Migrations in Entity Framework in a collaborative

问题:

We have multiple developers working on a project that uses Entity Framework 5.0. Every developer uses his own local SQL 2012 database so he can develop and test without impeding others.

At first, we used a hybrid of automatic migrations and code-based migrations. That didn't work well at all so we decided to disable automatic migrations and to only allow code-based. I should add that we started again with a clean database without a 'corrupted' _MigrationsHistory from all the automatic migrations.

So now the workflow is:

1. Developer changes his datamodel
2. Does add-migration <Name> and applies it to his database with update-database.
3. Checks in the datamodel change and the migration into Git.
4. Another developer pulls, receives the changes and applies it to his database.

So far, this worked well. However before today it was usually just me who made the migrations and the others applied them. But today there were migrations from three developers. I just pulled those migrations, did an update-database which went fine.

I also had a change to my own datamodel however so at the end of the update-database it gave me a warning that I still wasn't up to date so I did add-migration <my migration>. However when it scaffolded the migration, it gave me the changes of all the migrations I had already applied to the database. So: it tried to drop columns that had already been dropped, tried to create a table that already existed, etc.

How can that be? My assumption was that EF would just check the _MigrationsHistory table and find out which migrations weren't present in the table yet and apply those one by one ordered by the timestamp that's part of the name. But apparently not, because even when I undo my own changes and I have a clean environment it still complains my database isn't in sync with the model. But I just pulled those changes and applied them to my database. It is in sync. I can see the migrations that I just applied in the _MigrationsHistory table too.

The only thing I can think of is that I added a property to a datamodel that wouldn't result in a database change (I added a List<X> to datamodel Y where X is the many in the one-to-many relationship. This wouldn't result in a database change as X already had a foreign key to Y). Could that be it? If so, that's really fragile because there's no way to add a migration for that since there's no database change and I'm not sure how to fix this either.

I'm not sure how to deal with this, because I can of course just edit what it scaffolded and remove everything that has already been applied to my database. But then what? I check it in and then some other developer gets the same message that his database isn't up to date even after applying my new changes, scaffolds his own changes, gets the same nonsense scaffolding, edits it, checks it in and then the next developer gets it. It becomes a vicious circle and a similar one to what we had when we used automatic migrations and I thought we had fixed that by switching to code-based only. I can't trust it right now to do the right thing and it's a nightmare to work with like this.

What I also tried is adding the migrations I pulled from my coworkers one by one with update-database -t:201211091112102_<migrationname> but to no avail. It still gives me the erroneous scaffold.

So what did we do wrong here, or is EF simply not built for collaboration like this?

UPDATE

I created a reproducible test case, it's a bit of a lengthy dance though in order to simulate this multi user/multi database scenario.

https://github.com/JulianR/EfMigrationsTest/

Steps to reproduce when you have the above project (these steps are also present in the code):

1. add-migration Init
2. update-database (on database 'TestDb')
3. Change connection string to point to TestDb1
4. update-database on TestDb1
5. Uncomment property Foo on class Test
6. add-migration M1 to add property Foo to TestDb1
7. Comment out Test.Foo again
8. Change connection string to point to TestDb2
9. Exclude migration M1 from project so it doesn't get applied to TestDb2
10. Uncomment property Bar on class Test
11. update-database to apply Init migration to TestDb2
12. add-migration M2 to add property Bar to TestDb2
13. Change connection string to point to the original TestDb again
14. Include migration M1 into the project again
15. Uncomment property Foo on class Test
16. Uncomment property SomeInt on class Test
17. update-database
18. add-migration M3
19. update-database, get an error because M3 tries to add column Foo to database TestDb which was already just added by migration M1.

The above is to simulate three users, where user 1 inits his database, the other two use his initialization to create their database as well. Then user 2 and user 3 both make their own change to the datamodel and add it to source control together with the migrations needed to apply the changes. Then user 1 pulls the changes of user 2 and 3 while user 1 has also made a change to the database himself. Then user 1 calls update-database to apply the changes of user 2 and 3. He then scaffolds his own migration which then erroneously adds a change from user 2 or 3 to the scaffolded migration which causes an error when applied to user 1's database.

回答1:

You need to manually resolve migration conflicts just like you would code conflicts. If you update and there are new migrations, you need to ensure that the metadata behind the last migration matches the current model. To update the metadata of the migration, re-issue the Add-Migration command for it.

For example, before step 17 (Update-Database) in your scenario, you should issue the following command

Add-Migration M2

This will update the metadata to bring it in sync with your current model. Now when you try and add M3, it should be blank since you have not made any further model changes.

回答2:

You need to add a blank "merge" migration that will reset the snapshot of the latest migration in the .resx file. Do this using the IgnoreChanges switch:

Add-Migration <migration name> -IgnoreChanges

See here for an explanation

回答3:

Option 1: Add a blank ‘merge’ migration

1. Ensure any pending model changes in your local code base have been written to a migration. This step ensures you don’t miss any legitimate changes when it comes time to generate the blank migration.
2. Sync with source control.
3. Run Update-Database to apply any new migrations that other developers have checked in. ** Note:****if you don’t get any warnings from the Update-Database command then there were no new migrations from other developers and there is no need to perform any further merging.
4. Run Add-Migration –IgnoreChanges (e.g. Add-Migration Merge –IgnoreChanges). This generates a migration with all the metadata (including a snapshot of the current model) but will ignore any changes it detects when comparing the current model to the snapshot in the last migrations (meaning you get a blank Up and Down method).
5. Continue developing, or submit to source control (after running your unit tests of course).

