𓆤DBay

Table of Contents generated with DocToc

• 𓆤DBay
  • Introduction
  • Documentation
    • Main
      • Using Defaults
      • Automatic Location
      • Randomly Chosen Filename
      • Using Parameters
    • Opening and Closing DBs
      • Opening / Attaching DBs
      • Closing / Detaching DBs
    • Transactions and Context Handlers
    • Query
      • Use the alternative Connection to Avoid Connection Busy Errors
      • SQL Tag Function for Better Embedded Syntax
      • Executing SQL
    • User-Defined Functions (UDFs)
    • Standard Library of SQL Functions (StdLib)
      • List of Functions
      • Use Case for DBay Exceptions and Assertions: Enforcing Invariants
      • Use Case for DBay Variables: Parametrized Views
    • Safe Escaping for SQL Values and Identifiers
      • Purpose
      • Escaping Identifiers, General Values, and List Values
      • Statement Interpolation
    • SQL Statement Generation
      • Insert Statement Generation
      • Insert Statements with a returning Clause
    • Random
  • Concurrent Writes
  • Macros for SQL
  • Notes on User Defined Functions (UDFs)
    • (Outline for a) Draft for a Stored Procedure Feature Request
  • Note on Package Structure
    • better-sqlite3 an 'Unsaved' Dependency
  • To Do
  • Is Done

𓆤DBay

DBay is built on better-sqlite3, which is a NodeJS adapter for SQLite. It provides convenient access to in-process, on-file and in-memory relational databases.

DBay is the successor to and a re-write of ICQL-DBA. It is under development and nearing feature-parity with its predecessor while already providing some significant improvements in terms of ease of use and simplicity of implementation.

Introduction

DBay provides

• In-Process,
• In-Memory & File-Based
• Relational Data Processing
• for NodeJS
• with SQLite;
• being based on better-sqlite3,
• it works (almost) exclusively in a synchronous fashion.

Documentation

• Benchmarks

Main

Using Defaults

In order to construct (instantiate) a DBay object, you can call the constructor without any arguments:

{ DBay }  = require 'dbay'
db        = new DBay()

The db object will then have two properties db.sqlt1 and db.sqlt2 that are better-sqlite3 connections to the same temporary DB in the 'automatic location'.

The db object will then have a (non-enumerable) property db.sqlt1 which is a better-sqlite3 connection to a temporary DB in the 'automatic location'.

Automatic Location

The so-called 'automatic location' is either

• the directory /dev/shm on Linux systems that support SHared Memory (a.k.a a RAM disk)
• the OS's temporary directory as announced by os.tmpdir()

In either case, a file with a random name will be created in that location.

Randomly Chosen Filename

Format dbay-NNNNNNNNNN.sqlite, where N is a digit [0-9].

Using Parameters

You can also call the constructor with a configuration object that may have one or more of the following fields:

• cfg.path (?non-empty text): Specifies which file system path to save the DB to; if the path given is relative, it will be resolved in reference to the current directory (process.cwd()). When not specified, cfg.path will be derived from DBay.C.autolocation and a randomly chosen filename.

• cfg.temporary (?boolean): Specifies whether DB file is to be removed when process exits or db.destry() is called explicitly. cfg.temporary defaults to false if cfg.path is given, and true otherwise (when a random filename is chosen).

Opening and Closing DBs

Opening / Attaching DBs

• db.open cfg: Attach a new or existing DB to the db's connections (db.sqlt1, db.sqlt1). (db.sqlt1).

• cfg:

  • schema (non-empty string): Required property that specifies the name under which the newly attached DB's objects can be accessed as; having attached a DB as, say, db.open { schema: 'foo', path: 'path/to/my.db', }, one can then run queries like db "select * from foo.main;" against it. Observe that
    • the DB opened at object creation time (db = new DBay()) always has the implicit name main, and schema temp is reserved for temporary databases.
  • path (string): FS path to existing or to-be-created DB file; for compatibility, this may also be set to one of the special values that indicates a in-memory DB, although that is not recommended.
  • temporary (boolean): Defaults to false when a path is given, and to true otherwise.

• The custom SQLite library that is compiled when installing DBay has its SQLITE_LIMIT_ATTACHED compilation parameter set to the maximum allowed value of 125 (instead of the default 10). This allows developers to assemble a DB application from dozens of smaller pieces when desired.

Closing / Detaching DBs

▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊

Transactions and Context Handlers

▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊

Query

▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊

Use the alternative Connection to Avoid Connection Busy Errors

SQLite imposes certain restrictions on what one can and cannot do in a concurrent fashion, one restriction being that within the same connection, one cannot at the same time iterate over query results and insert values—not even into a completely unrelated table. There are ways to get around that limitation. For one, the standard recommendation of the maker of better-sqlite3 is to just fetch all the needed rows from the DB and then iterate over the list of values. This is totally doable and the simplest and most transparent solution, but of course a nagging thought remains—what if the dataset gets really huge? In reality, this may turn out never to be a problem, realistically, but that consideration won't make that nagging thought vanish.

