Stray changes detected only sometimes with revision --autogenerate

89 views
Skip to first unread message

Daniel Krebs

unread,
Sep 29, 2020, 4:17:57 AM9/29/20
to sqlalchemy-alembic
Hi,

we're having rather strange problems with Alembic 1.4.2 and Postgres 12, detecting stray changes sometimes but also sometimes not. I already dug through the code but I increasingly get the feel that this is rooted somewhere in the interaction between alembic and Postgres.

But let me explain our setup first. We maintain a set of SQLAlchemy model definitions that we evolve over time and use alembic to migrate the database accordingly. For every customer, we add a new Postgres schema to the database identified by a corresponding UUID (e.g. a3d74dcc-1634-33a5-ff74-235f3a7c6322). See [1] for our (stripped) env.py.

Since we want all customers to use the same DDL, we want common migrations that disregard the schema name altogether. It seems that this is not a supported use-case of alembic, so I hacked around it such that alembic only sees schema-less objects and I run SET search_path TO "uuid" before each migration (see env.py [1]). The patch should be rather straight-forward and can be found at [2].

Now the issue that we're facing is, that sometimes autogenerate detects changes in one or two customer schemas that are not real [3]. Deleting the migration and creating a new one often doesn't detect the changes anymore or for a different customer/schema. The tables that are incorrectly found to have changed also change over time.

My current workaround is to autogenerate migrations until alembic "does the right thing". I know that patching alembic is not the best base upon which to ask for support, but to be honest I am running out of theories what is going wrong here. Hope someone can help or point me into the right direction :)

Cheers,
Daniel




Mike Bayer

unread,
Sep 29, 2020, 10:35:34 AM9/29/20
to sqlalchem...@googlegroups.com


On Tue, Sep 29, 2020, at 9:17 AM, Daniel Krebs wrote:
Hi,

we're having rather strange problems with Alembic 1.4.2 and Postgres 12, detecting stray changes sometimes but also sometimes not. I already dug through the code but I increasingly get the feel that this is rooted somewhere in the interaction between alembic and Postgres.

But let me explain our setup first. We maintain a set of SQLAlchemy model definitions that we evolve over time and use alembic to migrate the database accordingly. For every customer, we add a new Postgres schema to the database identified by a corresponding UUID (e.g. a3d74dcc-1634-33a5-ff74-235f3a7c6322). See [1] for our (stripped) env.py.

Since we want all customers to use the same DDL, we want common migrations that disregard the schema name altogether. It seems that this is not a supported use-case of alembic, so I hacked around it such that alembic only sees schema-less objects and I run SET search_path TO "uuid" before each migration (see env.py [1]). The patch should be rather straight-forward and can be found at [2].

This is supported by SQLAlchemy more robustly using the schema_translate_map feature, see https://docs.sqlalchemy.org/en/13/core/connections.html#schema-translating .   this feature allows schema names to render explicitly without the need to manipulate search_path.

However, when using that approach you'd normally be running autogenerate from just one schema, since you are looking to generate migrations just once and then run them on every tenant.

so it's not clear why autogenerate is being run on every tenant explicitly - you'd have just one "model" schema that's the one where you actually run autogenerate upon.   there would be no need to consider other schemas and include_schemas would be set to False.

As far as being able to filter out schemas across many, the current approach is the include_object hook, which is insufficient for many schemas as it does not block the reflection of all the tables.   a new hook include_name is being added in the coming weeks that allows for pre-emptive inclusion or exclusion of specific schema names.




Now the issue that we're facing is, that sometimes autogenerate detects changes in one or two customer schemas that are not real [3]. Deleting the migration and creating a new one often doesn't detect the changes anymore or for a different customer/schema. The tables that are incorrectly found to have changed also change over time.

I don't have much insight onto this other than looking at concurrent activities on the database.   "sometimes" usually indicates a race condition of some sort, and the "autogenerate" process is strictly one of reading data. 




My current workaround is to autogenerate migrations until alembic "does the right thing". I know that patching alembic is not the best base upon which to ask for support, but to be honest I am running out of theories what is going wrong here. Hope someone can help or point me into the right direction :)

Cheers,
Daniel





--
You received this message because you are subscribed to the Google Groups "sqlalchemy-alembic" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sqlalchemy-alem...@googlegroups.com.

