Improved CSV import speed by several orders of magnitude.

pokedex/__init__.py

@@ -27,8 +27,7 @@ def command_csvimport(engine_uri, directory='.'):
 
     from sqlalchemy.orm.attributes import instrumentation_registry
 
-    # Use autocommit in case rows fail due to foreign key incest
+    session = connect(engine_uri)
-    session = connect(engine_uri, autocommit=True, autoflush=False)
 
     metadata.create_all()
 
@@ -74,11 +73,18 @@ def command_csvimport(engine_uri, directory='.'):
         reader = csv.reader(csvfile, lineterminator='\n')
         column_names = [unicode(column) for column in reader.next()]
 
-        # Self-referential tables may contain rows with foreign keys of
+        # Self-referential tables may contain rows with foreign keys of other
-        # other rows in the same table that do not yet exist.  We'll keep
+        # rows in the same table that do not yet exist.  Pull these out and add
-        # a running list of these and try inserting them again after the
+        # them to the session last
-        # rest are done
+        # ASSUMPTION: Self-referential tables have a single PK called "id"
-        failed_rows = []
+        deferred_rows = []  # ( row referring to id, [foreign ids we need] )
+        seen_ids = {}       # primary key we've seen => 1
+
+        # Fetch foreign key columns that point at this table, if any
+        self_ref_columns = []
+        for column in table_obj.c:
+            if any(_.references(table_obj) for _ in column.foreign_keys):
+                self_ref_columns.append(column)
 
         for csvs in reader:
             row = table_class()
@@ -101,33 +107,44 @@ def command_csvimport(engine_uri, directory='.'):
 
                 setattr(row, column_name, value)
 
-            try:
+            # May need to stash this row and add it later if it refers to a
-                session.add(row)
+            # later row in this table
-                session.flush()
+            if self_ref_columns:
-            except IntegrityError, e:
+                foreign_ids = [getattr(row, _.name) for _ in self_ref_columns]
-                failed_rows.append(row)
+                foreign_ids = [_ for _ in foreign_ids if _]  # remove NULL ids
 
-        # Loop over the failed rows and keep trying to insert them.  If a loop
+                if not foreign_ids:
-        # doesn't manage to insert any rows, bail.
+                    # NULL key.  Remember this row and add as usual.
-        do_another_loop = True
+                    seen_ids[row.id] = 1
-        while failed_rows and do_another_loop:
-            do_another_loop = False
 
-            for i, row in enumerate(failed_rows):
+                elif all(_ in seen_ids for _ in foreign_ids):
-                try:
+                    # Non-NULL key we've already seen.  Remember it and commit
-                    session.add(row)
+                    # so we know the old row exists when we add the new one
-                    session.flush()
+                    session.commit()
+                    seen_ids[row.id] = 1
 
-                    # Success!
+                else:
-                    del failed_rows[i]
+                    # Non-NULL future id.  Save this and insert it later!
-                    do_another_loop = True
+                    deferred_rows.append((row, foreign_ids))
-                except IntegrityError, e:
+                    continue
-                    pass
 
-        if failed_rows:
+            session.add(row)
-            print len(failed_rows), "rows failed"
+
-        else:
+        session.commit()
-            print 'loaded'
+
+        # Attempt to add any spare rows we've collected
+        for row, foreign_ids in deferred_rows:
+            if not all(_ in seen_ids for _ in foreign_ids):
+                # Could happen if row A refers to B which refers to C.
+                # This is ridiculous and doesn't happen in my data so far
+                raise ValueError("Too many levels of self-reference!  "
+                                 "Row was: " + str(row.__dict__))
+
+            session.add(row)
+            seen_ids[row.id] = 1
+            session.commit()
+
+        print 'loaded'
 
 def command_csvexport(engine_uri, directory='.'):
     import csv