sqlbind-t

sqlbind-t allows to safely bind parameters in text based raw SQL queries using t-string templates.

>>> email = 'some@domain.com'
>>> query = t'SELECT * FROM users WHERE email = {email}'
>>> raw_sql, params = render(query)
>>> raw_sql
'SELECT * FROM users WHERE email = ?'
>>> params
['some@domain.com']
>>> connection.execute(raw_sql, params)  # your db connection instance
['results...']

Supports all DBAPI parameter styles. Isn't limited by DBAPI compatible drivers and could be used with anything accepting raw SQL query and parameters in some way. For example sqlbind-t could be used with SQLAlchemy textual queries. Or with clickhouse-driver's non-DBAPI interface.

Table of contents

• Installation
• Motivation
• Quick start
• Static queries
• Dynamic queries
• Trusted SQL fragments and identifiers
• Older python version support

Installation

pip install sqlbind-t

Motivation

ORMs are great and could be used effectively for a huge number of tasks. But after many years with SQLAlchemy I've noticed some repeating patterns:

• It's really not an easy task to decipher complex SQLAlchemy expression back into SQL. Especially when CTEs, sub-queries, nested queries or self-referential queries are involved. It composes quite well but it takes too much effort to write and read SQLAlchemy queries. For novices it could be a hard time to deal with it.

• Most of reporting queries are big enough already not to be bothered with ORMs and use raw SQL anyway. This kind of SQL often requires dynamic constructs and becomes string fiddling contraption.

• For a few tasks ORMs bring too much overhead and the only solution is to get down to raw DBAPI connection and raw SQL.

• (Minor personal grudge, please ignore it) For some ORMs (like Django ORM) your SQL intuition could be useless and requires deep ORM understanding.

It boils down to one thing: from time to time you have to write raw SQL queries. I could highlight 3 types of queries:

1. Fixed queries. They don't contain any parameters. For example SELECT id, name FROM users ORDER BY registered DESC LIMIT 10. In general fixed queries or fixed query parts compose well and don't require any special treatment. Python's f-strings are enough.

2. Static queries. They contain parameters but structure is fully known beforehand. For example SELECT id, name FROM users WHERE email = :email LIMIT 1. They are also could be composed without large issues, especially for connection drivers supporting named parameters (:param, %(param)s) and accepting dicts as parameters. Although for positional connection drivers (%s, ?) composition requires careful parameter tracking and queries could be fragile to change.

3. Dynamic queries. Query part presence could depend on parameter value or external condition. For example to provide result on input filter you have to add CTE and corresponding JOIN to a query. Or add filters only for non None input values. ORMs are effective for composing such queries. Using raw SQL are almost impossible for abstraction and leads to a complex boilerplate heavy code.

Note: here and in following sections I deliberately use simple examples. In real life there is no need to use sqlbind-t for such kind of queries.

Note: by composing I mean ability to assemble a final query from parts which could be abstracted and reused.

sqlbind-t tries to address issues with static and dynamic query types. It tracks parameter binds and could help with dynamic query parts.

Quick start

Some things to consider:

• sqlbind-t provides an API for a simple composition of raw SQL. On high level user operates with t-strings (Template objects) or thin wrappers around it. As a last step before execution user renders template into actual raw SQL and execution parameters.

• sqlbind-t doesn't parse SQL in t-strings. It only concatenates stuff and extracts interpolations as execution parameters.

• There is a large set of functions/methods to address dynamic queries but you haven't use it inline in a single query string. You could use variables to keep query parts and stitch resulted SQL from these parts.

• This README misses large portions of API. Feel free to explore doc strings with examples.

