → CQL::Query"]
    build --> render["Render<br/>AST → Expression::Generator<br/>with the adapter's dialect"]
    render --> execute["Execute<br/>SQL and bound values on the pool<br/>or the open transaction"]
    execute --> db[("SQLite · PostgreSQL<br/>MySQL · CockroachDB")]
    db --> hydrate["Hydrate<br/>DB::Serializable models"]

Loading

Each query method records one clause. At execution, Query#build turns the clauses into an AST, and the Expression::Generator visitor writes SQL in the adapter's dialect, with placeholders (? or $1) where the values go. How CQL Keeps Queries Fast walks through each stage and what it costs.

Database Support

Database	Adapter	URL	Driver	How CQL checks it
SQLite	CQL::Adapter::SQLite	sqlite3://	sqlite3 shard	the spec suite runs live
PostgreSQL	CQL::Adapter::Postgres	postgres://, postgresql://	built in	generated-SQL specs; driver live specs on PostgreSQL 11 to 18
MySQL	CQL::Adapter::MySql	mysql://	mysql shard + require "cql/mysql"	live MySQL 8.0/8.4 transaction and authentication contracts
CockroachDB	CQL::Adapter::CockroachDB	cockroachdb://	built in	generated-SQL specs; driver live specs on CockroachDB 24.3

Feature differences, such as upsert, RETURNING, DISTINCT ON, locking, JSON and the minimum versions for CTEs and window functions, are listed in Database Support.

Native PostgreSQL Driver

require "cql" registers CQL's driver with crystal-db for postgres:// and postgresql:// URLs, so CQL::Schema and DB.open use it without the pg shard. The native driver guide covers running alongside pg and opting in to cockroachdb:// for DB.open.

CQL::Postgres::AsyncEngine has its own fiber-safe pool and is a CQL::Executor, so any statement runs on it with using:

DATABASE_URL = "postgres://app:secret@db.internal/app?sslmode=verify-full&pool_size=10"
engine = CQL::Postgres::AsyncEngine.connect(DATABASE_URL)

User.query.where(name: "Ada").using(engine).to_a                                         # ... WHERE users.name = $1
AppDB.query.from(:users).select(:name).where("id > ?", 10).using(engine).all(as: String) # ... WHERE id > $1

engine.transaction(isolation: :serializable) do |tx|
  AppDB.insert.into(:users).values(name: "Ada").using(tx).commit
  tx.savepoint { |inner| AppDB.insert.into(:posts).values(title: "Draft", user_id: 1_i64).using(inner).commit }
end

# 500 rows per fetch through a portal, not the whole result in memory
engine.stream(AppDB.query.from(:users).select(:id), fetch_size: 500) do |rows|
  rows.each { puts rows.read(Int64) }
end

# Several statements, one round trip
results = engine.pipeline([AppDB.query.from(:users).select(:name), AppDB.query.from(:posts).select(:title)])
results[0].as(CQL::Postgres::PipelineResult).all(as: String)

# A whole schema (models, migrations, transactions) on the engine's pool
ShopDB = CQL::Schema.new(:shop, DATABASE_URL, CQL::Adapter::Postgres, source: CQL::Postgres::EngineSource.connect(DATABASE_URL))

Guide	Covers
Native PostgreSQL Driver	entry points, switching from the pg shard
AsyncEngine and pool options	pool_size, health checks, setup:/reset:, errors
A Schema on the engine's pool	EngineSource
Streaming and pipelining	stream, pipeline, timeouts
COPY and binary parameters	copy_in, copy_out, binary_params=true
TLS, poolers and timeouts	sslmode, PgBouncer and other poolers, cancel, isolation
Use CockroachDB	CockroachDB differences

Measured Compatibility

The driver passes its live specs on PostgreSQL 11 to 18 and CockroachDB 24.3, and has been run behind PgBouncer, Odyssey, pgcat and pgpool-II.¹ CI also runs spec/postgres/live_spec.cr against a postgres service container on every push.

