Introduction
What PL/Lua isPL/Lua is an implementation of Lua as a loadable procedural language for PostgreSQL: with PL/Lua you can use PostgreSQL functions and triggers written in the Lua programming language.
Procedural languages offer many extra capabilities to PostgreSQL, similar to C language extensions: control structures, more complex computations than allowed by SQL, access to user-defined types and database functions and operators, and restriction to trusted execution.
PL/Lua brings the power and simplicity of Lua to PostgreSQL, including: small memory footprint, simple syntax, lexical scoping, functions as first-class values, and coroutines for non-preemptive threading. As a simple example, consider the following hello function:
# CREATE FUNCTION hello(name text) RETURNS text AS $$
return string.format("Hello, %s!", name)
$$ LANGUAGE pllua;
CREATE FUNCTION
# SELECT hello('PostgreSQL');
hello
--------------------
Hello, PostgreSQL!
(1 row)
The next sections present more examples where other features are used. In the Languages section the two flavors of PL/Lua are described; the Functions section details how functions are registered in PL/Lua and how arguments are treated; Database access presents the SPI interface to PL/Lua; and Triggers shows how triggers can be declared.
PL/Lua is licensed under the same license as Lua -- the MIT license -- and so can be freely used for academic and commercial purposes. Please refer to the Installation section for more details.
Languages
Trusted and Untrusted PL/LuaPL/Lua is available as either a trusted (pllua) or an
untrusted (plluau) language. In plluau the
user has access to a full-blown Lua environment, similar to the regular
interpreter: all libraries are loaded, the user can access the global table
freely, and modules can be loaded. Only database superusers are allowed to
create functions using this untrusted version of PL/Lua.
Unprivileged users can only create functions using the trusted version of
PL/Lua, pllua. The environment in pllua is more
restricted: only table, string, and
math libraries are fully loaded, the os library is
restricted, the package library is not available, that is, there
is no module system (including require), and the global table is
restricted for writing. The following table summarizes the differences:
| Table | plluau |
pllua |
table, string, math |
All functions | All functions |
os |
All functions | date, clock, time, difftime |
package (module system) |
All functions | None |
_G (global environment) |
Free access | Writing is restricted |
Even though the module system is absent in pllua, PL/Lua
allows for modules to be automatically loaded after creating the environment:
all entries in table pllua.init are
require'd at startup.
To facilitate the use of PL/Lua and following the tradition of other PLs,
the global table is aliased to shared. Moreover, write access to
the global table in pllua is restricted to avoid pollution;
global variables should then be created with
setshared.
Finally, errors in PL/Lua are propagated to the calling query and the
transaction is aborted if the error is not caught. Messages can be emitted by
log, info, notice,
and warning at log levels LOG, INFO, NOTICE, and WARNING
respectively. In particular, print emits log messages of level
INFO.
log(msg)
Emits message msg at log level LOG. Similar functions
info, notice, and warning have the same
signature but emit msg at their respective log levels.
setshared(varname [, value])
Sets global varname to value, which defaults to
true. It is semantically equivalent to
shared[varname] = value.
Types
Data values in PL/LuaPL/Lua makes no conversion of function arguments to string/text form between Lua and PostgreSQL. Basic data types are natively supported, that is, converted directly, by value, to a Lua equivalent. The following table shows type equivalences:
| PostgreSQL type | Lua type |
bool |
boolean |
float4, float8, int2, int4 |
number |
text, char, varchar |
string |
| Base, domain | userdata |
| Arrays, composite | table |
Base and domain types other than the ones in the first three rows in the
table are converted to a raw datum userdata in Lua with a suitable
__tostring metamethod based on the type's output function.
Conversely, fromstring takes a type name and a string and returns
a raw datum from the provided type's input function. Arrays are converted to
Lua tables with integer indices, while composite types become tables with keys
corresponding to attribute names.
fromstring(tname, s)
Returns a raw datum userdata for s of type tname
using tname's input function to convert s.
Functions
Functions in PL/LuaPL/Lua functions are created according to the following prototype:
CREATE FUNCTION func(args) RETURNS rettype AS $$ -- Lua function body $$ LANGUAGE [pllua | plluau];
where args are usually named arguments. The value
returned by func is converted to a datum of type
rettype unless rettype is void.