General use case looks like:

import sqlbind_t.dialect
from sqlbind_t import AnySQL
from sqlbind_t.query_params import QMarkQueryParams

# A global alias to a dialect used by connection backend.
# There is DB specific dialect (`sqlbind_t.sqlite.Dialect` for example)
dialect = sqlbind_t.dialect.Dialect()

def execute_query(query: AnySQL):
    # Render query template into raw SQL and corresponding parameters
    # using explicit query parameter style (qmark: `?`).
    raw_sql, params = dialect.render(query, QMarkQueryParams())
    with connection.cursor() as cursor:  # use your DBAPI connection
        return cursor.execute(raw_sql, params).fetchall()

def get_user(email: str):
    # Use t-string to capture query values
    query = t'SELECT * FROM users WHERE email = {email}'

    return execute_query(query)

As a shortcut you could use sqlbind_t.dialect.render function as a default dialect render.

Parameter marker styles

To render a query for a specific DBAPI marker style pass corresponding params container:

>>> import sqlbind_t.query_params as qp
>>> query = t'SELECT {10}, {20}'
>>> render(query, qp.QMarkQueryParams())
('SELECT ?, ?', [10, 20])
>>> render(query, qp.FormatQueryParams())
('SELECT %s, %s', [10, 20])
>>> render(query, qp.NumericQueryParams())
('SELECT :1, :2', [10, 20])
>>> render(query, qp.DollarQueryParams())
('SELECT $1, $2', [10, 20])
>>> render(query, qp.NamedQueryParams())
('SELECT :p0, :p1', {'p0': 10, 'p1': 20})
>>> render(query, qp.PyFormatQueryParams())
('SELECT %(p0)s, %(p1)s', {'p0': 10, 'p1': 20})

Static queries

Just use t-strings directly. Interpolation parts would be treated as parameters. The nice part it's quite hard to inject unprocessed data this way.

>>> date = "2023-01-01"
>>> render(t'SELECT * FROM users WHERE registered > {date}')
('SELECT * FROM users WHERE registered > ?', ['2023-01-01'])

Dynamic queries

Here begins a fun part. We can't use simple binds for dynamic queries. For example we have a function returning recently registered users:

def get_fresh_users(registered_since: datetime):
    query = t'''\
        SELECT * FROM users
        WHERE registered > {registered_since}
        ORDER BY registered
    '''
    return execute_query(query)

And later there is a new requirement for the function. It should return only enabled or only disabled users if corresponding argument is passed.

def get_fresh_users(registered_since: datetime, enabled: Optional[bool] = None):
    if enabled is not None:
        enabled_filter = t' AND enabled = {enabled}'
    else:
        enabled_filter = t''

    query = t'''\
        SELECT * FROM users
        WHERE registered > {registered_since} {enabled_filter}
        ORDER BY registered
    '''
    return execute_query(query)

It looks almost fine. You have to use t-string for enabled_filter as well otherwise query would treat it as a string parameter. From safety perspective it's great. But you can predict where we are going. Another one or two additional filters and it would be a complete mess. Take note how WHERE lost AND between two filters.

Conditional markers

sqlbind_t.cond injects a special UNDEFINED value on false condition and forces template to return an empty string on rendering. UNDEFINED values are processed per template instance, for example if some template A includes another nested template B and B contains UNDEFINED value then only B would be rendered as empty string.

>>> enabled = True
>>> render(t'AND enabled = {cond(enabled is not None)/enabled}')
('AND enabled = ?', [True])
>>> enabled = None
>>> render(t'AND enabled = {cond(enabled is not None)/enabled}')
('', [])

cond is a generic form. To remove a repetition (enabled is not None/enabled) when value is used both in a condition and as a parameter value there are helpers for most common cases:

• sqlbind_t.not_none: to check value is not None.
• sqlbind_t.truthy: to check value's trueness (bool(value) is True).
• sqlbind_t.required: forces parent template to be empty if sub-template is empty.