There's a (seemingly) better way. Commit 57e062a: make alt an on-demand clone of present instance introduces the new (non-enumerable) property db.alt which represents a clone of the db object. Previous versions had two underlying DB connections sqlt1 and sqlt2 which could be used for the purposes described in this section, but their drawback was that one falls back to the underlying better-sqlite3 API which can be a little confusing.

Let's have a look at a toy DB and see how to use db.alt. This is the definition, two tables with one integer field each:

#.................................................................................
db SQL"""
  create table foo ( n integer );
  create table bar ( n integer );"""
for n in [ 10 .. 12 ]
  db SQL"insert into foo ( n ) values ( $n );", { n, }

And here's what we want to accomplish: read data from one table and, based on that data, insert records into another one. Naïvely one would perhaps write it this way (the transaction being added because we need it later anyway):

db.with_transaction =>
  for row from db SQL"select * from foo order by n;"
    info '^806-2^', row
    db SQL"insert into bar values ( $n );", { n: n ** 2, }
    return null

This will not run, however, but fail with TypeError: This database connection is busy executing a query. This is where db.alt comes in: we have to use one connection for the iteration and another one for the insertion; this works:

#.................................................................................
# (1)
db.with_transaction =>
  for { n, } from db.alt SQL"select * from foo order by n;"
    #             ^^^^^^
    db SQL"insert into bar values ( $n ) returning *;", { n: n ** 2, }
  return null

The following points should be kept in mind: Explicit transactions and explicit prepared statements are they key factors for speedy inserts. Since explicit transactions are crucial for concurrent inserts, it's recommended to do all inserts within explicit transactions.

Therefore, because it's good practice to use explicit transactions and explicit prepared statements when doing inserts, most of the time inserts should take on the form shown in snippets (2) or (4):

#.................................................................................
# (2)
insert_into_bar = db.prepare SQL"insert into bar values ( $n ) returning *;"
db.with_transaction =>
  for { n, } from db.alt SQL"select * from foo order by n;"
    insert_into_bar.run { n: n ** 2, }
  return null

Observe that we have to explicitly exhaust the iterator that is returned from insert ... returning statements; to do so, either use db.first_row() or call a prepared statement's .get() (instead of .run()) method:

#.................................................................................
# (3)
db.with_transaction =>
  for { n, } from db.alt SQL"select * from foo order by n;"
    new_row = db.first_row SQL"insert into bar values ( $n ) returning *;", { n: n ** 2, }
  return null

#.................................................................................
# (4)
insert_into_bar = db.prepare SQL"insert into bar values ( $n ) returning *;"
db.with_transaction =>
  for { n, } from db.alt SQL"select * from foo order by n;"
    new_row = insert_into_bar.get { n: n ** 2, }
  return null

SQL Tag Function for Better Embedded Syntax

Mixing SQL and application code has the drawback that instead of editing SQL in your SQL-aware text editor, now you are editing bland string literals in your SQL-aware editor. If there only was a way to tell the editor that some strings contain SQL and should be treated as such!—Well, now there is. The combined power of [JavaScript Tagged Templates] (https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Template_literals#tagged_template_literals) and an (exprimental proof-of-concept level) [set of Sublime Text syntax definitions called coffeeplus] (https://github.com/loveencounterflow/coffeeplus) makes it possible to embed SQL into JavaScript (and CoffeeScript) source code. The way this works is by providing a 'tag function' that can be prepended to string literals. The name of the function together with the ensuing quotes can be recognized by the editor's hiliter so that constructs like SQL"...", SQL"""...""" and so will trigger switching languages. The tag function does next to nothing; here is its definition:

class DBay
  @SQL: ( parts, expressions... ) ->
    R = parts[ 0 ]
    for expression, idx in expressions
      R += expression.toString() + parts[ idx + 1 ]
    return R

It can be used like this:

{ DBay } = require 'dbay'
{ SQL  } = DBay

db = new DBay { path: 'path/to/db.sqlite', }

for row from db SQL"select id, name, price from products order by 1;"
              #    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
              #    imagine proper embedded hiliting etc here
  console.log row.id, row.name, row.price

Be aware that coffeeplus is more of an MVP than a polished package. As such, not even reckognizing backticks has been implemented yet so is probably best used with CoffeeScript.