Option 2: Update the model snapshot in the last migration

1. Ensure any pending model changes in your local code base have been written to a migration. This step ensures you don’t miss any legitimate changes when it comes time to generate the blank migration.
2. Sync with the source control.
3. Run Update-Database to apply any new migrations that other developers have checked in. ** Note:****if you don’t get any warnings from the Update-Database command then there were no new migrations from other developers and there is no need to perform any further merging.
4. Run Update-Database –TargetMigration (in the example we’ve been following this would be Update-Database –TargetMigration AddRating). This roles the database back to the state of the second last migration – effectively ‘un-applying’ the last migration from the database. ** Note:****This step is required to make it safe to edit the metadata of the migration since the metadata is also stored in the __MigrationsHistoryTable of the database. This is why you should only use this option if the last migration is only in your local code base. If other databases had the last migration applied you would also have to roll them back and re-apply the last migration to update the metadata.
5. Run Add-Migration (in the example we’ve been following this would be something like Add-Migration 201311062215252_AddReaders). ** Note:****You need to include the timestamp so that migrations knows you want to edit the existing migration rather than scaffolding a new one. This will update the metadata for the last migration to match the current model. You’ll get the following warning when the command completes, but that’s exactly what you want. “Only the Designer Code for migration '201311062215252_AddReaders' was re-scaffolded. To re-scaffold the entire migration, use the -Force parameter.”
6. Run Update-Database to re-apply the latest migration with the updated metadata.
7. Continue developing, or submit to source control (after running your unit tests of course).

MSDN have a great article on this. Please go through it.

Entity Framework Code First Migrations in Team Environments

回答4:

We are having similar issues in our environment, here is what we've figured out so far and how we got around it:

When you have changes that you have applied (update-database) but not checked in, and then you receive changes from another developer who doesn't have your changes, this is where things seem to get out of sync. In our experience, it seems like the meta data that is saved for your own changes get over written by the meta-data from the other developer when you do the update-database process. The other developer doesn't have your changes, so the meta-data that gets saved is no longer a real reflection of your database. When EF does a comparison after that, it 'thinks' that your changes are actually new again because of the meta data change.

A simple, admittedly ugly workaround is to do another migration, and wipe out it's contents so you have empty up() and empty down() methods. Apply that migration and check it into source control and let everyone sync to that. This simply syncs up all of the meta data so everyone has all of the changes accounted for.

回答5:

I have added an issue on codeplex, this issue causes many a head scratching in our team too.

The link is https://entityframework.codeplex.com/workitem/1670

回答6:

I have put some thought into this and I hope I will contribute to the different opinions and practices presented here.

Consider what your local migrations actually represent. When working locally with a dev database, I use migrations to update the database in the most convenient way possible when adding columns etc to tables, adding new entities etc.

So, Add-Migration checks my current model (let's call it model b) against my previous model (model a) and generates a migration to go from a => b in the database.

To me it makes very little sense to try and merge my migrations with anyone elses migrations, if everyone indeed has their own database and there then exists some kind of stage / test / dev / production database servers in the organization. This all depends on how the team has it set up, but it makes sense to insulate each other from changes that other people make if you want to truly work in a distributed manner.

Well, if you work distributed and have some entity, Person, for example, that you work on. For some reason, lots of other people are also working on it. So, you add and remove properties on Person as needed for your particular story in the sprint (we're all working agile here, aren't we?), like Social Security number that you first made into an integer because you aren't that bright and then to a string etc.

You add FirstName And LastName.

You are then done and you have ten weird up and down migrations (you probably removed some of them while working since they were just crap) and you fetch some changes from the central Git repo. Wow. Your colleague Bob also needed some names, maybe you should've talked to each other?

Anyways, he has added NameFirst and NameLast, I guess... so what do you do? Well, you merge, refactor, change so it has more sane names... like FirstName and LastName, you run your tests and check his code, and then you push to the central.

But what about the migrations? Well, now would be the time to make a migration moving the central repo, or the branch "test" more specifically, contain a nice little migration from its model a => model b. This migration will be one and only one migration, not ten weird ones.

Do you see what I'm getting at? We are working with nice little pocos and the comparisons of them constitute the actual migrations. So, we shouldn't merge migrations at all, in my opinion, we should have migrations-per-branch or something like that.

In fact, do we even need to create the migration in the branch after merge? Yes, if this database is updated automatically, we need to.

Another thing to consider is to never actually creating a migration before doing a pull from the central repo. That means you will both get the other team members' migration code and their changes to the model before creating your migration.

Gotta work some more, those are my thoughts on this, at least.

回答7:

The solution I was able to come up with (at least for 2 users, haven't tested for 3) is:

1. merging migrations to sync up the meta-data run update-database (this should fail), then
2. add-database and then
3. delete all of the generated code in up() and down() methods

this will still be run by update database but won't do anything, just bringing the metadata up to sync.

回答8:

I agree with @LavaEater. The core of the issue, it would seem, is that migration scaffolding should be centralised. Perhaps as part of some automated/integrated build process each time a push occurs? Thereafter the resulting migrations can be pulled from the server by team-members.

This means that their own migration scripts should not be pushed to the server.