>>> enabled = True
>>> render(t'AND enabled = {not_none/enabled}')
('AND enabled = ?', [True])
>>> enabled = None
>>> render(t'AND enabled = {not_none/enabled}')
('', [])

required is useful when wrapper SQL should be omitted for empty nested templates:

>>> enabled = True
>>> nested = E.enabled == not_none/enabled
>>> render(t'WHERE {required/nested}')
('WHERE enabled = ?', [True])
>>> enabled = None
>>> nested = E.enabled == not_none/enabled
>>> render(t'WHERE {required/nested}')
('', [])

Let's try it in the function:

from sqlbind_t import not_none

def get_fresh_users(registered_since: datetime, enabled: Optional[bool] = None):
    enabled_filter = t'AND enabled = {not_none/enabled}'

    query = t'''\
        SELECT * FROM users
        WHERE registered > {registered_since} {enabled_filter}
        ORDER BY registered
    '''
    return execute_query(query)

Hmm. But really nothing was changed. You could write previous code with ternary if/else and it would look the same from semantic standpoint. Maybe use it inline?

from sqlbind_t import not_none

def get_fresh_users(registered_since: datetime, enabled: Optional[bool] = None):
    query = t'''\
        SELECT * FROM users
        WHERE registered > {registered_since}
              {t'AND enabled = {not_none/enabled}'}
        ORDER BY registered
    '''
    return execute_query(query)

It's somewhat palatable but uses nested t-strings. Let's find some improvements.

WHERE prepender

One approach is to extract filters outside of the query and use sqlbind_t.WHERE prepender to assemble WHERE expression from non-empty parts.

It could help with readability of long complex filters.

from sqlbind_t import not_none, WHERE

def get_fresh_users(registered_since: datetime, enabled: Optional[bool] = None):
    filters = [
        t'registered > {registered_since}',
        t'enabled = {not_none/enabled}',
    ]

    query = t'SELECT * FROM users {WHERE(*filters)} ORDER BY registered'
    return execute_query(query)

There are also other prependers: WITH, GROUP_BY, ORDER_BY, SET. They all omit empty parts or are rendered as empty string if all parts are empty.

Expressions

Expressions (sqlbind_t.E) allow to drop excessive quoting and generate templated results with infix operators.

Attribute access would render a qualified name:

>>> render(t'{E.field}')
('field', [])
>>> render(t'{E.table.field}')
('table.field', [])

Real DB tables/columns could use quite peculiar names. You could use E(name) to construct expression from any string:

>>> render(t'{E('"weird table"."weird column"')}')
('"weird table"."weird column"', [])

Expression objects define a set of infix operators allowing to bind a right value:

>>> render(t'{E.field > 10}')
('field > ?', [10])
>>> render(t'{E.table.field == 20}')
('table.field = ?', [20])
>>> render(t'{E('"my column"') != None}')
('"my column" IS NOT NULL', [])
>>> render(t'{E.field <= not_none/None}')  # conditional marks also work!
('', [])
>>> render(t'{E.field.IN(not_none/[10])}') # BTW sqlbind has workaround for SQLite to deal with arrays in IN
('field IN ?', [[10]])

It could look like a hack and feel ORM-ish but there is no any expression trees and tree compilation passes. Expressions are immediately emit templates with interpolations and simple to reason about.

Let's use expressions with the function:

from sqlbind_t import not_none, WHERE, E

def get_fresh_users(registered_since: datetime, enabled: Optional[bool] = None):
    filters = [
        E.registered > registered_since,
        E.enabled == not_none/enabled,
    ]

    query = t'SELECT * FROM users {WHERE(*filters)} ORDER BY registered'
    return execute_query(query)

Also you could use & operator to join filters to assemble expression without a list:

>>> filters = (E.registered > '2023-01-01') & (E.enabled == not_none/True)
>>> render(WHERE(filters))
('WHERE (registered > ? AND enabled = ?)', ['2023-01-01', True])

Technically all infix operators on expressions and majority of sqlbind-t helpers return sqlbind_t.SQL instances which define logic operators. Logic operators themselves return SQL instances as well.

>>> type(E.enabled == True)
<class 'sqlbind_t.SQL'>

SQL type is a thin wrapper around t-string template objects providing additional methods and operators. To obtain SQL instance there is sqlbind_t.sql (from a t-string) and sqlbind_t.text (from safe static text) helpers:

>>> sql(t'enabled = {True}')
SQL('enabled = ', Interpolation(True, 'True', None, ''))
>>> text('registered_since IS NOT NULL')
SQL('registered_since IS NOT NULL')

Note: Internally SQL instances are treated as free from UNDEFINED interpolations and sqlbind_t.sql ensures it. Please avoid calling SQL constructor directly.

