▲ 82 points
back
21 comments
There is another problem unmentioned in this article with his simple table modification example... When you modify a table the DB needs to write the entire table to disk, locking it all the time. Depending on how much data is in the table and how fast your IO is this can be hours and hours of downtime on the live database, for a migration that only took a few seconds locally on your test database.
Running migrations on live databases with lots of data in them is hard. There are many strategies to work around this problem, but generally running an 'alter table' on your primary db server on a huge table while it is in use should be your very last option.
This is true, but note that the article was written in the context of PostgreSQL instead of MySQL. PostgreSQL can do many types of table migrations with no downtime (dropping columns, renaming, adding columns that default to NULL, and creating indexes). Before I made the switch from MySQL to Postgres I did some quick benchmarks which support these claims on a 5 million row test table: https://gist.github.com/1620133.
Guilty as charged for immediately assuming that the problems of the MySQL family apply to other databases. Thank you for pointing out that PostgreSQL does not necessarily have all the same limitations.
Absolutely! The author was very thorough with the change process, which is great, but downtime during ALTER should absolutely be broached as well.
A staging system with a db server provisioned similarly to production lets you "smoke test" the migrations and deploy before you take down prod -- but especially with MySQL, that happily blocks all reads during an ALTER, you should really watch out for any table ~> 1M rows (or if they're wide, even fewer than that).
If you're using postgres on Heroku you can get a fork of your production db to test things like these:
http://devcenter.heroku.com/articles/heroku-postgresql#fork_...
The issue is not performance. The issue is backward compatibility.
ALTER TABLE performance will eventually improve. PostgreSQL 9.1 lifts-off the lock-up limitation.
But what a performant ALTER TABLE will not improve is its intrusiveness to applications. When you change the schema you break the app.
That's the problem.
If you're writing your code to work against multiple schemas, how do you test it? Do you run your test suite twice, once with the pre-migration schema, and again with the post-migration schema?
For sure.
Most of the time migrations are developed in a feature branch, so the owner is responsible for splitting his branch into two or three before sending pull requests (as many as required to address issues with hot compatibility).
This means that we have two or more merges (and deploys) from master, each requiring a full run of the test suite.
Andy Delcambre wrote a post two months ago about how Engine Yard handles zero-downtime deploys with migrations. Link here: http://www.engineyard.com/blog/2011/zero-downtime-deploys-wi...
This is old news, but we successfully use https://github.com/freels/table_migrator in production not on heroku. It creates a copy of the table, performs the schema changes, copies the data over, then renames the tables for almost (i.e. 1-2 seconds) no downtime.
How does this deal with foreign key references? I thought that if you rename a table the FK references from other tables will still point at the renamed table? Discussion here: http://dev.mysql.com/doc/refman/5.0/en/rename-table.html
The Percona pt-online-schema-change tool goes to great lengths to avoid this kind of problem.
Basically, it doesn't - because it doesn't need to. Rails/ActiveRecord doesn't use those either. Problem solved :)
Interesting. As you have probably guessed by now, I'm not a Rails developer and therefore did not know this. I was surprised to read about this after your comment and find that Active Record went the lowest common denominator route with this and therefore gave up any native foreign key integrity support.
There is nothing to stop you using foreign key constraints with rails, the activerecord migration api includes methods to create them for the major db adapters and there are plugins to automate the process to some extent. It's not the Rails Way™ because it sacrifices some database-agnosticity, and therefore almost noone does it. People achieve the same behaviour with application-level validations in the model.
I'd wager that a lot of the big professional rails deployments are doing FK constraints though.
The gem 'foreigner' makes them painless, and supports Postgres, Oracle, and I think MSSQL.
I have never missed foreign keys on the projects I've worked on that don't use them.
This is because those projects typically sit on databases that are only expected to be accessed via web API and not directly from some other source.
Which is basically the Rails philosophy. Only let the app talk to the DB, and let the only outside interface to the DB be through the ORM that has to do most of the data validations anyway.
I use percona tools pt-online-schema-change http://www.percona.com/doc/percona-toolkit/2.0/pt-online-sch..., wrote a custom gem for it while ago
At SoundCloud we developed Large Hadron Migrator [1] to work around the problem of locked tables with ALTER and MySQL. It creates a table with the new schema, copies the data from the "old" table to the new one in chunks (utilizing INSERT IGNORE and triggers) and then switches the two tables. The switching is not yet atomic, as there was a bug with RENAME TABLE.
[1] https://github.com/soundcloud/large-hadron-migrator
Edit: URL
Does it make sense to add extra meta-dummy columns in your tables for the future? in case you do need to add an alter to add a column you can just reuse those extra columns instead?
What's the Django equivalent of this?
You pretty much have to do the same thing - release code using new models before migrating the db. Django tends to spell out all columns in its SQL statements, so when you delete one from the db, but fail to upgrade the code you will run into a db exception.
I'd also recommend using South for migrations http://south.aeracode.org/