Report how many changes were done in the interface

plans/02-write-statistics.md

@@ -1,5 +1,30 @@
 # Write statistics
 
+## Status
+
+Done. Queries that can write report `stats` in the final `PULL` summary, with the
+node, relationship and label counters exact, and `properties-set` approximate as
+described below. `CREATE DATABASE` and `DROP DATABASE` report system updates.
+
+The spike and testing changed some details below:
+
+- The counters do **not** reset at commit: they stay until PostgreSQL flushes them,
+  some time later. That makes the difference only reliable within one transaction,
+  which is still why each write runs in its own. Before, query and after are sent as
+  one script, so this costs no extra round trips.
+- The counters are read with `pg_stat_get_xact_tuples_*`, not the
+  `pg_stat_xact_user_tables` view.
+- Label tables are found through `pg_class` and `pg_inherits`, not `ag_label`: reading
+  `ag_label` made writes fail for roles without `SELECT` on it.
+- `labels-removed` is not reported. Deleting a node does not count in Neo4j, and AGE
+  cannot remove a label.
+- `SET n.a = 1, n.b = 2` is one update per node, so `properties-set` is the number of
+  nodes and relationships changed, not of properties.
+
+The price is about 1 ms per query that can write (2.1 ms to 3.0 ms for a single-row
+`CREATE` on a local database). A statement that writes many rows pays this once.
+Caching a graph's label tables could lower it.
+
 ## Problem
 
 Neo4j reports what a write query changed in the `stats` entry of the final `PULL`

plans/README.md

@@ -28,9 +28,9 @@ plans small:
    transaction-local counters, streaming (06) needs a transaction to keep a server-side
    cursor open, and `PROFILE` (01) needs a savepoint. All three want `Session.run` to
    wrap auto-commit queries in `BEGIN ... COMMIT` instead of relying on the
-   connection's autocommit mode. Plan 06 only did this for read queries, which run
-   in a cursor's transaction; queries that can write still use autocommit, so 02 still
-   has to add it for them.
+   connection's autocommit mode. This is done: reads run in a cursor's transaction
+   (06), and queries that can write in one around the statistics (02). `Session._script`
+   runs such a transaction in one round trip.
 3. **Unit tests.** The only test is the end-to-end `tests/smoke_test.py`. Most of
    this work is text rewriting in `cypher.py`, which is cheap to unit test without a
    database. Add `tests/test_cypher.py` (pytest) alongside the first plan that