Executing SQL

One thing that sets DBay apart from other database adapters is the fact that the object returned from new DBay() is both the representative of the database opened and a callable function. This makes executing statements and running queries very concise. This is an excerpt from the DBay test suite:

{ DBay }            = require H.dbay_path
db                  = new DBay()
db ->
  db SQL"drop table if exists texts;"
  db SQL"create table texts ( nr integer not null primary key, text text );"
  db SQL"insert into texts values ( 3, 'third' );"
  db SQL"insert into texts values ( 1, 'first' );"
  db SQL"insert into texts values ( ?, ? );", [ 2, 'second', ]
  #.......................................................................................................
  T?.throws /cannot start a transaction within a transaction/, ->
    db ->
#.........................................................................................................
T?.throws /UNIQUE constraint failed: texts\.nr/, ->
  db ->
    db SQL"insert into texts values ( 3, 'third' );"
#.........................................................................................................
rows = db SQL"select * from texts order by nr;"
rows = [ rows..., ]
T?.eq rows, [ { nr: 1, text: 'first' }, { nr: 2, text: 'second' }, { nr: 3, text: 'third' } ]

Note In the above SQL has been set to String.raw and has no further effect on the string it precedes; it is just used as a syntax marker (cool because then you can have nested syntax hiliting).

As shown by benchmarks, a crucial factor for getting maximum performance out of using SQLite is strategically placed transactions. SQLite will not ever execute a DB query outside of a transaction; when no transaction has been explicitly opened with begin transaction, the DB engine will precede each query implicitly with (the equivalent of) begin transaction and follow it with either commit or rollback. This means when a thousand insert statements are run, a thousand transactions will be started and committed, leavin performance pretty much in the dust.

To avoid that performance hit, users are advised to always start and commit transactions when doing many consecutive queries. DBay's callable db object makes that easy: just write db -> many; inserts; here; (JS: db( () -> { many; inserts; here; })), i.e. pass a function as the sole argument to db, and DBay will wrap that function with a transaction. In case an error should occur, DBay guarantees to call rollback (in a try ... finally ... clause). Those who like to make things more explicit can also use db.with_transaction ->. Both formats allow to pass in a configuration object with an attribute mode that may be set to one of 'deferred', 'immediate', or 'exclusive', the default being 'deferred'.

Another slight performance hit may be caused by the logic DBay uses to (look up an SQL text in a cache or) prepare a statement and then decide whether to call better-sqlite3's' Database::execute(), Statement::run() or Statement::iterate(); in order to circumvent that extra work, users may choose to fall back on to better-sqlite3 explicitly:

insert = db.prepare SQL"insert into texts values ( ?, ? );" # returns a `better-sqlite3` `Statement` instance
db ->
  insert.run [ 2, 'second', ]

User-Defined Functions (UDFs)

▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊

Standard Library of SQL Functions (StdLib)

List of Functions

• Strings

  • std_str_reverse()
  • std_str_join()
  • std_str_split()
  • std_str_split_re()
  • std_str_split_first()
  • std_re_matches()

• XXX

  • std_generate_series()

• Output

  • std_echo()
  • std_debug()
  • std_info()
  • std_warn()

• Exceptions and Assertions

  • std_raise( message )—unconditionally throw an error with message given.
  • std_raise_json( facets_json )—unconditionally throw an error with informational properties encoded as a JSON string.
  • std_assert( test, message )—throw an error with message if test is falsy.
  • std_warn_if( test, message )—print an error message if test is truthy.
  • std_warn_unless()—print an error message if test is falsy.

• Variables

  • std_getv()
  • std_variables()

• Dates, Time, Durations, Timestamps

• dt_dbayts_from_isots

• dt_parse

• dt_format

• dt_isots_from_dbayts

Use Case for DBay Exceptions and Assertions: Enforcing Invariants

• std_assert: ( test, message ) -> throws error if test is false(y)
• std_warn_unless: ( test, message ) -> prints warning if test is false(y)
• often one wants to ensure a given SQL statement returns / affects exactly zero or one rows
• easy to do if some rows are affected, but more difficult when no rows are affected, because a function in the statement won't be called when there are no rows.
• The trick is to ensure that at least one row is computed even when no rows match the query, and the way to do that is to include an aggregate function such as count(*).
• May want to include limit 1 where appropriate.

select
    *,
    std_assert(
      count(*) > 0,
      '^2734-1^ expected one or more rows, got ' || count(*) ) as _message
  from nnt
  where true
    and ( n != 0 );

select
    *,
    std_assert(
      count(*) > 0, -- using `count(*)` will cause the function to be called
                    -- even in case there are no matching rows
      '^2734-2^ expected one or more rows, got ' || count(*) ) as _message
  from nnt
  where true
    and ( n != 0 )
    and ( t = 'nonexistant' ); -- this condition is never fulfilled

Use Case for DBay Variables: Parametrized Views

• An alternative for user-defined table functions where those functions would perform queries against the DB, which is tricky.
• Inside the view definition, use std_getv( name ) to retrieve variable values which must have been set immediately prior to accessing the view.
• Downside is that it's easy to forget to update a given value, so best done from inside a specialized method in your application.

Safe Escaping for SQL Values and Identifiers

Purpose

• Facilitate the creation of securely escaped SQL literals.
• In general not thought of as a replacement for the value interpolation offered by DBay::prepare(), DBay::query() and so, except when
  • one wants to parametrize DB object names (e.g. use table or column names like variables),
  • one wants to interpolate an SQL values list, as in select employee from employees where department in ( 'sales', 'HR' );.