Mike Bayer

unread,
Sep 29, 2020, 10:46:54 AM9/29/20
to sqlalchem...@googlegroups.com


On Tue, Sep 29, 2020, at 3:35 PM, Mike Bayer wrote:


On Tue, Sep 29, 2020, at 9:17 AM, Daniel Krebs wrote:
Hi,

we're having rather strange problems with Alembic 1.4.2 and Postgres 12, detecting stray changes sometimes but also sometimes not. I already dug through the code but I increasingly get the feel that this is rooted somewhere in the interaction between alembic and Postgres.

But let me explain our setup first. We maintain a set of SQLAlchemy model definitions that we evolve over time and use alembic to migrate the database accordingly. For every customer, we add a new Postgres schema to the database identified by a corresponding UUID (e.g. a3d74dcc-1634-33a5-ff74-235f3a7c6322). See [1] for our (stripped) env.py.

Since we want all customers to use the same DDL, we want common migrations that disregard the schema name altogether. It seems that this is not a supported use-case of alembic, so I hacked around it such that alembic only sees schema-less objects and I run SET search_path TO "uuid" before each migration (see env.py [1]). The patch should be rather straight-forward and can be found at [2].

This is supported by SQLAlchemy more robustly using the schema_translate_map feature, see https://docs.sqlalchemy.org/en/13/core/connections.html#schema-translating .   this feature allows schema names to render explicitly without the need to manipulate search_path.

this feature may not work completely with Alembic however, I'm evaluating that now.

still doesnt shed any light on your sporadic problem, however.



However, when using that approach you'd normally be running autogenerate from just one schema, since you are looking to generate migrations just once and then run them on every tenant.

so it's not clear why autogenerate is being run on every tenant explicitly - you'd have just one "model" schema that's the one where you actually run autogenerate upon.   there would be no need to consider other schemas and include_schemas would be set to False.

As far as being able to filter out schemas across many, the current approach is the include_object hook, which is insufficient for many schemas as it does not block the reflection of all the tables.   a new hook include_name is being added in the coming weeks that allows for pre-emptive inclusion or exclusion of specific schema names.




Now the issue that we're facing is, that sometimes autogenerate detects changes in one or two customer schemas that are not real [3]. Deleting the migration and creating a new one often doesn't detect the changes anymore or for a different customer/schema. The tables that are incorrectly found to have changed also change over time.

I don't have much insight onto this other than looking at concurrent activities on the database.   "sometimes" usually indicates a race condition of some sort, and the "autogenerate" process is strictly one of reading data. 




My current workaround is to autogenerate migrations until alembic "does the right thing". I know that patching alembic is not the best base upon which to ask for support, but to be honest I am running out of theories what is going wrong here. Hope someone can help or point me into the right direction :)

Cheers,
Daniel





--
You received this message because you are subscribed to the Google Groups "sqlalchemy-alembic" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sqlalchemy-alem...@googlegroups.com.


--
You received this message because you are subscribed to the Google Groups "sqlalchemy-alembic" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sqlalchemy-alem...@googlegroups.com.

Daniel Krebs

unread,
Sep 29, 2020, 11:01:09 AM9/29/20
to sqlalchem...@googlegroups.com
Hi Mike,

thanks a bunch for the quick response!

>> This is supported by SQLAlchemy more robustly using the
>> schema_translate_map feature, see
>> https://docs.sqlalchemy.org/en/13/core/connections.html#schema-translating
>> .   this feature allows schema names to render explicitly without the
>> need to manipulate search_path.
>
> this feature may not work completely with Alembic however, I'm
> evaluating that now.
>
> still doesnt shed any light on your sporadic problem, however.

It's been some time already since I've implemented this, so can't recall
the exact reason why I went down that road but I know for sure that I
was trying to make it work with schema_translate_map but didn't succeed.
I would have very much preferred a solution without patching alembic it
seemed to my last resort at the time. I'd be more than happy to find a
solution here :)


> However, when using that approach you'd normally be running
autogenerate from just one schema, since you are looking to generate
migrations just once and then run them on every tenant.
>
> so it's not clear why autogenerate is being run on every tenant
explicitly - you'd have just one "model" schema that's the one where you
actually run autogenerate upon. there would be no need to consider
other schemas and include_schemas would be set to False.

Indeed that sounds like a good idea. So you'd suggest to keep one schema
just for autogeneration purposes and then apply the migrations to all
customer schemas? Is that possible from within alembic or would you wrap
some tooling around alembic in order to apply migrations one-by-one to
every schema?


