RFC 1024: Index ordering

text/1024-index-ordering.rst

@@ -0,0 +1,95 @@
+::
+
+    Status: Draft
+    Type: Feature
+    Created: 2023-08-09
+    Authors: Victor Petrovykh <[email protected]>
+
+========================
+RFC 1024: Index Ordering
+========================
+
+This RFC outlines an improvement to EdgeDB indexes that should make it easier
+for users to write queries that actually use the index speedups.
+
+
+Motivation
+==========
+
+In Postgres the indexes support "ordering" specificaion, so one can write:
+``CREATE INDEX test3_desc_index ON test3 (id DESC NULLS LAST);`` or ``CREATE
+INDEX test3_xy_index ON test3 (x ASC, y DESC);``. This allows fine-tuning
+indexes to the types of queries that you expect. We should replicate this
+capability in EdgeDB.
+
+
+Specification
+=============
+
+Instead of treating the ``on (...)`` clause as an expression, we can upgrade
+it to actually be an ordering clause for the index with syntax much like order
+by::
+
+  index on (
+      <expr> [ acs | desc ] [ empty { first | last } ] [ then ... ]
+  )
+  [ except ( <except-expr> ) ]
+  [ "{" <annotation-declarations> "}" ] ;
+
+For example::
+
+  type Foo {
+    text: str;
+    x: int64;
+    required y: int64;
+
+    index on (.text asc empty last);
+    index on (.x desc empty last then .y asc);
+  }
+
+Notice that this approach works well with indexes that use more than one field
+and potentially need different ordering for different fields.
+
+Additionally, since this syntax naturally allows specifying several fields for
+indexing ``index on (.a then .b then .c)``, we no longer need to use a tuple
+expression to do that, removing the special case and making the index
+expression more semantically consistent.
+
+
+Implementation
+--------------
+
+This change maps very naturally onto Postgres indexes. In particular, it
+removes the need to have a special case handling for tuples as index
+expression as we no longer need to use that format to create an index over
+multiple fields::
+
+  index on (.x desc empty last then .y asc);
+
+can be translated into::
+
+  CREATE INDEX foo_index ON "Foo" (x DESC NULLS LAST, y ASC);
+
+
+Backwards Compatibility
+=======================
+
+This change is *syntactically* backwards compatible since all of the extra
+specification for the index are optional. However, sematically, we should not
+interpret index on a tuple expression as index on several fields anymore. We
+should urge people to update their schemas to use this new version of index
+instead of a tuple expression. In theory this would involve a migration that
+only actually changes the schema text (and therefore the hash), but doesn't
+affect existing indexes.
+
+
+Security Implications
+=====================
+
+There are no security implications.
+
+
+Rejected Alternative Ideas
+==========================
+
+. . .