Escaping Identifiers, General Values, and List Values

• db.sql.I: ( name ) ->: returns a properly quoted and escaped SQL Identifier.
• db.sql.L: ( x ) ->: returns a properly quoted and escaped SQL Value. Note that booleans (true, false) will be converted to 1 and 0, respectively.
• db.sql.V: ( x ) ->: returns a bracketed SQL list of values (using db.sql.V() for each list element).

Statement Interpolation

db.interpolate( sql, values ) -> accepts a template (a string with placeholder formulas) and a list or object of values. It returns a string with the placeholder formulas replaced with the escaped values.

# using named placeholders
sql     = SQL"select $:col_a, $:col_b where $:col_b in $V:choices"
d       = { col_a: 'foo', col_b: 'bar', choices: [ 1, 2, 3, ], }
result  = db.sql.interpolate sql, d
# > """select "foo", "bar" where "bar" in ( 1, 2, 3 )"""

# using positional placeholders
sql     = SQL"select ?:, ?: where ?: in ?V:"
d       = [ 'foo', 'bar', 'bar', [ 1, 2, 3, ], ]
result  = db.sql.interpolate sql, d
# > """select "foo", "bar" where "bar" in ( 1, 2, 3 )"""

# using an unknown format
sql     = SQL"select ?:, ?X: where ?: in ?V:"
d       = [ 'foo', 'bar', 'bar', [ 1, 2, 3, ], ]
result  = db.sql.interpolate sql, d
# throws "unknown interpolation format 'X'"

SQL Statement Generation

DBay offers limited support for the declarative generation of a small number of recurring classes of SQL statements. These facilities are in no way intended to constitute or grow into a full-blown Object-Relational Mapper (ORM); instead, they are meant to make working with relational data less of a repetitive chore.

Insert Statement Generation

To pick one case in point, SQL insert statements when called from a procedural language have a nasty habit of demanding not two, but three copies of a table's column names:

db SQL"""
  create table xy (
    a   integer not null primary key,
    b   text not null,
    c   boolean not null );"""
db SQL"insert into xy ( b, c ) values ( $b, $c )", { b, c, }
#                     ^^^^^^^^        ^^^^^^^^^^   ^^^^^^^^^

Instead, we implement facilities to cover the most frequent use cases and offer opportunities to insert SQL fragments at strategic points.

Often, when an insert statement is being called for, one wants to insert full rows (minus generated columns, for which see below) into tables. This is the default that DBay makes easy: A call to db.prepare_insert() with the insertion target identified with into will return a prepared statement that can then be used as first argument to the db callable:

insert_into_xy = db.prepare_insert { into: 'xy', }
db insert_into_xy, { a, b, c, }

Observe that named parameters (as opposed to positional ones) are used, so values must be passed as an object (as opposed to a list).

In case the actual SQL text of the statement is needed, call db.create_insert() instead:

insert_sql = db.create_insert { into: 'xy', }
# 'insert into "main"."xy" ( "a", "b", "c" ) values ( $a, $b, $c );'

When one or more columns in a table are autoincremented or have a default value, then those columns are often intended not to be set explicitly. What's more, columns with generated values must not be set explicitly. For this reason, db.create_insert() (and, by extension, db.prepare_insert()) will skip generated columns and allow to explicitly specify either included columns (as fields) or else excluded columns (as exclude):

db SQL"""
  create table t1(
    a integer primary key,
    b integer,
    c text,
    d integer generated always as (a*abs(b)) virtual,
    e text generated always as (substr(c,b,b+1)) stored );"""
insert_into_t1 = db.create_insert { into: 't1', }

### Observe `d` and `e` are left out because they're generated, but `a` is present: ###
# 'insert into "main"."t1" ( "a", "b", "c" ) values ( $a, $b, $c );'

### You probably want either this: ###
insert_into_t1 = db.create_insert { into: 't1', fields: [ 'b', 'c', ], }
# 'insert into "main"."t1" ( "b", "c" ) values ( $b, $c );'

### Or this: ###
insert_into_t1 = db.create_insert { into: 't1', exclude: [ 'a', ], }
# 'insert into "main"."t1" ( "b", "c" ) values ( $b, $c );'

There's a subtle yet important semantic difference in how the fields and exclude settings are handled: When fields are explicitly given, the table does not have to exist when generating the SQL; however, when fields is not given, the table must already exist at the time of calling create_insert().

In either case, prepare_insert() can only succeed when all referenced object in an SQL statement have already been created.

The next important thing one often wants in inserts is resolving conflicts. DBay create_insert() supports setting on_conflict to either (1) an arbitrary string that should spell out a syntactically valid SQL on conflict clause, or (2) an object { update: true, } to generate SQL that updates the explicitly or implicitly selected columns. This form has been chosen to leave the door open to future expansions of supported features.