The function body is composed as below to become a typical Lua chunk:
local _U, func -- _U is upvalue func = function(argnames) -- Lua function body end return func
Note the upvalue _U that can be later declared in the
function body (see examples below.)
If any of the arguments provided to create function is not
named then argnames gets substituted to ..., that
is, func becomes vararg.
The resulting chunk is then compiled and stored in the registry of the PL/Lua state as a function with the same name. It is important to have the above structure in mind when writing PL/Lua functions. As an example, consider the following function:
CREATE FUNCTION max(a integer, b integer) RETURNS integer AS $$ if a == nil then return b end -- first arg is NULL? if b == nil then return a end -- second arg is NULL? return a > b and a or b -- return max(a, b) $$ LANGUAGE pllua;
Note that max is not strict and returns NULL when
both a and b are NULL.
Since functions in PL/Lua are stored with their declared names, they can be recursive:
CREATE FUNCTION fib(n int) RETURNS int as $$
if n < 3 then
return n
else
return fib(n - 1) + fib(n - 2)
end
$$ LANGUAGE pllua;
Moreover, as can be seen in the composition of func above,
PL/Lua functions are actually closures on the upvalue
_U. The user can think of _U as
local cache to func that could — and should! — be
used instead of the global state to store values. Quick example:
CREATE FUNCTION counter() RETURNS int AS $$
while true do
_U = _U + 1
coroutine.yield(_U)
end
end
do
_U = 0 -- counter
counter = coroutine.wrap(counter)
$$ LANGUAGE pllua;
Function counter is similar to an iterator, returning
consecutive integers every time it is called, starting at one. Note that we
need to add end to finish the function definition body and
do to start a new block since the process of function composition
always appends an end. It is important to observe that what
actually gets defined as counter is a wrapper around a
coroutine.
From Types we know that composite types can be accessed as tables with keys corresponding to attribute names:
CREATE TYPE greeting AS (how text, who text); CREATE FUNCTION makegreeting (g greeting, f text) RETURNS text AS $$ return string.format(f, g.how, g.who) $$ LANGUAGE pllua;
Set-returning functions (SRFs) are implemented in PL/Lua using coroutines.
When a SRF func is first called a new Lua thread is created
and func is pushed along with its arguments onto the new thread's
stack. A new result is then returned whenever func yields and
func is done when the coroutine suspends or finishes. Using our
composite type from above, we can define
CREATE FUNCTION greetingset (how text, who text[])
RETURNS SETOF greeting AS $$
for _, name in ipairs(who) do
coroutine.yield{how=how, who=name}
end
$$ LANGUAGE pllua;
with this usage example:
# SELECT makegreeting(greetingset, '%s, %s!') FROM
(SELECT greetingset('Hello', ARRAY['foo', 'bar', 'psql'])) AS q;
makegreeting
--------------
Hello, foo!
Hello, bar!
Hello, psql!
(3 rows)
Now, to further illustrate the use of arrays in PL/Lua, we adapt an example from Programming in Lua:
CREATE FUNCTION perm (a text[]) RETURNS SETOF text[] AS $$
_U(a, #a)
end
do
_U = function (a, n) -- permgen in PiL
if n == 0 then
coroutine.yield(a) -- return next SRF row
else
for i = 1, n do
a[n], a[i] = a[i], a[n] -- i-th element as last one
_U(a, n - 1) -- recurse on head
a[n], a[i] = a[i], a[n] -- restore i-th element
end
end
end
$$ LANGUAGE pllua;
As stated in Languages, it is possible to access
the global table of PL/Lua's state. However, as noted before, since PL/Lua
functions are closures, creating global variables should be restricted to
cases where data is to be shared between different functions. The following
simple example defines a getter-setter pair to access a shared variable
counter:
CREATE FUNCTION getcounter() RETURNS integer AS $$
if shared.counter == nil then -- not cached?
setshared("counter", 0)
end
return counter -- _G.counter == shared.counter
$$ LANGUAGE pllua;
CREATE FUNCTION setcounter(c integer) RETURNS void AS $$
if shared.counter == nil then -- not cached?
