Rewrite the "sync_local" query.

Author: rkistner

crates/core/src/sync_local.rs

@@ -55,50 +55,98 @@ pub fn sync_local(db: *mut sqlite::sqlite3, _data: &str) -> Result<i64, SQLiteEr
 
     // Query for updated objects
 
+    // Be careful with modifying this query - it is critical for performance. When modifying, make sure to check
+    // performance of the query with a large number of rows, and also with a large number of duplicate rows (same row_id).
+    //
+    // This form uses a subquery with max(r.op_id) instead of a JOIN to get the latest oplog entry for each updated row.
+    // The subquery is because:
+    // 1. We need the GROUP BY to execute _before_ looking up the latest op_id for each row, otherwise
+    //    we get terrible performance if there are lots of duplicate ids (O(N^2) performance).
+    // 2. We want to avoid using a second GROUP BY, which would use a secondary TEMP B-TREE.
+    //
+    // It does not appear to be feasible to avoid the single TEMP B-TREE here.
+    //
+    // The query roughly does the following:
+    // 1. Filter oplog by the ops added but not applied yet (oplog b). These are not unique.
+    // 2. Use GROUP BY to get unique rows. This adds some overhead because of the TEMP B-TREE, but is necessary
+    //    to cover cases of duplicate rows. DISTINCT would do the same in theory, but is slower than GROUP BY in practice.
+    // 3. For each op, find the latest version of the data. This is done using a subquery, with `max(r.op_id)`` to
+    //    select the latest version.
+    //
+    // The subquery instead of a JOIN is because:
+    // 1. We need the GROUP BY to execute _before_ looking up the latest op_id for each row, otherwise
+    //    we get terrible performance if there are lots of duplicate ids (O(N^2) performance).
+    // 2. We want to avoid using a second GROUP BY, which would use a second TEMP B-TREE.
+    //
+    // The `ifnull(data, max(op_id))` clause is a hack to pick the row with the largest op_id, but only select the data.
+    //
+    // QUERY PLAN
+    // |--CO-ROUTINE updated_rows
+    // |  `--COMPOUND QUERY
+    // |     |--LEFT-MOST SUBQUERY
+    // |     |  |--SCAN buckets USING COVERING INDEX ps_buckets_name
+    // |     |  `--SEARCH b USING INDEX ps_oplog_opid (bucket=? AND op_id>?)
+    // |     `--UNION ALL
+    // |        `--SCAN ps_updated_rows
+    // |--SCAN b
+    // |--USE TEMP B-TREE FOR GROUP BY
+    // `--CORRELATED SCALAR SUBQUERY 3
+    //    `--SEARCH r USING INDEX ps_oplog_row (row_type=? AND row_id=?)
+
     // language=SQLite
     let statement = db
         .prepare_v2(
             "\
--- 1. Filter oplog by the ops added but not applied yet (oplog b).
---    SELECT DISTINCT / UNION is important for cases with many duplicate ids.
 WITH updated_rows AS (
-  SELECT DISTINCT b.row_type, b.row_id FROM ps_buckets AS buckets
-    CROSS JOIN ps_oplog AS b ON b.bucket = buckets.id
-  AND (b.op_id > buckets.last_applied_op)
-  UNION SELECT row_type, row_id FROM ps_updated_rows
+    SELECT b.row_type, b.row_id FROM ps_buckets AS buckets
+        CROSS JOIN ps_oplog AS b ON b.bucket = buckets.id
+        AND (b.op_id > buckets.last_applied_op)
+    UNION ALL SELECT row_type, row_id FROM ps_updated_rows
 )
 
--- 3. Group the objects from different buckets together into a single one (ops).
-SELECT b.row_type as type,
-    b.row_id as id,
-    r.data as data,
-    count(r.bucket) as buckets,
-    /* max() affects which row is used for 'data' */
-    max(r.op_id) as op_id
--- 2. Find *all* current ops over different buckets for those objects (oplog r).
-FROM updated_rows b
-    LEFT OUTER JOIN ps_oplog AS r
-               ON r.row_type = b.row_type
-                 AND r.row_id = b.row_id
--- Group for (3)
-GROUP BY b.row_type, b.row_id",
+SELECT
+    b.row_type,
+    b.row_id,
+    (
+        SELECT ifnull(r.data, max(r.op_id))
+                 FROM ps_oplog r
+                WHERE r.row_type = b.row_type
+                  AND r.row_id = b.row_id
+    ) as data
+    FROM updated_rows b;
+    GROUP BY b.row_type, b.row_id;
+    ",
         )
         .into_db_result(db)?;
 
-    // TODO: cache statements
+    // An alternative form of the query is this:
+    //
+    // SELECT r.row_type as type,
+    //     r.row_id as id,
+    //     r.data as data,
+    //     max(r.op_id) as op_id
+    // FROM ps_oplog r
+    // GROUP BY r.row_type, r.row_id;
+    //
+    // This form is simple and fast, but does not filter only on updated rows. It also ignores ps_updated_rows.
+    // We could later add heuristics to use this form on initial sync, or when a large number of rows have been re-synced.
+    //
+    // QUERY PLAN
+    // `--SCAN r USING INDEX ps_oplog_row
+
+    // TODO: cache individual statements
 
     while statement.step().into_db_result(db)? == ResultCode::ROW {
         let type_name = statement.column_text(0)?;
         let id = statement.column_text(1)?;
-        let buckets = statement.column_int(3)?;
         let data = statement.column_text(2);
 
         let table_name = internal_table_name(type_name);
 
         if tables.contains(&table_name) {
             let quoted = quote_internal_name(type_name, false);
 
-            if buckets == 0 {
+            if data.is_err() {
                 // DELETE
                 let delete_statement = db
                     .prepare_v2(&format!("DELETE FROM {} WHERE id = ?", quoted))
@@ -115,7 +163,7 @@ GROUP BY b.row_type, b.row_id",
                 insert_statement.exec()?;
             }
         } else {
-            if buckets == 0 {
+            if data.is_err() {
                 // DELETE
                 // language=SQLite
                 let delete_statement = db