When choosing the first option, observe that whatever string is passed in, create_insert() will prepend 'on conflict ' to it; therefore, to create an insert statement that ignores insert conflicts, and according to the upsert syntax railroad diagram: —

— the right thing to do is to call db.create_insert { into: table_name, on_conflict: 'do nothing', }. Assuming table t1 has been declared as above, calling

db.create_insert { into: 't1', exclude: [ 'a', ], on_conflict: "do nothing", }

will generate the (unformatted but properly escaped) equivalent to:

insert into main.t1 ( b, c )
  values ( $b, $c )
  on conflict do nothing;
  --          |<------>|
  --        inserted string

while calling

db.create_insert { into: 't1', exclude: [ 'a', ], on_conflict: { update: true, }, }

wiil generate the (unformatted but properly escaped) equivalent to:

insert into main.t1 ( b, c )
  values ( $b, $c )
  on conflict do update set  --| conflict resolution clause
    b = excluded.b,          --| mandated by { update: true, }
    c = excluded.c;          --| containing same fields as above

Insert Statements with a returning Clause

It is sometimes handy to have insert statements that return a useful value. Here's a toy example that demonstrates how one can have a table with generated columns:

db SQL"""
  create table xy (
    a   integer not null primary key,
    b   text not null,
    c   text generated always as ( '+' || b || '+' ) );"""
insert_into_xy_sql = db.create_insert { into: 'xy', on_conflict: SQL"do nothing", returning: '*', }
# -> 'insert into "main"."xy" ( "a", "b" ) values ( $a, $b ) on conflict do nothing returning *;'
db.single_row insert_into_xy_sql, { a: 1, b: 'any', } # -> { a: 1, b: 'any', c: '+any+' }
db.single_row insert_into_xy_sql, { a: 2, b: 'duh', } # -> { a: 2, b: 'duh', c: '+duh+' }
db.single_row insert_into_xy_sql, { a: 3, b: 'foo', } # -> { a: 3, b: 'foo', c: '+foo+' }

Generally, the returning clause must be defined by a non-empty string that is valid SQL for the position after returning and the end of the statement. A star * will return the entire row that has been inserted; we here use db.single_row() to eschew the result iterator that would be returned by default.

Random

▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊ ▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊▌▊

Concurrent Writes

• with_deferred_write: ( f ) ->—will call the function as f write, where write may be used like the db object. Each call to write() will put any arguments into a cache; typically, this will be calls of the form write my_prepared_statement, my_data. When f() has finished, the items in cache will be used to call the db object as in db whatever... for whatever in buffer. These repeated calls will happen inside an implicit transaction in case no transaction is already open.

• db.alt—the db.alt property, when first retrieved, opens a second better-sqlite3 connection to the same database file. Using pragma journal_mode = wal (which is the default as of DBay v14.16), this (or any other secondary) connection can be used for concurrent writes.

• observe that insert statements generated with db.prepare_insert() are now implicitly bound to db.alt to improve DBay's concurrency story.

• Observe that rows inserted in the same transaction are not visible to the alternative connection until those rows have been committed. Generally, one will want to insert data in one transaction, finish it, and then begin a new (explicit or implicit) transaction in order to iterate over existing and possibly inserting new data. It is obvious that one does not want this newly inserted data to be visible just yet because that could cause an infinite loop (just like appending to an array while stepping over its elements would create an infinite loop).

Macros for SQL

Because User-Defined Functions have several shortcomings in SQLite (as discussed under Notes on User Defined Functions (UDFs), below), an alternative mechanism named dbay-sql-macros has been conceived to work around those issues. It has been integrated into DBay and is accessible via two methods, declare() and resolve(), under the macros key of a db instance. Furthermore, when constructing the DBay instance, one can pass in { macros: true, } to get implicit macro resolution. An example:

{ DBay }          = require 'dbay'
{ SQL  }          = DBay
db                = new DBay { macros: true, }
#.........................................................................................................
### NOTE exact syntax subject to change; for now, this works: ###
db.macros.declare SQL"""@secret_power( @a, @b ) = power( @a, @b ) / @b;"""
#.........................................................................................................
result  = db.all_rows SQL"""select @secret_power( 3, 2 ) as p;"""
# [ { p: 4.5 } ]

For more details, head over to the documentation for dbay-sql-macros.

Notes on User Defined Functions (UDFs)

• SQLite principally differs from client/server RDBMSes in that it allows User Defined Functions (UDFs) only on the DB connection of the host application

• these UDFs are written in the language and run in the environment of the host application

• I believe UDFs are, as such, a huge boon because they help to push many chores to the DB and allow for such nifty things as having generated columns whose contents are not stored but indexed and that can call into existing code of the hosting app without code duplication and without network roundtrips (both of which are hallmarks of client/server architectures)

• but the downside of connection-defined UDFs is that SQLite DBs created with UDFs will break when the environment changes (e.g. when openening a second connection to the same DB without recreating all UDFs or openening the DB with other tools such as the sqlite3 command line tool); an abortive error will occur as soon as any statement (such as selecting from a view whose definition includes a call to a UDF) is encountered. While one may be able to perform some operations that do not cause a function to be called, it is not a hallmark of a safe operations regime when things work sometimes without any warning only to break under certain conditions. I'm not aware of a way to conveniently check an SQLite DB for the use or lack of use of UDFs so apparently the best thing one can do is proofread the DDL statements and/or do selects from each relation.