setshared("counter", c)
else
counter = c -- _G.counter == shared.counter
end
$$ LANGUAGE pllua;
Examples
Let's revisit our (rather inefficient) recursive Fibonacci function
fib. A better version uses tail recursion:
CREATE FUNCTION fibt(n integer) RETURNS integer AS $$
return _U(n, 0, 1)
end
_U = function(n, a, b) -- tail recursive
if n < 1 then
return b
else
return _U(n - 1, b, a + b)
end
$$ LANGUAGE pllua;
We can also use the upvalue _U as a cache to store previous
elements in the sequence and obtain a memoized version:
CREATE FUNCTION fibm(n integer) RETURNS integer AS $$
if n < 3 then
return n
else
local v = _U[n]
if not v then
v = fibm(n - 1) + fibm(n - 2)
_U[n] = v
end
return v
end
end
do _U = {} -- memoize
$$ LANGUAGE pllua;
Finally, we can implement an iterator similar to counter:
CREATE FUNCTION fibi() RETURNS integer AS $$
while true do
_U.curr, _U.next = _U.next, _U.curr + _U.next
coroutine.yield(_U.curr)
end
end
do
_U = {curr = 0, next = 1}
fibi = coroutine.wrap(fibi)
$$ LANGUAGE pllua;
Anonymous blocks
Anonymous code blocks are also supported in PL/Lua. The following prototype
DO $$ -- Lua chunk $$ LANGUAGE [pllua | plluau];
compiles and executes the Lua chunk. Here are some examples:
DO $$ print(_VERSION) $$ LANGUAGE pllua;
DO $$
local ffi = assert(require("ffi")); -- LuaJIT
ffi.cdef[[ double lgamma (double); ]]
mathx = ffi.load("m")
$$ LANGUAGE plluau; -- note: untrusted due to "require"
CREATE FUNCTION lfactorial (n int) RETURNS double precision AS $$
return mathx.lgamma(n + 1)
$$ LANGUAGE plluau;
Database Access
Server interface in PL/LuaThe server interface in PL/Lua comprises the methods in table
server and userdata plan, cursor,
tuple, and tupletable. The entry point to the
SPI is the table
server: server.execute executes a SQL command,
server.find retrieves a cursor, and
server.prepare prepares, but does not execute, a SQL command into
a plan.
A tuple represents a composite type, record, or row. It can be
accessed similarly to a Lua table, by simply indexing fields in the composite
type as keys. A tuple can be used as a return value, just like a table, for
functions that return a complex type. Tuple sets, like the ones returned by
server.execute, plan:execute, and
cursor:fetch, are stored in a tupletable. A tupletable
is similar to an integer-keyed Lua table.
server.execute(cmd, readonly [, count])
Executes the SQL statement cmd for count rows. If
readonly is true, the command is assumed to be
read-only and execution overhead is reduced. If count is zero
then the command is executed for all rows that it applies to; otherwise at
most count rows are returned. count defaults to
zero. server.execute returns a tupletable.
server.rows(cmd)
Returns a function so that the construction
for row in server.rows(cmd) do -- body end
iterates over the tuples in the read-only SQL statement
cmd.
server.prepare(cmd, argtypes)
Prepares and returns a plan from SQL statement cmd. If
cmd specifies input parameters, their types should be specified
in table argtypes. The plan can be executed with
plan:execute. The returned plan
should not be used outside the current invocation of
server.prepare since it is freed by SPI_finish. Use
plan:save if you wish to store the plan
for latter application.
server.find(name)
Finds an existing cursor with name name and returns a cursor
userdatum or nil if the cursor cannot be found.
Plans
Plans are used when a command is to be executed repeatedly,
possibly with different arguments. In this case, we can prepare a plan with
server.prepare and execute it later with
plan:execute (or using a cursor). It is also possible to save a
plan with plan:save if we want to keep the plan for longer than
the current transaction.
plan:execute(args, readonly [, count])
Executes a previously prepared plan with parameters in table
args. readonly and count have the same
meaning as in server.execute.
plan:getcursor(args, readonly [, name])
Sets up a cursor with name name from a prepared plan. If
name is not a string a random name is selected by the system.
readonly has the same meaning as in
server.execute.
plan:rows(args)
Returns a function so that the construction
for row in plan:rows(args) do -- body end
iterates over the tuples in the execution of a previously prepared
read-only plan with parameters in table args. It is
semantically equivalent to:
function plan:rows (cmd)
local c = self:getcursor(nil, true) -- read-only
return function()
local r = c:fetch(1)
if r == nil then
c:close()
return nil
else
return r[1]
end
end
end
plan:issaved()
Returns true if plan is saved and false
otherwise.
plan:save()
Saves a prepared plan for subsequent invocations in the current session.