> As far as being able to filter out schemas across many, the current
approach is the include_object hook, which is insufficient for many
schemas as it does not block the reflection of all the tables. a new
hook include_name is being added in the coming weeks that allows for
pre-emptive inclusion or exclusion of specific schema names.

I tried the include_object hook back then, but I was facing reflection
issues if I remember correctly. We're using TimescaleDB [1] in the same
database which adds some schemas of its own that alembic doesn't seem to
be able to handle.


> I don't have much insight onto this other than looking at concurrent
activities on the database. "sometimes" usually indicates a race
condition of some sort, and the "autogenerate" process is strictly one
of reading data.

Indeed I'm suspecting the same thing. However, I'm not the most
knowledgable DBA nor do I have an understanding of how alembic actually
compares the current state of the DB with the SA models. Could you maybe
provide a more concrete point what and where to look for?


Thank you for taking the time to help!


Cheers,
Daniel

[1] https://www.timescale.com/

Mike Bayer

unread,
Sep 29, 2020, 1:04:59 PM9/29/20
to sqlalchem...@googlegroups.com


Hi, so I added a quick recipe to the site just now just so that the "set search path" idea is documented to some extent, that is at https://alembic.sqlalchemy.org/en/latest/cookbook.html#rudimental-schema-level-multi-tenancy-for-postgresql-databases .

Re: autogenerate, if you have many schemas that all contain an exact copy of the same schema, then autogenerate implies you'd be running this exactly once for just one of the schemas, or a "master" schema, such that you generate the .py files that will then be applicable to all the schemas.  The recipe above discusses this, and in particular it works in a completely schema agnostic fashion, you would set include_schemas=False in the config and allow the PostgreSQL search_path to handle all movement between schemas.

then for the actual migration "upgrade" run, you would run those files against each schema in sequence, again making use of search_path in order to select each schema.

as far as how alembic is looking at current tables you would need to turn on SQL logging, using the "debug" level, such as in alembic.ini (if you're using that):

[logger_sqlalchemy]
level = DEBUG
handlers =
qualname = sqlalchemy.engine



this will emit a lot of queries and result sets.  You'd have to capture that logging and then grep through for the "nonexistent" schema - it will be present in a result set that Alembic is capturing, most fundamentally the one it uses to get all the schema names.
-- 
You received this message because you are subscribed to the Google Groups "sqlalchemy-alembic" group.
To unsubscribe from this group and stop receiving emails from it, send an email to sqlalchemy-alem...@googlegroups.com.

Daniel Krebs

unread,
Sep 30, 2020, 8:52:52 AM9/30/20
to sqlalchem...@googlegroups.com
Hi Mike,

I adapted our setup to the new cookbook recipe and it basically fixed
both problems! I don't need to maintain my
patches anymore (thanks to setting
`connection.dialect.default_schema_name`, this did the trick!) and the
detected stray changes are also gone. I did some investigation with
increased log level as you suggested and came to the conclusion that it
must have been a bad idea to manipulate the `conn_table_names` object
in-place. But since that's gone now, everything is working as expected \o/

Thank you so much for your help!

Cheers,
Daniel
>> <mailto:sqlalchemy-alem...@googlegroups.com>.
>> To view this discussion on the web visit
>> https://groups.google.com/d/msgid/sqlalchemy-alembic/fc2bc5e4-274b-a9e7-0566-3b745c4e5fc5%40enlyze.com.
>>
>
> --
> You received this message because you are subscribed to a topic in the
> Google Groups "sqlalchemy-alembic" group.
> To unsubscribe from this topic, visit
> https://groups.google.com/d/topic/sqlalchemy-alembic/ZlIG6qOULcA/unsubscribe.
> To unsubscribe from this group and all its topics, send an email to
> sqlalchemy-alem...@googlegroups.com
> <mailto:sqlalchemy-alem...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/sqlalchemy-alembic/cc68de85-ebf3-4eb7-bf83-35b8082a80e4%40www.fastmail.com
> <https://groups.google.com/d/msgid/sqlalchemy-alembic/cc68de85-ebf3-4eb7-bf83-35b8082a80e4%40www.fastmail.com?utm_medium=email&utm_source=footer>.

Mike Bayer

unread,
Sep 30, 2020, 8:57:46 AM9/30/20
to sqlalchem...@googlegroups.com
Hi Daniel -

I didn't expect it to go so well, but that's great!    
To unsubscribe from this group and stop receiving emails from it, send an email to sqlalchemy-alem...@googlegroups.com.


Reply all
Reply to author
Forward
0 new messages