• this means that a DB created with UDFs will not be amenable for any of the helpful tools that exist (like ER diagrammers and so on)

• one can also not pass a DB file around for other people to have a gander into the data—using UDFs means your host application (or all the relevant parts of it) has to be reproduced on the other machine, and even then, only the host application can provide access to the data—again, no external tooling here

• because UDFs are so useful, it's probably worthwhile to think about how to work around the limitations. Possible solutions include:

  • put all the logic in the application—this is the most straightforward and classical approach. If a prospective generated field can not be readily or reasonably computed using only SQLite's built in functions, use an ordinary table field and precompute the value before inserting rows. If the UDF would be called from a view, turn that view into a table and insert the rows from you application. Rating: +1 because it's so straightforward and classical.
  • open a feature request against SQLite with a view to enable support for something like SQL CREATE FUNCTION; the body of such a function could be formulated in much the same way as the already existing syntax for CREATE TRIGGER which likewise allows to define a block with a sequence of SQL statements. In its simplest form, CREATE FUNCTION would allow for parametrized views, which would be incredibly useful. Such a feature request has been issued before (not listed on the Open Feature Requests page) in 2021 and lead to an extended and informative discussion (see triggers, below), but so far, nothing has come of it. Rating: +0 because while everyone is encouraged to do it, hopes are not high IMHO; however see below for a draft that I think could have some chances.
  • compile your UDFs into a loadable SQLite extension—this can solve part of the problem, but only just so. Most tools simply have no way or concept to load an SQLite extension, one exception being the sqlite3 command line tool, but even then, you must ship the extension alongside with the DB file, and the receiving partner will have to do a bit more work to open the DB (and be willing to use the command line). They must also be on a compatible system or your *.so/*.dll will not work, or else you must compile the extension for multiple systems. Rating: -1 because who wants to do the authoring and testing and compiling stuff when so little universal usability comes out of it.
  • use triggers—this is a somewhat promising workaround that only occurred to me when reading the OP of the aforementioned feature request: put your functionality inside a trigger. Of course, this makes only sense if your function is readily expressable in terms of SQLite's built in functions, but potentially you can better bundle your functionalities. Rating +1 because this is another classical technique; its main downside is that you still have to (re-)produce your functionalities in pure SQL (with built in functions), and in case the same functionality is required in more than one place, there is no other way than to do copy/paste. Maybe code duplication could be avoided with code generation?

All of these solutions suck in one or the other way.

(Outline for a) Draft for a Stored Procedure Feature Request

• minimal: the extension to SQL consists in introducing a CREATE FUNCTION statement; its body would be much like the existing syntax for triggers. The default (and, initially, only) language to be used is SQL.

Example:

CREATE FUNCTION product( a number, b number ) RETURNS number LANGUAGE SQL
  BEGIN
    SELECT a * b; /* or RETURN a * b; */
  END;
END;

Notes:

• Type annotations and the RETURNS clause should be optional as in CREATE TABLE statements, whereas LANGUAGE SQL should initially be made mandatory to avoid premature fixation of a bad default.
• Initially at least, functions should not be multi-dispatch (i.e. a name can only appear at most once).
• Nice-to-have: CREATE OR REPLACE, IF NOT EXISTS, DROP FUNCTION.

This form is already useful because now you can bundle and name recurrent expressions—anything that can appear in a scalar (single-values) SELECT statement can be named and collected into libraries.

Extension Add to this RETURNS SETOF $TYPE, RETURNS SETOF ROW, and now you can have table-valued functions a.k.a. parametrized views!

Extension Add to this statement sequences.

Extension Add to this branching (IF/THEN and/or CASE/WHEN). This extension would cross the line where language inside a function declaration is significantly different from that outside. OTOH branching could conceivably work outside of functions, much like C's preprocessor directives.

Extension Add to this LOOP/BREAK/BREAK IF loops.

Extension Add to this YIELD for use in table-valued functions.

Extension Add to this (function-local) variables. These are in principle already implemented in the form of function parameters.

Extension Some support for dynamic SQL that could potentially be much less clunky than what PostgreSQL offers; at first one would only need a syntax to signify safe interpolation as identifier, e.g. SELECT * FROM @table_name; or similar.