Cursors
Cursors execute previously prepared plans. Cursors provide a more powerful abstraction than simply executing a plan, since we can fetch results and move in a query both forward and backward. Moreover, we can limit the number of rows to be retrieved, and so avoid memory overruns for large queries in contrast to direct plan execution. Another advantage is that cursors can outlive the current procedure, living to the end of the current transaction.
cursor:fetch([count])
Fetches at most count rows from a cursor. If
count is nil or zero then all rows are fetched. If
count is negative the fetching runs backward.
cursor:move([count])
Skips count rows in a cursor, where count
defaults to zero. If count is negative the moving runs
backward.
cursor:close()
Closes a cursor.
Examples
Let's start with a simple example using cursors:
CREATE TABLE sometable ( sid int, sname text, sdata text);
CREATE FUNCTION get_rows (i_name text) RETURNS SETOF sometable AS $$
if _U == nil then -- plan not cached?
local cmd = "SELECT sid, sname, sdata FROM sometable WHERE sname=$1"
_U = server.prepare(cmd, {"text"}):save()
end
local c = _U:getcursor({i_name}, true) -- read-only
while true do
local r = c:fetch(1)
if r == nil then break end
r = r[1]
coroutine.yield{sid=r.sid, sname=r.sname, sdata=r.sdata}
end
c:close()
$$ LANGUAGE pllua;
This SRF works as a pipeline: it uses _U to store a saved
plan, while local variable c is a cursor that we use to fetch, at
each loop iteration, a row from _U and then yield a new row.
Note that local r is a tupletable and we need to access
r[1].
A more concise version uses plan:rows():
CREATE FUNCTION get_rows (i_name text) RETURNS SETOF sometable AS $$
if _U == nil then -- plan not cached?
local cmd = "SELECT sid, sname, sdata FROM sometable WHERE sname=$1"
_U = server.prepare(cmd, {"text"}):save()
end
for r in _U:rows{i_name} do
coroutine.yield(r) -- yield tuple
end
$$ LANGUAGE pllua;
Now, for a more elaborate example, let's store a binary tree:
CREATE TABLE tree (id int PRIMARY KEY, lchild int, rchild int);
which we can fill using:
CREATE FUNCTION filltree (t text, n int) RETURNS void AS $$
local p = server.prepare("insert into " .. t .. " values($1, $2, $3)",
{"int4", "int4", "int4"})
for i = 1, n do
local lchild, rchild = 2 * i, 2 * i + 1 -- siblings
p:execute{i, lchild, rchild} -- insert values
end
$$ LANGUAGE pllua;
Local variable p stores a prepared plan for insertion with
three parameters as values, while the actual insertion is executed in the
loop.
We can perform a preorder traversal of the tree with:
CREATE FUNCTION preorder (t text, s int) RETURNS SETOF int AS $$
coroutine.yield(s)
local q = server.execute("select * from " .. t .. " where id=" .. s,
true, 1) -- read-only, only 1 result
if q ~= nil then
local lchild, rchild = q[1].lchild, q[1].rchild -- store before next query
if lchild ~= nil then preorder(t, lchild) end
if rchild ~= nil then preorder(t, rchild) end
end
$$ LANGUAGE pllua;
The traversal is recursive and we simply execute a query in every call and
store its result in tupletable q. It is important to store the
fields in q[1] in locals before next query, since q
gets updated in the next query.
In preorder we executed a query many times. For our postorder
traversal below we prepare a plan, save it, and cache in a
_U table. Instead of executing the plan, we get a cursor
from it and fetch only one row, as before.
