Make it easy to upsert on traditional RDBMS like MySQL, PostgreSQL, and SQLite3—hey look NoSQL!. Transparently creates (and re-uses) stored procedures/functions when necessary.
You pass it a bare-metal connection to the database like Mysql2::Client (from mysql2 gem on MRI) or Java::OrgPostgresqlJdbc4::Jdbc4Connection (from jdbc-postgres on Jruby).
As databases start to natively support SQL MERGE (which is basically upsert), this library will take advantage (but you won't have to change your code).
Does not depend on ActiveRecord.
70–90%+ faster than emulating upsert with ActiveRecord.
Supports MRI and JRuby.
You pass a selector that uniquely identifies a row, whether it exists or not. You also pass a setter, attributes that should be set on that row.
Syntax inspired by mongo-ruby-driver's update method.
connection = Mysql2::Client.new([...])
table_name = :pets
upsert = Upsert.new connection, table_name
# N times...
upsert.row({:name => 'Jerry'}, :breed => 'beagle', :created_at => Time.now)The created_at and created_on columns are used for inserts, but ignored on updates.
So just to reiterate you've got a selector and a setter:
selector = { :name => 'Jerry' }
setter = { :breed => 'beagle' }
upsert.row(selector, setter)By organizing your upserts into a batch, we can do work behind the scenes to make them faster.
connection = Mysql2::Client.new([...])
Upsert.batch(connection, :pets) do |upsert|
# N times...
upsert.row({:name => 'Jerry'}, :breed => 'beagle')
upsert.row({:name => 'Pierre'}, :breed => 'tabby')
endBatch mode is tested to be about 80% faster on PostgreSQL, MySQL, and SQLite3 than other ways to emulate upsert (see the tests, which fail if they are not faster).
require 'upsert/active_record_upsert'
# N times...
Pet.upsert({:name => 'Jerry'}, :breed => 'beagle')Pull requests for any of these would be greatly appreciated:
- Cache JDBC PreparedStatement objects.
- Sanity check my three benchmarks (four if you include activerecord-import on MySQL). Do they accurately represent optimized alternatives?
- Provide
require 'upsert/debug'that will make sure you are selecting on columns that have unique indexes - Test that
Upsertinstances accept arbitrary columns, even within a batch, which is what people probably expect. - @antage's idea for "true" upserting: (from seamusabshere#17)
selector = { id: 15 }
update_setter = { count: Upsert.sql('count + 1') }
insert_setter = { count: 1 }
upsert.row_with_two_setter(update_setter, insert_setter, selector)
We use upsert for big data processing at Brighter Planet and in production at
Originally written to speed up the data_miner data mining library.
| * | MySQL | PostgreSQL | SQLite3 |
|---|---|---|---|
| MRI | mysql2 | pg | sqlite3 |
| JRuby | jdbc-mysql | jdbc-postgres | jdbc-sqlite3 |
See below for details about what SQL MERGE trick (emulation of upsert) is used, performance, code examples, etc.
(assuming that one of the other three supported drivers is being used under the covers)
Upsert.new Pet.connection, Pet.table_nameDepends on the driver being used!
Depends on the driver being used!
On MRI, use the mysql2 driver.
require 'mysql2'
connection = Mysql2::Connection.new(:username => 'root', :password => 'password', :database => 'upsert_test')
table_name = :pets
upsert = Upsert.new(connection, table_name)On JRuby, use the jdbc-mysql driver.
require 'jdbc/mysql'
java.sql.DriverManager.register_driver com.mysql.jdbc.Driver.new
connection = java.sql.DriverManager.get_connection "jdbc:mysql://127.0.0.1/mydatabase?user=root&password=password"From the tests (updated 11/7/12):
Upsert was 82% faster than find + new/set/save
Upsert was 85% faster than find_or_create + update_attributes
Upsert was 90% faster than create + rescue/find/update
Upsert was 46% faster than faking upserts with activerecord-import
Thanks to Dennis Hennen's StackOverflow response!!
CREATE PROCEDURE upsert_pets_SEL_name_A_tag_number_SET_name_A_tag_number(`name_sel` varchar(255), `tag_number_sel` int(11), `name_set` varchar(255), `tag_number_set` int(11))
BEGIN
DECLARE done BOOLEAN;
REPEAT
BEGIN
-- If there is a unique key constraint error then
-- someone made a concurrent insert. Reset the sentinel
-- and try again.
DECLARE ER_DUP_UNIQUE CONDITION FOR 23000;
DECLARE ER_INTEG CONDITION FOR 1062;
DECLARE CONTINUE HANDLER FOR ER_DUP_UNIQUE BEGIN
SET done = FALSE;
END;
DECLARE CONTINUE HANDLER FOR ER_INTEG BEGIN
SET done = TRUE;
END;
SET done = TRUE;
SELECT COUNT(*) INTO @count FROM `pets` WHERE `name` = `name_sel` AND `tag_number` = `tag_number_sel`;
-- Race condition here. If a concurrent INSERT is made after
-- the SELECT but before the INSERT below we'll get a duplicate
-- key error. But the handler above will take care of that.
