Error when listing columns for model backed by different DB

[owst]

We have a couple of models that are backed by a different DB to the majority of models in our app. A rough example is something like this:

class MainDBThings < ApplicationRecord
end

class ExternalDBThings < ApplicationRecord
  establish_connection :external_db
end

It seems that a couple of places in active_record_doctor assume that all tables can be accessed from a connection to a single DB (ActiveRecord::Base.connection); when the "different DB" table names are accessed, a "table missing" error is raised because the connection used is to the wrong DB.

For example, this line:

active_record_doctor/lib/active_record_doctor/detectors/missing_non_null_constraint.rb

Line 35 in f3b8ba5

connection.columns(table).each do |column|

uses connection, which raises an error along the lines of:

ActiveRecord::StatementInvalid: PG::UndefinedTable: ERROR:  relation "ExternalDBThings" does not exist

A possible fix might be to use models.first.columns instead of connection.columns(table) so that the per-model connection is used?

Similarly, this line has the same issue:

active_record_doctor/lib/active_record_doctor/detectors/base.rb

Line 218 in f3b8ba5

connection.columns(model.table_name).each do |column|

I'm happy to have a stab at a PR for this, but would like some guidance as to any thoughts or suggested approaches before I start. I expect the fix will be to not use connection without reference to a particular model instance.

Thanks!

[fatkodima]

Suggested changes will solve this issue, but this gem still do not have proper support for multiple databases, see #63.

Please, do open a PR.

[gregnavis]

Thanks for opening the issue, @owst! I'm working hard on adding support for multi-database systems, so expect an update soon. Would you be willing to test a release candidate?

[owst]

Sure thing @gregnavis, that sounds good - thanks!

[brunoarueira]

@gregnavis I would like to test a release candidate too, call me in! Btw, thanks for this gem 😃

[owst]

Hi @gregnavis I just came across this issue again preventing us from upgrading active_record_doctor, have you managed to make any progress on multi-DB support and do you maybe have a release candidate to test?

[gregnavis]

@owst, actually, yes, I've made some progress, but it's not done yet. I released transient_record 2.0 that supports multiple databases. That was a prerequisite for adding multi-database test cases to active_record_doctor.

[owst]

Hi @gregnavis - just following up on this again - any further progress since your last comment?

[gregnavis]

@owst, some progress, but need to put more time into that. I'm preparing 1.15 right now and created a milestone for 2.0. Multi-database support should land in 2.0.

[mculp]

@gregnavis is there a relatively simple workaround?

[owst]

Hi @gregnavis - has there been any further progress on this issue?

[tmaier]

Still hitting this on v2.0.1. Our scenario: a Rails app on PostgreSQL with a bundled gem whose models use establish_connection to a separate SQLite database.

Reproduction

# Gem's models — connected to a separate SQLite DB
module MyGem
  class ApplicationRecord < ActiveRecord::Base
    self.abstract_class = true
  end

  class Database
    def self.establish_connection!
      ApplicationRecord.establish_connection(adapter: 'sqlite3', database: 'gem.sqlite3')
    end
  end

  class Identifier < ApplicationRecord  # table "identifiers" in SQLite
    belongs_to :publication
  end
end

The main app has no identifiers table in PostgreSQL.

Error

ActiveRecord::StatementInvalid: PG::UndefinedTable: ERROR: relation "identifiers" does not exist
LINE 10:  WHERE a.attrelid = '"identifiers"'::regclass

# Crash is in missing_non_null_constraint detector:
active_record_doctor/detectors/missing_non_null_constraint.rb:47
  → connection.columns(table)
active_record_doctor/detectors/base.rb:113
  → @connection ||= ActiveRecord::Base.connection   # ← always PG

Why it crashes

missing_non_null_constraint#detect (line 36) calls models.select(&:table_exists?). This uses each model's own connection — so MyGem::Identifier.table_exists? returns true (it checks its SQLite connection). The models are then grouped by table_name, and line 47 calls connection.columns(table) which resolves to ActiveRecord::Base.connection (the PG connection). PG has no identifiers table → crash.

The core issue is the connection mismatch: discovery uses the model's connection, inspection uses the hardcoded default connection.

Our workaround

Adding the gem's tables to ignore_tables and its models to ignore_models in .active_record_doctor.rb:

ActiveRecordDoctor.configure do
  global :ignore_models, [/^MyGem::/]
  global :ignore_tables, [
    'identifiers',
    'other_gem_tables'
  ]
end

This works but requires knowing all table names from external gems, and ignore_tables doesn't apply to model-based detectors while ignore_models doesn't apply to table-based detectors.

Potential fix ideas

1. Filter out models with non-default connections early (minimal, safe)

The simplest fix — skip models whose connection pool differs from the primary one in Base#models:

def models
  base_pool = ActiveRecord::Base.connection_pool
  ActiveRecord::Base.descendants
    .reject { |model| model.connection_pool != base_pool }
    .sort_by(&:name)
end

This is a one-line guard that prevents all detectors from seeing external-DB models. It matches the implicit assumption already baked into the codebase (single connection method). Low risk, no API changes needed.

2. Use model-specific connections in detectors (targeted)

As @owst suggested in the original report, detectors that iterate models could use model.connection instead of Base#connection. For example in missing_non_null_constraint.rb:

# Before (line 47):
connection.columns(table).each do |column|

# After — use the connection from the models grouped under this table:
models.first.connection.columns(table).each do |column|

And similarly in each_attribute (base.rb:217):

# Before:
connection.columns(model.table_name).each do |column|

# After:
model.connection.columns(model.table_name).each do |column|

This is more correct but requires auditing every detector to ensure consistency.

3. Group detectors by connection pool (comprehensive, future-proof)

For full multi-database support, the runner could group all models by connection_pool, then run each detector once per pool with the appropriate connection:

def models_by_pool
  ActiveRecord::Base.descendants
    .reject(&:abstract_class?)
    .group_by(&:connection_pool)
end

# Each detector receives a scoped set of models and the matching connection
models_by_pool.each do |pool, pool_models|
  detector.run(connection: pool.connection, models: pool_models)
end

This would properly analyze all databases, not just the primary one. Biggest change, but the only approach that actually checks external-DB models rather than silently skipping them.

---

Approach 1 would be a quick win that unblocks everyone hitting this today. Approach 3 might be the proper long-term solution for the v2.0 multi-database goal.

Happy to help test any of these if there's a branch to try.