CREATE FUNCTION postorder (t text, s int) RETURNS SETOF int AS $$
local p = _U[t]
if p == nil then -- plan not cached?
p = server.prepare("select * from " .. t .. " where id=$1", {"int4"})
_U[t] = p:save()
end
local c = p:getcursor({s}, true) -- read-only
local q = c:fetch(1) -- one row
if q ~= nil then
local lchild, rchild = q[1].lchild, q[1].rchild -- store before next query
c:close()
if lchild ~= nil then postorder(t, lchild) end
if rchild ~= nil then postorder(t, rchild) end
end
coroutine.yield(s)
end
do _U = {} -- plan cache
$$ LANGUAGE pllua;
Triggers
Triggers in PL/LuaTriggers can be defined in PL/Lua as usual by just creating a function
returning trigger. When a function returns a trigger, PL/Lua
creates a (global) table trigger containing all
necessary information. The trigger table is described below.
| Key | Value |
name |
trigger name |
when |
"before" if trigger fired before or
"after" if trigger fired after
|
level |
"row" if row-level trigger or
"statement" if statement-level trigger
|
operation |
"insert", "update", "delete",
or "truncate" depending ontrigger operation |
relation |
Lua table describing the relation with keys: name is
relation name (string)and attributes is a table with
relation attributes as string keys
|
row |
Tuple representing the row-level trigger's target: in update
operations holds the new row, otherwise holds the old row. row is nil instatement-level triggers. |
old |
Tuple representing the old row in an update before row-level operation. |
Trigger functions in PL/Lua don't return; instead, only for
row-level-before operations, the tuple in trigger.row is read for
the actual returned value. The returned tuple has then the same effect for
general triggers: if nil the operation for the current row is
skipped, a modified tuple will be inserted or updated for insert and update
operations, and trigger.row should not be modified if none of the
two previous outcomes is expected.
Example
Let's restrict row operations in our previous binary tree example: updates
are not allowed, deletions are only possible on leaf parents, and insertions
should not introduce cycles and occur only at leaves. We store closures in
_U that have prepared plans as upvalues.
create function treetrigger() returns trigger as $$
local row, operation = trigger.row, trigger.operation
if operation == "update" then
trigger.row = nil -- updates not allowed
elseif operation == "insert" then
local id, lchild, rchild = row.id, row.lchild, row.rchild
if lchild == rchild or id == lchild or id == rchild -- avoid loops
or (lchild ~= nil and _U.intree(lchild)) -- avoid cycles
or (rchild ~= nil and _U.intree(rchild))
or (_U.nonemptytree() and not _U.isleaf(id)) -- not leaf?
then
trigger.row = nil -- skip operation
end
else -- operation == "delete"
if not _U.isleafparent(row.id) then -- not both leaf parent?
trigger.row = nil
end
end
end
do
local getter = function(cmd, ...)
local plan = server.prepare(cmd, {...}):save()
return function(...)
return plan:execute({...}, true) ~= nil
end
end
_U = { -- plan closures
nonemptytree = getter("select * from tree"),
intree = getter("select node from (select id as node from tree "
.. "union select lchild from tree union select rchild from tree) as q "
.. "where node=$1", "int4"),
isleaf = getter("select leaf from (select lchild as leaf from tree "
.. "union select rchild from tree except select id from tree) as q "
.. "where leaf=$1", "int4"),
isleafparent = getter("select lp from (select id as lp from tree "
.. "except select ti.id from tree ti join tree tl on ti.lchild=tl.id "
.. "join tree tr on ti.rchild=tr.id) as q where lp=$1", "int4")
}
$$ language pllua;
Finally, we set the trigger on table tree:
create trigger tree_trigger before insert or update or delete on tree for each row execute procedure treetrigger();
Installation
How to obtain and install PL/LuaPL/Lua is distributed as a source package and can be obtained at PgFoundry. Depending on how Lua is installed in your system you might have to edit the Makefile. After that the source package is installed like any regular PostgreSQL module, that is, after downloading and unpacking, just run:
$ make && sudo make install $ psql -c "CREATE EXTENSION pllua" mydb
The pllua extension installs both trusted and untrusted
flavors of PL/Lua and creates the module table pllua.init.
Alternatively, a systemwide installation though the PL template facility can
be achieved with:
INSERT INTO pg_catalog.pg_pltemplate
VALUES ('pllua', true, 'pllua_call_handler', 'pllua_validator', '$libdir/pllua', NULL);
INSERT INTO pg_catalog.pg_pltemplate
VALUES ('plluau', false, 'plluau_call_handler', 'plluau_validator', '$libdir/pllua', NULL);
License
Copyright (c) 2008 Luis Carvalho
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