Note on Package Structure

better-sqlite3 an 'Unsaved' Dependency

Since DBay depends on better-sqlite3 with a custom-configured build of the SQLite C engine, it is (for whatever reason) important that better-sqlite3 must not be listed under package.json#dependencies; otherwise, compilation will not work properly. The build script will run npm install better-sqlite3@'^7.4.3' but with an added --no-save flag.

## Use npm, Not pnpm

Also, at the time of this writing (2021-09), while the project compiles fine using npm v7.21.1 (on NodeJS v16.9.1 on Linux Mint), but it fails using pnpm v6.14.6 with Unknown options: 'build-from-source', 'sqlite3'. Yarn has not been tried.

Note—These considerations only concern those who wish to fork/clone DBay to work on the code. Those who just want to use DBay as a dependency of their project can both either run npm install dbay or pnpm add dbay, both package managers work fine.

To Do

• [–] port foundational code from hengist &c

• [–] at construction time, allow dbnick when path is given and ram is false

• [–] to solve the table-UDF-with-DB-access conundrum, consider

  • [+] https://github.com/mapnik/mapnik/issues/797, where connection parameters are discussed (see also https://www.sqlite.org/c3ref/open.html); nothing of interested AFAICS
  • [–] mirroring a given DB into a second (RAM or file) location, taking care to replay any goings-on on both instances. This is probably unattractive from a performance POV.
  • [–] using NodeJS worker threads to perform updates; maybe one could even continuously mirror a RAM DB on disk to get a near-synchronous copy, obliviating the necessity to explicitly call db.save(). See https://github.com/JoshuaWise/better-sqlite3/blob/master/docs/threads.md
  • [–] Observe that, seemingly, only table-valued UDFs hang while with shared-cache we already can issue selects from inside UDFs, so maybe there's a teeny, fixable difference between how both are implemented that leads to the undesirable behavior

• [–] let users choose between SQLite-only RAM DBs and tmpfs-based in-memory DBs (b/c the latter allow pragma journal_mode = WAL for better concurrent access). Cons include: tmpfs-based RAM DBs necessitate mounting a RAM disk which needs sudo rights, so might as well just instruct users to mount RAM disk, then use that path? Still, it would be preferrable to have some automatic copy-to-durable in place.

• [–] implement context handler for discardable / temporary file

• [–] allow to call DBay::do -> ... with an asynchronous function

• [–] implement DBay::open(), DBay::close()

• [–] ensure how cross-schema foreign keys work when re-attaching DBs / schemas one by one

• [–] demote random from a mixin to functions in helpers.

• [–] implement db.truncate() / db.delete(); allow to retrieve SQL.

• [–] implement DBay::insert_into.<table> [ 'field1', 'field2', ..., ], { field1, field2, ..., }; allow to retrieve SQL.

• [–] clarify whether UDFs get called at all when any argument is null b/c it looks like they don't get called which would be unfortunate

• [–] add schematic to clarify terms like database, schema, connection; hilite that UDFs are defined on connections (not schemas or databases as would be the case in e.g. PostgreSQL).

• [–] allow to transparently treat key/value tables as caches

• [–] implement escaping of dollar-prefixed SQL placeholders (needed by create_insert()).

• [–] implement

  • [–] db.commit()
  • [–] db.rollback()

• [–] allow to use sets with sql.V()

• [–] implement export/snapshot function that generates a DB with a simplified structure:

  • replace generated fields, results from function calls by constants
  • remove strict and similar newer attributes
  • DB should be readable by tools like sqlite3 command line, visualize-sqlite

• [+] consider to implement trash() as trash_to_sql() (path optional), trash_to_sqlite() (path optional) trash functionality now moved to DeSQL

• [–] rewrite all uses of plain E to @E

• [–] limit support for schemas, especially in plugins; require a separate instance of DBay for each DB file (so that all DB objects are in the default main namespace and the SQL"#{schema}.xxx" constructs can become SQL"xxx"). Complex DBs can still be assembled with db.open(), but one must keep in mind that in SQLite, foreign keys do not work across schemas, only joins so, so that limits the usefulness of multi-schema connections.

• [–] consider to change call argument in UDFs to callee

• [–] add fields to std_re_matches():

• [–] consider to change construction method of DBay instances to returning a proxy over a function (as done in guy.obj.Strict_proprietor.get())

    db.create_table_function
      name:           prefix + '_re_matches'
      columns:        [ 'match', 'capture', 'start', 'stop', ]
      parameters:     [ 'text', 'pattern', ]
      rows: ( text, pattern ) ->
        regex = new RegExp pattern, 'g'
        while ( match = regex.exec text )?
          [ m, c, ] = match
          yield [ m, ( c ? null ), start: match.index, stop: match.index + m.length, ]
        return null

• [–] update to an SQLite version that includes #9430ead7ba433cbf to fix an issue with window functions

• [–] write a chapter about application architecture best practices, including:

