Rob B. wrote:
On Apr 22, 2010, at 9:06 AM, Marnen Laibow-Koser wrote:
Rob B. wrote:
[…]
Matt, Pito, Marnen, and anyone else,
- The opinion on whether db/schema.rb goes into the source
repository
has changed over time.
No. I’ve used Rails since 1.2.6. Every version has put a comment in
the schema.rb file that recommends putting it into version control.
I’ve been using Rails since 0.13 and I’ll restate that this opinion
has changed.
I can’t find it at the moment, but there WAS a version that
specifically said in
that comment to NOT put the file into source control. At least in
1.2.2, there
is no comment either way.
I’ll get out my old versions of Rails and look. However, it’s pretty
much irrelevant, because look what current versions say:
Note that this schema.rb definition is the authoritative source for
your database schema. If you need
to create the application database on another system, you should be
using db:schema:load, not running
all the migrations from scratch. The latter is a flawed and
unsustainable approach (the more migrations
you’ll amass, the slower it’ll run and the greater likelihood for
issues).
It’s strongly recommended to check this file into your version control
system.
That’s pretty definitive.
[…]
But you should never be running old migrations in the first place. If
you need version 1000 of the schema for a new installation, then don’t
start at zero and run 1000 migrations – just do rake db:schema:load
and
have done with it. This is the core team’s recommendation, and I
think
it’s a good one.
I have helped other developers who have created several migrations in
development, which were applied approximately when created, that were
subsequently unable to run when deployed to production.
And why was this so?
The very same
recommendations that guard against this kind of problem will make it
possible to run those 1000 migrations (or any subsequence) without
problem.
No. There is a difference between running a days-old migration and a
years-old migration.
merge the schema.rb files. If not, get a better VCS.
Yikes! No! The database itself holds the official version of the
schema. If I merge changes from a master branch into my development
branch, I will run any new migrations, but I certainly don’t want some
merge tool to give me a new schema.rb.
For changing the schema on an existing installation, I agree with you
– run the migrations. For a new installation, use rake db:schema:load.
That implies that you need schema.rb in the VCS. How much more clearly
can I say this?
Depending on the actual
content of the migrations on different branches and the order in which
they are run, the actual schema might be slightly different due to
the rules for where new columns are placed on a table.
The order of columns on a DB table is immaterial. You should be able to
run rake db:schema:load for a new installation and immediately get a
usable DB.
You’ve
got it completely backwards.
You assume that I need to run db:schema:load, which I don’t.
Why do you believe you don’t? Because you don’t understand what it’s
for? Because you never do new installations?
I’ll submit that you do need to run db:schema:load, and you just don’t
know it because you don’t understand what it’s good for.
If you
think about why the migration numbering (file naming) was changed
from
sequence number to timestamp, you’ll realize that the practice of
such
“interleaved” migrations was a much bigger pain-point than what to do
about db/schema.rb.
I don’t really understand what you’re getting at here.
When migrations were sequentially numbered, two developers on separate
branches might both create migration 005 for different purposes. This
was a problem. The chance that two developers both create migration
20100422105524 is acceptable small.
OK, now I see what you mean. Yes.
Of course, they will probably be
executed in a different order and if they both add a column to the
same table, those columns will likewise be in a different order.
So what? Column order is immaterial.
If db/
schema.rb is in the repository, then lots of commits will have
effectively meaningless changes
How do you figure that?
and unless the current HEAD has my
version is it not going to truly represent what’s in my database
schema.
Sure it is – modulo immaterial things like column order.
- Unless you’re initializing (inserting) data via migration, running
all the migrations is really not much different than doing a
db:schema:load because all the migrations are operating on empty
tables.
How can you say this with a straight face? There is no reason at
all
to run lots of migrations rather than doing a simple schema load.
The reason is deploying to an existing production database. You can’t
do a schema load.
Right. That’s what migrations are for. That’s the only thing
migrations are for – changing an existing database.
The proper way to apply the “new” migrations (which
might be kinda old if the last production deploy wasn’t so recent)
Then you’ve got bigger problems. If you can’t do lots of little
deploys, then you will wind up in integration hell anyway, and
migrations will be the least of your worries.
is
a db:migrate. As I’ve stated, you can get into trouble with mismatches
between migrations (which don’t change after being created and
shouldn’t if created properly) and models (which obviously change
over time).
If you’re deploying frequently enough, there will be no issue.
I have projects with 210 and 204 migrations as well as many with
fewer. Some of the migrations deal with rather nasty data
manipulations to maintain data relationships when the associations are
flipped around. It’s not something that I would recommend, but the
definition of “reasonable” can change dramatically when a client
shifts the way he thinks about the data and its evolution.
As an experiment, I set up a new environment for the 204 migration
project and ran the migrations from scratch. It takes about 8
minutes.
That’s a long time, but that’s the least of the problems.
There is about 1 minute of startup time, there are several
migrations that load some data including one that takes a bit over 3
minutes to put a few tens of thousands of research datapoints into a
set of tables.
Wow, it just gets worse. Seed data never, ever belongs in migrations
– partly because it makes schema loading impossible, and partly because
migrations are only about the schema, not the data. Use seed-fu or
Rails 2.3’s built-in seeding.
I’m OK with that amount of time for something as
significant as creating a new environment.
Yes, it’s a one-time task. But it’s still wrong.
(Besides, if you have to “scale up your app”, you probably
aren’t adding a new empty database, but creating a master-slave or
sharding for performance.)
Another red herring.
Well, if you can “scale up your app” by starting from an empty schema,
go ahead. Perhaps initializing the shards, but then I’d start with the
db/schema.rb from production, not something from the repository which
almost certainly reflects a development environment however close to
production that might be.
Still a red herring. I want new installations to be easy to create –
not for scalability, which would certainly use DB replication, but for
people actually creating a new, independent instance of the app – say,
a test environment, or simply someone else installing a Rails app that’s
meant for external use.
Rob, I know you know a lot about the Rails framework, but your advice
here will make dealing with databases far more difficult than it needs
to be.
If I help someone who recalls some of these nuggets of my wisdom and
experience at a time where migrations give them trouble, then I will
have made a positive difference.
However, it is that same experience that has led me to the conclusion
that keeping db/schema.rb in the source repository is wrong. It is
derived data and I would no more put it into the repository than I
would have someone put their object files compiled from C or their
class files compiled from Java in there.
schema.rb is certainly derived data to a point. However, it is also the
only reliable source of information about what the DB schema should look
like at a given point in history. I probably would not have thought to
put it into version control had I not seen that note from the core team,
but having done so and thought about it, I believe that it must go into
version control. It contains information that cannot be reliably
derived except by going through the process that it was designed to
circumvent.
To continue the compilation analogy: you might not put your binaries
into the repository with the source code – but you’d surely want to
make them available for people who didn’t care to build from source.
There are many good reasons to have schema.rb in version control. There
are few if any to do otherwise.
Best,
Marnen Laibow-Koser
http://www.marnen.org
[email protected]