All SQL instances could be composed with &, | and ~ (negate) operations:

>>> render(~E.enabled | text('registered_since IS NOT NULL'))
('(NOT enabled OR registered_since IS NOT NULL)', [])

Sadly due to python's' precedence rules you have to wrap expressions into additional parenthesis to make it work.

Inline filters

What about inlining filters into the query though? You could use AND_ (there is also 'OR_') prepender:

from sqlbind_t import not_none, E, AND_

def get_fresh_users(registered_since: datetime, enabled: Optional[bool] = None):
    query = t'''
        SELECT * FROM users
        WHERE registered > {registered_since}
         {AND_(E.enabled == not_none/enabled)}
        ORDER BY registered
    '''
    return execute_query(query)

It could be a viable alternative for extracted filters. It looks almost acceptable without nested t-strings and quoting.

That's it. I have no more tricks to make it more readable or pretty.

Inline vs extracted filters

Inline expressions scale really bad for a large set of filters especially with complex boolean logic or heavy on conditional markers. To make it manageable try to extract as much logic and use only not_none conditional marker.

IMHO instead of:

>>> now = None
>>> show_only_enabled = True
>>> query = t'SELECT * FROM users WHERE registered > {(now or datetime.now()) - timedelta(days=30)} {AND_(E.enabled == cond(show_only_enabled)/1)}'

please consider to use:

>>> now = None
>>> show_only_enabled = True
>>> registered_since = (now or datetime.now()) - timedelta(days=30)
>>> enabled = 1 if show_only_enabled else None
>>> query = t'SELECT * FROM users WHERE registered > {registered_since} {AND_(E.enabled == not_none/enabled)}'

Trusted SQL fragments and identifiers

By default, values inside {...} are treated as query parameters and are bound safely. But some API entry points deliberately embed SQL text directly. Use them only with trusted strings (not user input):

• text('...'): inserts a raw SQL fragment.
• E(name) / E.table.column: emits identifier-like SQL as-is.
• Keyword-based field names in WHERE(**kwargs), SET(**kwargs) and VALUES(**kwargs) are inserted into SQL directly.

If field names or SQL fragments are influenced by external input, validate them against an explicit allowlist before passing them into these APIs.

Older python version support

Python before 3.14 has no t-strings but sqlbind-t has alternative templatelib implementation and some workarounds to mimic t-string syntax to use with older Python.

tf-strings

Instead of using t-strings:

from sqlbind_t.dialect import render

user_id = 'some-id'
query = t'SELECT * FROM users where user_id = {user_id}'
raw_sql, params = render(query)

You could use marked f-strings (tf-strings):

from sqlbind_t import sqlf
from sqlbind_t.dialect import render

user_id = 'some-id'
query = sqlf(f'@SELECT * FROM users where user_id = {user_id}')
raw_sql, params = render(query)

It works by using custom import hook which replaces all f-strings with @ prefix in resulted AST. It has almost no overhead and semantically similar to t-strings. For example, for interpolation expressions, static type checkers see and validate the same Python code as with t-strings.

sqlf works as type wrapper for static type checkers and ensures argument is a transformed template.

It's required to install the AST transformer before any imports with tf-strings.

myapp/__init__.py:

import sqlbind_t.tfstring

# list modules which require transformations
sqlbind_t.tfstring.init(['myapp.db.queries'])

myapp/db/queries.py:

from sqlbind_t import sqlf

def get_user(email):
    query = sqlf(f'@SELECT * FROM users WHERE email = {email}')
    return execute_query(query)

You could use glob patterns as module names. * matches single part in qualified name, ** matches recursively.

Parse and eval

As a last-resort fallback there is sqlbind_t.sqls wrapper. Which parses template and evaluates interpolations using eval. Big downside is static type checkers can't check interpolation content.

from sqlbind_t import sqls

def get_user(email):
    query = sqls('SELECT * FROM users WHERE email = {email}')
    return execute_query(query)