databricks-tellr-app 0.4.3.dev7__tar.gz → 0.4.3.dev9__tar.gz

{databricks_tellr_app-0.4.3.dev7 → databricks_tellr_app-0.4.3.dev9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: databricks-tellr-app
-Version: 0.4.3.dev7
+Version: 0.4.3.dev9
 Summary: Tellr application package for Databricks Apps
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown

{databricks_tellr_app-0.4.3.dev7 → databricks_tellr_app-0.4.3.dev9}/databricks_tellr_app/run.py

@@ -51,6 +51,29 @@ def init_database(seed_databricks_defaults: bool = False) -> None:
         logger.error(f"Failed to initialize database tables: {e}\n{tb}")
         raise SystemExit(1) from e
 
+    # Data migrations/backfills that used to run in the FastAPI lifespan now run
+    # HERE, once, before the server forks its workers — the uvicorn workers must
+    # never execute migration code (4 of them racing the migration chain on boot
+    # wedged startup). init_db() above runs the schema/data migrations; these two
+    # convert legacy profile/session rows to the agent_config shape.
+    logger.info("Migrating profiles/sessions to agent_config...")
+    try:
+        from src.core.database import get_session_local
+        from src.core.migrate_profiles_to_agent_config import (
+            backfill_sessions,
+            migrate_profiles,
+        )
+        migrated = migrate_profiles(get_session_local())
+        if migrated:
+            logger.info(f"Migrated {migrated} profiles to agent_config")
+        backfilled = backfill_sessions(get_session_local())
+        if backfilled:
+            logger.info(f"Backfilled {backfilled} sessions with agent_config")
+    except Exception as e:
+        tb = traceback.format_exc()
+        logger.error(f"Failed to migrate profiles/sessions to agent_config: {e}\n{tb}")
+        raise SystemExit(1) from e
+
     # Seed default content
     logger.info(f"Seeding defaults (include_databricks={seed_databricks_defaults})...")
     try:

{databricks_tellr_app-0.4.3.dev7 → databricks_tellr_app-0.4.3.dev9}/databricks_tellr_app.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: databricks-tellr-app
-Version: 0.4.3.dev7
+Version: 0.4.3.dev9
 Summary: Tellr application package for Databricks Apps
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown

{databricks_tellr_app-0.4.3.dev7 → databricks_tellr_app-0.4.3.dev9}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "databricks-tellr-app"
-version = "0.4.3.dev7"
+version = "0.4.3.dev9"
 description = "Tellr application package for Databricks Apps"
 readme = "README.md"
 requires-python = ">=3.10"

{databricks_tellr_app-0.4.3.dev7 → databricks_tellr_app-0.4.3.dev9}/src/api/main.py

@@ -34,13 +34,10 @@ from src.api.routes.settings import (
 from src.api.services.export_job_queue import start_export_worker
 from src.api.services.job_queue import recover_stuck_requests, start_worker
 from src.core.database import (
-    get_session_local,
-    init_db,
     is_lakebase_environment,
     start_token_refresh,
     stop_token_refresh,
 )
-from src.core.migrate_profiles_to_agent_config import migrate_profiles, backfill_sessions
 
 logger = logging.getLogger(__name__)
 
@@ -92,8 +89,8 @@ async def lifespan(app: FastAPI):
     else:
         logger.info("Pytest detected: skipping MCP session manager startup")
 
-    # Start Lakebase token refresh if running in Databricks Apps
-    # Must happen before init_db() so OAuth token is ready for database connections
+    # Start Lakebase token refresh if running in Databricks Apps. Workers serve
+    # DB-backed requests, so they need fresh OAuth tokens.
     if is_lakebase_environment():
         try:
             await start_token_refresh()
@@ -102,24 +99,11 @@ async def lifespan(app: FastAPI):
             logger.error(f"Failed to start Lakebase token refresh: {e}")
             raise
 
-    # Initialize database tables (idempotent - only creates tables that don't exist)
-    is_pytest = os.getenv("PYTEST_CURRENT_TEST") is not None
-    if not is_pytest:
-        try:
-            init_db()
-            logger.info("Database tables initialized")
-
-            migrated = migrate_profiles(get_session_local())
-            if migrated:
-                logger.info(f"Migrated {migrated} profiles to agent_config")
-            backfilled = backfill_sessions(get_session_local())
-            if backfilled:
-                logger.info(f"Backfilled {backfilled} sessions with agent_config")
-        except Exception as e:
-            logger.error(f"Failed to initialize database: {e}")
-            raise
-    else:
-        logger.info("Pytest detected: skipping database initialization")
+    # NOTE: database migrations run ONCE before the server starts, in
+    # run.py::init_database (invoked as its own step by the Databricks Apps boot
+    # command, and by scripts/init_database.py locally) — NOT here. This lifespan
+    # runs in every uvicorn worker, and having 4 workers race the migration chain
+    # on boot wedged startup. Workers must never run migration code.
 
     if IS_PRODUCTION:
         logger.info("Production mode: serving frontend from package assets")