IF @count > 0 THEN
-- UPDATE table_name SET b = b_SET WHERE a = a_SEL;
UPDATE `pets` SET `name` = `name_set`, `tag_number` = `tag_number_set` WHERE `name` = `name_sel` AND `tag_number` = `tag_number_sel`;
ELSE
-- INSERT INTO table_name (a, b) VALUES (k, data);
INSERT INTO `pets` (`name`, `tag_number`) VALUES (`name_set`, `tag_number_set`);
END IF;
END;
UNTIL done END REPEAT;
ENDOn MRI, use the pg driver.
require 'pg'
connection = PG.connect(:dbname => 'upsert_test')
table_name = :pets
upsert = Upsert.new(connection, table_name)On JRuby, use the jdbc-postgres driver.
require 'jdbc/postgres'
java.sql.DriverManager.register_driver org.postgresql.Driver.new
connection = java.sql.DriverManager.get_connection "jdbc:postgresql://127.0.0.1/mydatabase?user=root&password=password"If you want to use HStore, make the pg-hstore gem available and pass a Hash in setters:
gem 'pg-hstore'
require 'pg_hstore'
upsert.row({:name => 'Bill'}, :mydata => {:a => 1, :b => 2})From the tests (updated 9/21/12):
Upsert was 72% faster than find + new/set/save
Upsert was 79% faster than find_or_create + update_attributes
Upsert was 83% faster than create + rescue/find/update
# (can't compare to activerecord-import because you can't fake it on pg)
Adapted from the canonical PostgreSQL upsert example:
CREATE OR REPLACE FUNCTION upsert_pets_SEL_name_A_tag_number_SET_name_A_tag_number("name_sel" character varying(255), "tag_number_sel" integer, "name_set" character varying(255), "tag_number_set" integer) RETURNS VOID AS
$$
DECLARE
first_try INTEGER := 1;
BEGIN
LOOP
-- first try to update the key
UPDATE "pets" SET "name" = "name_set", "tag_number" = "tag_number_set"
WHERE "name" = "name_sel" AND "tag_number" = "tag_number_sel";
IF found THEN
RETURN;
END IF;
-- not there, so try to insert the key
-- if someone else inserts the same key concurrently,
-- we could get a unique-key failure
BEGIN
INSERT INTO "pets"("name", "tag_number") VALUES ("name_set", "tag_number_set");
RETURN;
EXCEPTION WHEN unique_violation THEN
-- seamusabshere 9/20/12 only retry once
IF (first_try = 1) THEN
first_try := 0;
ELSE
RETURN;
END IF;
-- Do nothing, and loop to try the UPDATE again.
END;
END LOOP;
END;
$$
LANGUAGE plpgsql;I slightly modified it so that it only retries once - don't want infinite loops.
On MRI, use the sqlite3 driver.
require 'sqlite3'
connection = SQLite3::Database.open(':memory:')
table_name = :pets
upsert = Upsert.new(connection, table_name)On JRuby, use the jdbc-sqlite3 driver.
# TODO somebody please verify
require 'jdbc/sqlite3'
java.sql.DriverManager.register_driver org.sqlite.Driver.new
connection = java.sql.DriverManager.get_connection "jdbc:sqlite://127.0.0.1/mydatabase?user=root&password=password"From the tests (updated 9/21/12):
Upsert was 77% faster than find + new/set/save
Upsert was 80% faster than find_or_create + update_attributes
Upsert was 85% faster than create + rescue/find/update
# (can't compare to activerecord-import because you can't fake it on sqlite3)
Thanks to @dan04's answer on StackOverflow:
INSERT OR IGNORE INTO visits VALUES (127.0.0.1, 1);
UPDATE visits SET visits = 1 WHERE ip LIKE 127.0.0.1;In addition to correctness, the library's tests check that it is
- Faster than comparable upsert techniques
- Compatible with supported databases
As below, all you need is a raw database connection like a Mysql2::Connection, PG::Connection or a SQLite3::Database. These are equivalent:
# with activerecord
Upsert.new ActiveRecord::Base.connection, :pets
# with activerecord, prettier
Upsert.new Pet.connection, Pet.table_name
# without activerecord
Upsert.new Mysql2::Connection.new([...]), :petsYou could also use activerecord-import to upsert:
Pet.import columns, all_values, :timestamps => false, :on_duplicate_key_update => columnsactiverecord-import, however, only works on MySQL and requires ActiveRecord—and if all you are doing is upserts, upsert is tested to be 40% faster. And you don't have to put all of the rows to be upserted into a single huge array - you can batch them using Upsert.batch.
We don't have any logic to convert integers into strings, strings into integers, etc. in order to satisfy PostgreSQL/etc.'s strictness on this issue.
So if you try to upsert a blank string ('') into an integer field in PostgreSQL, you will get an error.
Datetimes are immediately converted to UTC and sent to the database as ISO8601 strings.
If you're using MySQL, make sure server/connection timezone is UTC. If you're using Rails and/or ActiveRecord, you might want to check ActiveRecord::Base.default_timezone... it should probably be :utc.
In general, run some upserts and make sure datetimes get persisted like you expect.
Copyright 2012 Seamus Abshere