PostgreSQL versions, CockroachDB, poolers and TLS

PostgreSQL	Protocol used	Cancel key	SCRAM, MD5, cleartext	live_spec.cr
11.22, 12.22, 13.23, 14.24, 15.19, 16.15, 17.11	3.0 (server answers NegotiateProtocolVersion)	4 bytes	pass	pass
18.6	3.2	32 bytes	pass	pass

• Versions: script/pg-matrix runs spec/postgres/matrix_spec.cr and live_spec.cr (23 examples) against each server. Every version also passed re-preparing after a schema change, pipelining, COPY in and out, cancel after a statement timeout, and per-transaction synchronous_commit: false. PostgreSQL 19rc1, built from source, passed the matrix (7 of 7), live_spec.cr and streaming_spec.cr (35 examples). Untested: a final PostgreSQL 19, non-Linux or non-x86_64 hosts, distribution-patched builds.
• CockroachDB 24.3.5, insecure and with TLS plus a password: live_spec.cr and streaming_spec.cr pass except the LISTEN/NOTIFY example (CockroachDB has no LISTEN).
• Poolers: PgBouncer 1.22, 1.24 and 1.26 (script/pgbouncer-matrix), Odyssey 1.5.2, pgcat 1.2.0 and pgpool-II 4.3.7 (script/pooler_probes.cr); see Poolers for what works on each.
• TLS refusal: script/tls-probes checks that a TLS requirement that cannot be met raises CQL::Postgres::TLSRefused and that crystal-db's pool does not retry it. spec/postgres/blackhole_spec.cr drops a TLS connection on a path that stops answering: 50 timed-out requests, 50 pings and 10 writes finished without hanging. Other TLS modes and Unix sockets were tried by hand against PostgreSQL 16; beyond that they have byte-level fixture and fake-server specs only.

Benchmarks: streaming and binary parameters

Two measurements on PostgreSQL 16 over loopback.²

• benchmarks/pg_wire_bench.cr, 200,000 rows of two columns: streaming 1000 rows per fetch held 1000 rows at once (+1.7 MB peak memory) and took 125 ms because of its 200 round trips; reading the whole result held 200,000 rows (+45 MB) and took 82 ms.
• benchmarks/pg_binary_params_bench.cr, statements per second with binary_params off / on (medians of 7 rounds of 5000 statements, encoding included):

Workload	Text	Binary
six scalars	15,751	18,358
one int8 / timestamptz / numeric	19,958 / 18,431 / 19,176	19,509 / 18,717 / 19,385
seven-column insert	12,902	13,654
int8[1000] / text[200] / numeric[200]	1,944 / 7,224 / 5,113	3,009 / 8,339 / 5,877
timestamptz[500]	904	4,929

A single value is within run-to-run noise, so binary_params stays off by default. To run both scripts on your own server, see benchmarks/README.md.

Performance

Model reads and writes on in-memory SQLite, from benchmarks/sqlite_bench.cr:³

Operation	Time	Allocated
ArUser.find(id)	1.13 µs	480 B
ArUser.last	3.31 µs	3.25 kB
ArUser.query.exists?(age: 30)	2.78 µs	3.1 kB
create!	3.57 µs	2.44 kB
update!	3.32 µs	2.05 kB
query.from(:users).where(id: 42).to_sql	1.27 µs	1.89 kB
join with order and limit, to_sql	2.28 µs	3.73 kB
100 parents, preload(:posts)	222 µs	165 kB

Bytes per operation repeat from run to run; time per operation depends on the machine. How CQL Keeps Queries Fast explains what each query costs and what is off by default.

Documentation

The guide is published at azutopia.gitbook.io/cql and organized by what you need:

Section	For	Start here
Tutorials	learning CQL step by step, from a first app to a blog	docs/tutorials/
How-to guides	one task each: models, queries, migrations, caching, performance, PostgreSQL	docs/how-to/
Reference	every option, DSL method and error, and a one-page cheat sheet	docs/reference/
Explanation	why CQL works the way it does: Active Record, guardrails, performance	docs/explanation/
Resources	FAQ and contributing	docs/resources/

Every public type and method is in the API reference, generated from the doc comments in src/ for each release. There are also installation notes, runnable examples and a playground.

Project Status

Important

CQL is pre-1.0 (0.0.x). CI tags and releases every merge to master, and a release can change behavior. Read the changelog before you upgrade.

Contributing

Bug reports, questions and pull requests are welcome. Open an issue to report a problem or discuss a change. To work on CQL itself:

shards install --ignore-crystal-version
crystal spec                                                    # SQLite runs live; PostgreSQL specs need DATABASE_URL
crystal tool format --check src spec
crystal run lib/ameba/bin/ameba.cr -- -f progress ./src ./spec  # lint, as CI runs it

Note

Without a server, the PostgreSQL and Redis specs are pending.

A bug fix comes with a spec that fails without it. A behavior change adds a CHANGELOG.md entry under [Unreleased] and updates the matching page in docs/. Contributing has the full workflow.

License

CQL is released under the MIT License.

Footnotes

1. Live specs and probe scripts run against real servers on one Linux x86_64 machine. Details are in the collapsed sections below. ↩

2. One machine, one run each. ↩

3. One run of crystal run --release benchmarks/sqlite_bench.cr: one fiber, on a shared 4-core Linux VM. ↩