{databricks_tellr_app-0.4.3.dev7 → databricks_tellr_app-0.4.3.dev9}/src/core/database.py

@@ -1126,6 +1126,41 @@ _DECK_PLACEHOLDER_COLUMNS: tuple[tuple[str, str], ...] = (
 )
 
 
+def _deck_columns_have_int_placeholder(conn, live_targets, _qual, is_sqlite) -> bool:
+    """Does any target column still hold an int-style ``{{image:<digits>}}`` placeholder?
+
+    The cheap gate that keeps :func:`_migrate_rewrite_deck_image_placeholders` a
+    near no-op on every startup after its one-time rewrite. Read-only and
+    short-circuiting (``LIMIT 1``), so it never takes a write lock and stops at the
+    first match.
+
+    - PostgreSQL / Lakebase (the real target): a POSIX regex, ``{{image:`` + one or
+      more digits + ``}}``. Precise — a token placeholder is
+      ``{{image:<token>}}`` where the token is ``secrets.token_urlsafe(32)`` (43
+      chars of ``[A-Za-z0-9_-]``), which cannot be all digits, so only genuine
+      legacy int placeholders match.
+    - SQLite (tests only): ``GLOB`` has no ``+`` quantifier, so the pattern matches
+      ``{{image:`` followed by a single digit. Deliberately BROAD — it can also
+      match a token placeholder that happens to start with a digit, which at worst
+      makes the idempotent rewrite loop run when it strictly need not. It never
+      misses a real int placeholder, so correctness is preserved; SQLite is not a
+      deployment target and its data is tiny.
+    """
+    from sqlalchemy import text
+
+    for table_name, column_name in live_targets:
+        q = _qual(table_name)
+        if is_sqlite:
+            stmt = text(f"SELECT 1 FROM {q} WHERE {column_name} GLOB :pat LIMIT 1")
+            params = {"pat": "*{{image:[0-9]*"}
+        else:
+            stmt = text(f"SELECT 1 FROM {q} WHERE {column_name} ~ :pat LIMIT 1")
+            params = {"pat": r"\{\{image:[0-9]+\}\}"}
+        if conn.execute(stmt, params).first() is not None:
+            return True
+    return False
+
+
 def _migrate_rewrite_deck_image_placeholders(conn, inspector, schema, _qual, is_sqlite) -> None:
     """Rewrite stored deck ``{{image:<int-id>}}`` placeholders to ``{{image:<token>}}`` (SDR-4437 F-TM-7).
 
@@ -1151,9 +1186,14 @@ def _migrate_rewrite_deck_image_placeholders(conn, inspector, schema, _qual, is_
       per-match lookup into ``image_assets``, not a fixed template.
     - **Prefix-safe.** The literal ``{{image:5}}`` (closing braces included) never
       matches inside ``{{image:51}}``, so there is no id-prefix collision.
-    - **Idempotent + cheap.** The ``LIKE`` guard means once the int placeholders are
-      gone (first run) re-runs match nothing; unreferenced or hard-deleted ids are
-      simply never touched. Cost is O(images) set-based UPDATEs, fine at startup.
+    - **Idempotent + cheap on repeat.** This runs on EVERY worker at EVERY startup,
+      so a fast gate (:func:`_deck_columns_have_int_placeholder`) short-circuits the
+      whole thing once no int-style placeholder remains — turning steady-state boots
+      into a few read-only ``LIMIT 1`` scans instead of O(images) unindexable
+      leading-wildcard ``LIKE`` UPDATEs per column (that unconditional per-boot cost
+      stalled real multi-worker startup on prod-scale data). The per-UPDATE ``LIKE``
+      guard is the second layer: within a run it skips rows without this id, and
+      unreferenced or hard-deleted ids are never touched.
     - **No jsonb cast.** Every target column (:data:`_DECK_PLACEHOLDER_COLUMNS`) is
       ``TEXT`` on both backends, so ``replace()`` applies directly. Tokens are
       ``[A-Za-z0-9_-]`` (JSON-safe), so text-level replacement inside a JSON blob
@@ -1187,6 +1227,16 @@ def _migrate_rewrite_deck_image_placeholders(conn, inspector, schema, _qual, is_
     if not live_targets:
         return
 
+    # Fast idempotency gate. This migration runs on EVERY worker at EVERY startup
+    # (init_db in the FastAPI lifespan), so once the one-time rewrite is done it must
+    # NOT re-issue the per-image UPDATEs: each is an unindexable leading-wildcard
+    # LIKE scan of a large-blob deck table, and O(images) of them per column across
+    # several concurrent workers stalls startup on prod-scale data. If no target
+    # column still holds an int-style {{image:<digits>}} placeholder, there is
+    # nothing to do — return before touching the image rows or the deck tables.
+    if not _deck_columns_have_int_placeholder(conn, live_targets, _qual, is_sqlite):
+        return
+
     rows = conn.execute(
         text(f"SELECT id, token FROM {_qual('image_assets')} WHERE token IS NOT NULL")
     ).fetchall()