  • limitations of using UDFs (tools, sqlite3 CLI will not work)
  • SQLite, like other popular DBs (i.e. Postgres) are notoriously bad at giving exact error locations. Using triggers (and generated columns) can exacerbate that problem (imagine an on insert trigger that performs inserts on another table using select from t; ahould an error occur in the last step, SQLite will still attribute it to the original insert statement (without giving it any kind of locality or mentioning t's role))
  • because errors are badly located by SQLite, prefer writing many small steps instead of few big ones (i.e. prefer db SQL"do this;", db SQL"do that;" over db SQL"do this; do that;")

• [–] implement select * from t SQL generation

• [–] could the SQL string annotation / tagged literal function be syntactically extended to allow simpler interpolation of escaped names? Could we instantiate it with a dictionary of values (implement in Guy)

• [–] would it be possible to keep the application code in its own tables? one could then ship the application by sending a single DB file and the instruction to run it using a standard DBay installation

• [–] provide API for pragma journal_mode; make wal the default

• [–] use GUY.datetime for dt features in stdlib

• [–] see whether we can support libSQL

  • [–] user-defined functions in WASM

• [–] review BEGIN CONCURRENT allows multiple writers

• [–] review the below, adjust default settings accordingly:

  • [–] SQLite Optimizations for Ultra High-Performance, HN, adjust default settings accordingly
  • [–] SQLite performance tuning: concurrent reads, multiple GBs and 100k SELECTs/s, HN

• [–] see whether could support libSQL, as described in WebAssembly functions for your SQLite-compatible database

• [–] ensure reasonable defaults:

  • pragma journal_mode=WAL;: always use WAL mode
  • always use IMMEDIATE transactions (needed so busy_timeout is always honored)
  • pragma busy_timeout = 60000; (1min waiting time for writers if a writer is busy)
  • see https://kerkour.com/sqlite-for-servers#use-immediate-transactions, https://lobste.rs/s/fxkk7v/why_does_sqlite_production_have_such_bad, https://til.simonwillison.net/sqlite/enabling-wal-mode

Is Done

• [+] implement DBay::do() as a method that unifies all of better-sqlite3's Statement::run(), Statement::iterate(), and Database::execute().
• [+] allow to call DBay::do -> ... with a synchronous function with the same semantics as DBay::with_transaction -> ....
• [+] allow to call DBay::do { mode: 'deferred', }, -> ....
• [+] make db = new DBay() an instance of Function that, when called, runs DBay::do() Database::execute(). statement = DBay::prepare.insert_into.<table> [ 'field1', 'field2', ..., ]
• [+] change classname(s) from Dbay to DBay to avoid spelling variant proliferation
• [+] let db.do() accept prepared statement objects.
• [+] make first_row(), all_rows() etc accept statements and strings
• [+] at the moment we use cfg.prefix for (inherently schema-less) UDF names (and require a trailing underscore to be part of the prefix), and cfg.schema for plugin-in-specific DB tables and views; in the future, we should use a single parameter for both (and make the underscore implicit). In addition, it should be possible to choose whether a plugin will create its objects with a prefix (in the same schema as the main DB) or within another schema.
• [+] fix generated SQL insert statements without explicit fields
• [+] add hidden E attribute to instance giving access to error classes (mainly for plugin use)
• [+] implement as_object: ( key, sql, P... ) ->
• [+] modify time stamp format to make it viable for use in file names on most systems
  • new format is YYYYMMDD-HHmmssZ, e.g. 20220426-171916Z is the time of this writing
• [+] fix datetime output to use different formats for input, output so output contains literal Z instead of numerical offset
• [+] fix build-sqlite3: Permission denied bug
  • occurs when publishing with pnpm version minor && pnpm publish --access public && git push
  • does not occur when publishing with npm version minor && npm publish --access public && git push
• [+] implement macros so one could write eg select * from foo( x ) as d; to get select * from ( select a, b, c from blah order by 1 ) as d (i.e. inline expansion)
  • done in separate project dbay-sql-macros
  • [+] implement 'pseudo-functions' / macros:
    • no internal logic (for now), just function composition
    • use common prefix e.g. @ as in @f := ( a, b ) -> g( a, @h( b ) ); (define function f() with two parameters, calls g(), @h(), where g() is a built-in SQLite function and @h() is another macro)
    • use re-writing such that definitions are removed / turned into comments, calls are resolved in-place. Could even consider to implement user-defined datatypes as in create table d ( name @nonempty_text, email @email ); where data type annotations are replaced with their basic types (both text here) and the statement is amended with check clauses.
      • might want to split this out into a separate dbay-sql project
      • consider to use https://github.com/Rich-Harris/code-red for parsing arguments part
      • decide whether declarations made within an aborted transaction should be undone as well (probably: yes)
• [+] concurrent writes w/ WAL mode:
  • [+] dbw = dbr?
  • [+] generated inserts to db.alt
  • [+] UDFs?
• [+] All UDFs are now created for both the primary (db.sqlt1) and secondary (db.alt.sqlt1) connections to avoid surprising messages like no such table: f when doing concurrent writes.