[arturonline]

Problem

Connecting to a SQL Server instance with many databases, without a default database set, makes sqlit unresponsive for minutes. It often has to be killed and restarted.

Reproduced against a real server with 236 user databases (about 280k tables, 209k views, 594k routines and 3M routine parameters in total).

Cause

When no database is selected and the provider supports cross-database queries, the autocomplete schema loader indexes every non-system database on the server (_load_schema_directly → get_databases → _on_databases_loaded(all_dbs)).

For each database it runs 3 metadata queries: tables, views, and routines with parameters. Each query is fast on its own (about 1s per database in total), but:

1. Executor starvation. All 3 × N queries (about 700 here) go through session.executor, which is the single thread that owns the connection. Anything the user does (expanding a tree node, running a query) waits behind them for several minutes.
2. UI thread load. Every table and view adds 3–4 entries to _table_metadata, and every routine is appended to the cache. That is millions of items processed in idle-scheduler batches on the main thread, plus several hundred worker threads spawned at once.

The result looks like a hang.

Fix

Add MAX_SERVER_WIDE_INDEX_DATABASES = 20. When the server-wide database list is longer than this, autocomplete indexing is skipped and a warning is shown:

Autocomplete not indexed: the server has N databases. Select a database in the explorer (Enter on it) to index it.

Selecting a database already calls _load_schema_cache() with that database as the only target, so indexing picks up there with no further changes. Servers with 20 or fewer databases, and connections with a database configured, behave exactly as before.

The guard is applied in both loading paths:

• _on_databases_loaded (threaded / idle-scheduler path). A selected database always arrives here as a single-item list, so it is never affected.
• _load_schema_cache_async, right after the database list is fetched.

Workaround for users on older versions

Set a database on the connection. Autocomplete then indexes only that database.

Tests

New tests/unit/test_autocomplete_many_databases.py:

• more than the limit: nothing is indexed, the spinner stops, and one warning is emitted
• at the limit: databases are indexed, no warning
• a single selected database is always indexed

The new tests pass when run directly. I couldn't run the full suite with pytest in my local environment.

Possible follow-ups (not in this PR)

• Make the limit configurable in settings.
• Index the active database first and the rest lazily or on demand, instead of all or nothing.

🤖 Generated with Claude Code