Update UPSERT comment now that native upserts are the default

changelog.d/13924.misc

@@ -0,0 +1 @@
+Update an innaccurate comment in Synapse's upsert database helper.

synapse/storage/database.py

@@ -1141,17 +1141,48 @@ async def simple_upsert(
         desc: str = "simple_upsert",
         lock: bool = True,
     ) -> bool:
-        """
+        """Insert a row with values + insertion_values; on conflict, update with values.
+
+        All of our supported databases accept the nonstandard "upsert" statement in
+        their dialect of SQL. We call this a "native upsert". The syntax looks roughly
+        like:
+
+            INSERT INTO table VALUES (values + insertion_values)
+            ON CONFLICT (keyvalues)
+            DO UPDATE SET (values); -- overwrite `values` columns only
+
+        If (values) is empty, the resulting query is slighlty simpler:
+
+            INSERT INTO table VALUES (insertion_values)
+            ON CONFLICT (keyvalues)
+            DO NOTHING;             -- do not overwrite any columns
+
+        This function is a helper to build such queries.
+
+        In order for upserts to make sense, the database must be able to determine when
+        an upsert CONFLICTs with an existing row. Postgres and SQLite ensure this by
+        requiring that a unique index exist on the column names used to detect a
+        conflict (i.e. `keyvalues.keys()`).
+
+        If there is no such index, we can "emulate" an upsert with a SELECT followed
+        by either an INSERT or an UPDATE. This is unsafe: we cannot make the same
+        atomicity guarantees that a native upsert can and are very vulnerable to races
+        and crashes. Therefore if we wish to upsert without an appropriate unique index,
+        we must either:
+
+        1. Acquire a table-level lock before the emulated upsert (`lock=True`), or
+        2. VERY CAREFULLY ensure that we are the only thread and worker which will be
+           writing to this table, in which case we can proceed without a lock
+           (`lock=False`).
+
+        Generally speaking, you should use `lock=True`. If the table in question has a
+        unique index[*], this class will use a native upsert (which is atomic and so can
+        ignore the `lock` argument). Otherwise this class will use an emulated upsert,
+        in which case we want the safer option unless we been VERY CAREFUL.
 
-        `lock` should generally be set to True (the default), but can be set
-        to False if either of the following are true:
-            1. there is a UNIQUE INDEX on the key columns. In this case a conflict
-            will cause an IntegrityError in which case this function will retry
-            the update.
-            2. we somehow know that we are the only thread which will be updating
-            this table.
-        As an additional note, this parameter only matters for old SQLite versions
-        because we will use native upserts otherwise.
+        [*]: This class is aware that some tables have unique indices added as
+             background updates. It checks at runtime to see if those updates are
+             pending: if not, they are deemed safe for use with native upserts.
 
         Args:
             table: The table to upsert into