Page store: only store the key when deleting, not the whole row.

h2/src/main/org/h2/index/PageBtreeNode.java

@@ -66,8 +66,8 @@ public class PageBtreeNode extends PageBtree {
      * Create a new b-tree node page.
      *
      * @param index the index
-     * @param data the data
      * @param pageId the page id
+     * @param parentPageId the parent page id
      * @return the page
      */
     public static PageBtreeNode create(PageBtreeIndex index, int pageId, int parentPageId) {

h2/src/main/org/h2/index/PageScanIndex.java

@@ -356,6 +356,16 @@ public class PageScanIndex extends PageIndex implements RowIndex {
     }
 
     public Row getRow(Session session, int key) throws SQLException {
+        return getRow(key);
+    }
+
+    /**
+     * Get the row with the given key.
+     *
+     * @param key the key
+     * @return the row
+     */
+    public Row getRow(long key) throws SQLException {
         PageData root = getPage(rootPageId, 0);
         return root.getRow(key);
     }

h2/src/main/org/h2/store/PageLog.java

@@ -251,21 +251,37 @@ public class PageLog {
                             undoAll.set(pageId);
                         }
                     }
-                } else if (x == ADD || x == REMOVE) {
+                } else if (x == ADD) {
                     int sessionId = in.readInt();
                     int tableId = in.readInt();
                     Row row = readRow(in, data);
-                    if (stage == RECOVERY_STAGE_UNDO && x == ADD) {
+                    if (stage == RECOVERY_STAGE_UNDO) {
                         store.allocateIfIndexRoot(pos, tableId, row);
                     } else if (stage == RECOVERY_STAGE_REDO) {
                         if (isSessionCommitted(sessionId, logId, pos)) {
                             if (trace.isDebugEnabled()) {
-                                trace.debug("log redo " + (x == ADD ? "+" : "-") + " table:" + tableId + " " + row);
+                                trace.debug("log redo + table:" + tableId + " " + row);
                             }
-                            store.redo(pos, tableId, row, x == ADD);
+                            store.redo(pos, tableId, row, true);
                         } else {
                             if (trace.isDebugEnabled()) {
-                                trace.debug("log ignore s:" + sessionId + " " + (x == ADD ? "+" : "-") + " table:" + tableId + " " + row);
+                                trace.debug("log ignore s:" + sessionId + " + table:" + tableId + " " + row);
+                            }
+                        }
+                    }
+                } else if (x == REMOVE) {
+                    int sessionId = in.readInt();
+                    int tableId = in.readInt();
+                    long key = in.readLong();
+                    if (stage == RECOVERY_STAGE_REDO) {
+                        if (isSessionCommitted(sessionId, logId, pos)) {
+                            if (trace.isDebugEnabled()) {
+                                trace.debug("log redo - table:" + tableId + " key:" + key);
+                            }
+                            store.redoDelete(pos, tableId, key);
+                        } else {
+                            if (trace.isDebugEnabled()) {
+                                trace.debug("log ignore s:" + sessionId + " - table:" + tableId + " key:" + key);
                             }
                         }
                     }
@@ -523,9 +539,13 @@ public class PageLog {
             out.write(add ? ADD : REMOVE);
             out.writeInt(session.getId());
             out.writeInt(tableId);
-            out.writeInt(row.getPos());
-            out.writeInt(data.length());
-            out.write(data.getBytes(), 0, data.length());
+            if (add) {
+                out.writeInt(row.getPos());
+                out.writeInt(data.length());
+                out.write(data.getBytes(), 0, data.length());
+            } else {
+                out.writeLong(row.getPos());
+            }
             flushOut();
         } catch (IOException e) {
             throw Message.convertIOException(e, null);

h2/src/main/org/h2/store/PageStore.java

@@ -79,25 +79,24 @@ import org.h2.value.ValueString;
  */
 public class PageStore implements CacheWriter {
 
+    // TODO delete: only log the key
+    // TODO update: only log the key and changed values
     // TODO fix page format of data overflow and so on
     // TODO implement checksum; 0 for empty pages
-    // TODO in log, don't store empty space between page head and page data
+    // TODO undo log: fully compress empty pages
+    // TODO undo log: don't store empty space between head and data
+    // TODO undo log: lzf compression
     // TODO long primary keys don't use delegating index yet (setPos(): int)
+    // TODO replace CRC32
+    // TODO maybe remove parent pointer
+    // TODO index creation: use less space (ordered, split at insertion point)
+    // TODO test running out of disk space (using a special file system)
 
     // TODO utf-x: test if it's faster
     // TODO after opening the database, delay writing until required
-    // TODO maybe remove parent pointer
-    // TODO replace CRC32
     // TODO optimization: try to avoid allocating a byte array per page
     // TODO optimization: check if calling Data.getValueLen slows things down
-    // TODO undo pages: don't store the middle zeroes
-    // TODO undo pages compression: try http://en.wikipedia.org/wiki/LZJB
     // TODO order pages so that searching for a key only seeks forward
-    // TODO completely re-use keys of deleted rows; maybe
-    // remember last page with deleted keys (in the root page?),
-    // and chain such pages
-    // TODO delete: only log the key
-    // TODO update: only log the key and changed values
 
     // TODO detect circles in linked lists
     // (input stream, free list, extend pages...)
@@ -105,16 +104,12 @@ public class PageStore implements CacheWriter {
     // synchronized correctly (on the index?)
     // TODO remove trace or use isDebugEnabled
     // TODO recover tool: don't re-do uncommitted operations
-    // TODO no need to log old page if it was always empty
     // TODO don't store default values (store a special value)
     // TODO split files (1 GB max size)
     // TODO add a setting (that can be changed at runtime) to call fsync
     // and delay on each commit
-    // TODO PageData and PageBtree addRowTry: try to simplify
-    // TODO test running out of disk space (using a special file system)
     // TODO check for file size (exception if not exact size expected)
     // TODO implement missing code for STORE_BTREE_ROWCOUNT (maybe enable)
-    // TODO store dates differently in Data; test moving db to another timezone
     // TODO online backup using bsdiff
 
     // TODO when removing DiskFile:
@@ -1047,6 +1042,20 @@ public class PageStore implements CacheWriter {
         }
     }
 
+    /**
+     * Redo a delete in a table.
+     *
+     * @param logPos the redo log position
+     * @param tableId the object id of the table
+     * @param key the key of the row to delete
+     */
+    void redoDelete(int logPos, int tableId, long key) throws SQLException {
+        Index index = metaObjects.get(tableId);
+        PageScanIndex scan = (PageScanIndex) index;
+        Row row = scan.getRow(key);
+        redo(logPos, tableId, row, false);
+    }
+
     /**
      * Redo a change in a table.
      *